# Phase-1 skeleton — the NL parser (v0 spec, 2026-07-05) **Status:** design settled in session 2026-07-05; NOTHING BUILT YET. This is the build plan for the first Alternator Phase-1: a frozen-trunk parser that turns templated NL into factor-graph deltas in the deducer's own vocabulary, closing the loop so the alternation exists at all. **No external API calls anywhere — the point of the project is the small-footprint local model.** Context: `docs/phase1_construction_brief.md`, `docs/phase1_prep_grounding.md`, CLAUDE.md §8. ## 0. Scope discipline (what this skeleton is and is not) - **It is the MECHANISM testbed** (Job A): does the Alternator loop work — cycles, deltas, notebook conditioning, NACK response, zero-LoRA null, parse-side silhouette. Domain = KenKen-in-words, where gold graphs are free, the deducer is trained, the verifier is exact, and the full NACK stack (calib head / VIOLATED flags / late-JSD 0.687) is live on the deduce side. - **It is NOT the value probe** (Job B — does parse→solve beat LLM-direct on real math). DEFERRED WITH A DATE, not dropped: the fully-local diagnostic is *decisions-per-problem* from the search tier over MATH-500-style problems (0 decisions = calculator band, >0 = engine band — maps where the bet lives with our own machinery only). **Milestone gate: run it once the skeleton's Brick-A measurable is in.** Job A must not silently absorb months. ## 1. The two jaws (the corrected metaphor — bake it in) NOT "one jaw for hard constraints, one for soft." On every problem the jaws cooperate in the validated division of labor: - **Symbolic jaw** (GAC / MRV / LCV / verifier / backtracking): DISPOSES — commits only what is logically forced. Needs exact predicates to bite. - **Neural jaw** (the deducer): PROPOSES — ordering priors, soft-graph inference, differentiable critique. Never commits. Hard-vs-soft determines **which jaw can bite at all**: on clean constraints the symbolic jaw dominates for free (the honest negative); on soft / learned / NL-uncertain constraints it has nothing to chew and the neural jaw is the only one working — the Alternator's target regime. Marine version: grouper + moray hunting COOPERATIVELY, each reaching where the other can't (the moray has the pharyngeal second jaws; the grouper signals it into the crevices). ## 2. Architecture - **Trunk:** Llama-3.2-1B **L0–L3, FROZEN**, 2048d (weights on disk: `.cache/llama-3.2-1b-weights`). Frozen is the falsifiability of the zero-LoRA null: all per-cycle behavior change must come from INPUT conditioning (notebook + NACK), not weight mutation. Fallback ladder if the null fails: rank-≤16 LoRA on the parser only (§8.4) — never the deducer trunk. - **Parse-side WAIST (built from birth, day one):** a readout bottleneck just before the delta head — 2048 → **512d**, importance-ordered dims, Matryoshka prefix mask **512→128** (two schedule axes per §8.6: training-time handicap + per-cycle coarse-narrow/fine-wide). This is the **parser's silhouette tap** — the object the perceiver segments and the BirdNET re-run reads. It mirrors the deducer's *tap* (readout probe), NOT its dormant in-loop waist. Capture hooks from birth (capture-once schema, token-position × waist-dim). - **Delta head:** slot-based parallel emission (§3). Trained; the only trained parameters besides the waist projection in v0. - **Deducer (Phase 2):** untouched, exactly §1–§2 of CLAUDE.md. Never sees NL. Budget: frozen L0–L3 + embeddings fp16 ≈ 1 GB; deducer fp32 ≈ 0.4 GB; fits 24 GB with room for activations. Smoke-test the combined residency before anything else. ## 3. The delta-head output format (v0 — the expensive-to-change object) **Slot-based, parallel, no autoregression** (the §6 no-mid-breath-token-generation law applies to the parser too: one structured emission per cycle, not a token stream). L_out fixed factor slots (DETR-style), each slot emits: | field | form | supervision | |---|---|---| | presence | 1 logit | BCE vs gold (slot count varies per cycle) | | type | logits over the REGISTRY MENU (v0 = the deducer's latent types: `row`, `col`, `cage`) | CE | | op (cage only) | logits over OP_VOCAB `{given,add,sub,mul,div}` (ids 0–4, existing) | CE | | target (cage only) | **digit-spaced**: 3 × 10-way digit heads (§6 law: never whole-number classes). Exact integer — the SYMBOLIC verifier needs exact targets; the inlet's log-buckets are derived downstream via the existing `target_to_bucket` | 3 × CE | | membership | pointer distribution over the s_max=49 cell slots (multi-hot) | per-cell BCE vs gold multi-hot | - **Attention-bootstrap law compliance (§6):** the membership pointer is a ~49-way attention pathway — it does NOT bootstrap from task gradient. It doesn't have to: gold deltas give **direct supervision** from step 0 (the sanctioned escape, observed 4× as the requirement). - **Slot↔gold matching:** v0 = CANONICAL ORDER (sort gold factors by (type, first-member cell index)) with positional supervision. Hungarian matching is the fallback if positional proves brittle — do not build it preemptively. - **Registry (SEMANTICS channel):** ONE menu, fixed meaning across cycles (§8.2). v0 menu == the deducer's existing vocabulary verbatim (format-definer role: the deducer's membership + inlet vocabulary IS the target). Centroids live in one shared embedding space; classification is cosine-to-centroid — the same space the perceiver's segmenter reads later. - **Ball (TOPOLOGY channel):** v0 = the parser emits MEMBERSHIP DIRECTLY; the Poincaré ball sits behind a flag as a strictly-additive upgrade (§7's relaxation is blocked; hard masks are the permanent fallback BY DESIGN — alternation must not wait on unsolved geometry). - **Notebook (TEMPORAL memory):** WRITTEN from the deducer's silhouette tap (readout-LN — where all evidence lives) + the NACK features; READ by the parser as prefix conditioning (the concrete mechanism the zero-LoRA null stands on). Ledger (append-only, committed facts: settled assignments, verified factors) + scratch (replace, provisional). NACK feature vector v0: per-factor VIOLATED flags (symbolic, exact, free) + the per-cell late-breath belief-JSD field (gold-free, AUC 0.687) + cycle index. ## 4. The template generator + gold labeling (the other expensive object) Generate KenKen with the existing builder, render to NL: - **Preamble** (one sentence): grid size + the row/col all-different rule ("Solve the 5×5 KenKen: every row and every column contains 1–5 exactly once."). The parser must emit ALL row+col factors from this ONE sentence — the deliberate one-sentence→many- factors case. - **One sentence per cage**, template bank with paraphrase variation ("The cage covering r1c1 and r1c2 multiplies to 12" / "Cells (1,1) and (1,2) have product 12" / …). **Givens** as their own sentences ("Row 3, column 4 is a 5."). - **GOLD = (char/token span ↔ induced factor(s)) alignment**, emitted by the generator for free. This single labeling scheme supervises THREE things at once: 1. the delta head (which factors, per cycle); 2. **band masking / remove-at-read on the INPUT side**: once a factor is matched to a registry centroid (committed to the ledger), its token span is masked at read — explained-away text drops out and each later cycle parses the RESIDUAL unexplained text (Law 7 applied to input; the polarized-sunglasses mechanism with no Fourier machinery); 3. segmentation gold for the parse-side BirdNET re-run (token spans ARE the calls). - **Curriculum knobs:** template diversity, paraphrase depth, sentence-order shuffling, distractor sentences, factors-per-cycle cap (forces genuine multi-cycle parses). - **NACK curriculum:** corrupted-parse variants (wrong target / wrong member / wrong op injected mid-session) with gold retransmissions — Brick-C-v0's training and eval data. ## 5. The energy wave (Phase-1 form — HYPOTHESIS, say so) Per-token residual energy across the SEQUENCE, per cycle — the parse-side analog of the deducer's per-breath wave. Hypotheses to test (not assume): - energy peaks align with semantically dense spans (quantities, relations); - the wave's spatial frequency ALONG THE TOKEN AXIS carries the band structure — coarse cycles the low-frequency envelope (document scaffold), fine cycles the high-frequency detail (exact operands); - band masking (§4.2) is the mechanism that *changes* the wave cycle-to-cycle. Parse-side priors genuinely favor the bird pipeline (text is narratively banded; loosely-coupled sentences compose far more linearly than joint deduction) — but the deduce-side lesson stands: MEASURE, the field not the summary statistic, before building any machinery on it. Re-run the exact capture-once protocol (`scripts/capture_silhouette_trajectories.py` pattern): banding + linearity on token-position × waist-dim, the moment the skeleton trains. ## 6. Build order + measurables + kill gates 1. **Residency smoke:** frozen Llama L0–L3 + deducer co-resident on the 7900 XTX. 2. **Template generator + gold labeling** (§4) — CPU, selftested. 3. **Delta head + waist** (§3, §2) supervised on single-cycle parses (no alternation yet): parse-accuracy vs gold deltas is the unit test. 4. **Brick-A (the zero-LoRA null):** cycle-conditioned input (blank notebook vs rich ledger + NACK features) through the SAME frozen trunk — does parse behavior change appropriately with NO weight mutation? Kill: if conditioning can't switch bands, engage the LoRA fallback ladder (§8.4) and say so. 5. **Brick-C-v0 (NACK response):** inject corruption → verifier VIOLATED + late-JSD flag it → NACK features next cycle → retransmission accuracy vs a no-NACK control. **The §8.3 gate is binding: if NACK-response does not beat staged parse-then-solve, the alternation folds back to the simpler staged design by its own spec.** 6. **Parse-side BirdNET re-run** (§5) — free once (3) trains. 7. **Milestone gate (Job B, local-only):** decisions-per-problem over MATH-500-style problems via the search tier — map the engine band. Runs no later than Brick-A completion. ## 7. Honest open risks - 1B-class comprehension may be too weak even for templated NL at high paraphrase diversity — the curriculum knobs are the instrument; find the ceiling honestly. - The zero-LoRA null may fail (§8.7 #1) — the fallback ladder exists; failing the null is a RESULT, not a defeat. - Slot-based emission with canonical-order matching may be brittle under sentence shuffling — Hungarian fallback documented above. - The engine-band question (Job B) remains existential and open — the milestone gate keeps it in the crosshairs. ## 8. Build log (results as they land) - **2026-07-05/06 — steps 1-3 BUILT.** Generator (`scripts/kenken_nl_gen.py`): span-SET gold + split-ref family + round-trip-as-generation-gate (4,140/4,140 samples pass; labeling bugs die at generation). Residency (`scripts/phase1_residency_smoke.py`): 2.9GB co-resident, ~21GB headroom, JIT replay 0.34s, no AM hazards, deducer unperturbed (0.745). Delta head (`scripts/phase1_delta_head.py`): 3.2M params, three loss-decomposition-diagnosed design iterations (TEXT-ORDER slots — canonical order made slot->sentence assignment a circular grid-sort; membership pos-weight 5.0; per-token SENTENCE-INDEX embedding — one attention hop matches a discrete code but cannot COUNT sentences: the attention-bootstrap law's quieter cousin). - **Data scale is the current lever:** 300 samples -> memorization (train mem loss 0.002, test F1 0.76); 3,060 samples -> reading (factor exact 0.60 -> **0.748**, op 0.944, target 0.865, member F1 0.891). One-shot solve rate 0/60 — on-model at p^~20 compounding; NOT the Alternator's operating point (see below). - **Matryoshka answer (parse side, first pass): the parse signal is LOW-DIMENSIONAL.** Width 128 ~= width 512 on every head (factor exact 0.724 vs 0.748; op/type flat; the ~2pt cost sits in the fine-detail heads, target/member). Head-aware waist scheduling not currently justified — uniform prefix suffices. - **ERROR TAXONOMY (2026-07-06, the pre-Brick-C gate): 60/60 parse errors are SYMBOLICALLY DETECTABLE — zero SILENT.** 44 UNSAT + 16 malformed; DETECT_multi 0 (over-constraint dominates); every failure involves a membership error. So at factor-exact 0.748, the one-shot pipeline solves 0% but the symbolic NACK has 100% recall on flagged problems — the alternation-earns-its-cost story stated quantitatively: staged one-shot 0%, NACK-recoverable ceiling 100%. CAVEATS: n=60; detectability~=1 leans on KenKen being densely over-constrained (rows/cols interlock every cage) — sparser domains will have a real SILENT class; UNSAT flags the GRAPH, not the factor — Brick-C-v0 needs a localization story (unsat-core-ish or per-factor blame) on top of the flag. Tier-3 late-JSD recall sits above this ceiling, unmeasured. - **NEXT:** membership exactness is THE lever (present in 100% of failures) — scale data further / targeted membership curriculum; then Brick-A (notebook/NACK conditioning, zero-LoRA null), Brick-C-v0 (retransmission vs the no-NACK control), parse-side BirdNet re-run on the captured waist silhouettes. - **THE EXPIRATION CONDITION on the 100% (write it before it bites):** zero-silent is partly a property of the current ERROR MIX, not only of KenKen's density — membership errors (100% of today's failures) are the structurally LOUD kind (a mispointed cage fights the row/col lattice). As membership approaches exactness, residual failures rotate toward target-digit/op errors — the plausible coherent-misreading class — and a SILENT class can be BORN exactly as the parser improves. Discipline: the taxonomy is cheap; RE-RUN AT EVERY ACCURACY CHECKPOINT and track detectable-fraction AS A FUNCTION OF factor-exactness. Pinned at 100% while the mix rotates = a much stronger claim; a silent class at 0.9+ = learned while the tier-0 confidence head can still be designed to catch it. - **Brick-C localization v0 = the DELETE-ONE-FACTOR BLAME SWEEP:** remove each parsed factor in turn, re-solve; SAT-on-removal fingers that factor as unsat-core member. O(F) search-tier calls at median-zero decisions ≈ less than one deducer breath; no new machinery (the taxonomy's bridge). Not minimal-core (overlapping errors smear blame — measure it). The tier-0 confidence head later slots in as the sweep ORDER (least-confident first): propose/dispose fractally repeated INSIDE the NACK. - **The guard: DETECTION IS NOT CORRECTION.** 100% is a detection ceiling; Brick-C must demonstrate that a NACK-conditioned re-parse FIXES the flagged region rather than re-emitting the same wrong membership. Encouraging structure (retransmission is an easier problem: ledger pins the verified, attention has fewer places to go) is exactly the kind of claim this project measures rather than assumes. - **BLAME SWEEP v0 MEASURED (2026-07-06) — an honest negative with a design lesson:** delete-one-factor re-solve on the 44 UNSAT parses: precision **1.000** (when it fires, it is right), recall **0.034** (it fired on 2/44 — at factor-exact 0.75 a parse carries ~5 wrong factors and single deletion cannot restore SAT). So delete-one is the SINGLE-ERROR-REGIME tool: its usefulness co-improves with parser exactness. At realistic multi-error density, SYMBOLIC localization alone fails — which MOTIVATES (by measurement, not taste) the neural NACK tiers: (a) the add-back sweep (start rows/cols/givens, add cages in tier-0-confidence order, blame additions that turn UNSAT), and (b) the deducer's soft-solve suspicion field (per-region late-JSD on the parsed graph — the tier-3 role the Alternator assigned it a priori). The propose/dispose fractal is now REQUIRED inside the NACK, not merely elegant. - **REGISTERED PREDICTION #2 (2026-07-06, ahead of the 40k curve):** the SILENT class is born first in TARGET-DIGIT errors on `add` cages (most compensating-coincidence room — many operand sets reach the same sum) and stays near-zero for membership as long as the row/col lattice polices. Where silents appear = where the tier-0 confidence head's calibration matters most. - **Brick-A operationalization (pre-specified as a FIELD, per the instrument lesson):** the zero-LoRA read is DIFFERENTIAL — same puzzle, same frozen weights, blank-notebook input vs ledger+NACK input; the measure is whether the parse DELTA (attention mass + emission changes) CONCENTRATES on the flagged region vs a global reshuffle. A scalar accuracy delta is the summary that will degenerate; specify the localization field from the start. - **CURVE POINTS 2-3 + THE FIRST SOLVE (2026-07-06, 40k corpus, 24k steps):** factor exact 0.753 -> **0.780** (op 0.967 / target 0.907 / member F1 0.924, all climbing) and **SOLVE RATE 1/60 — the first end-to-end NL -> parse -> symbolic solve -> correct grid**, at BOTH Matryoshka widths (128d factor exact 0.749; the solve survives narrow). Taxonomy: **third consecutive 100% detectable** (59/59; zero silent) while the mix rotates (target-digit errors growing share, 28/52 UNSAT; membership still in every failure; presence split: phantoms 9 vs dropped 5 — mixed, mildly favoring the tier-0 presence-gate preview). Prediction #2 remains properly unfalsified (trigger condition — mix clearing membership — not yet reached). **BLAME CO-IMPROVEMENT MEASURED:** delete-one recall 0.034 -> **0.067** (doubled, restored 5/52 vs 2/44; precision 1.0 -> 0.8) as the parser improved — the single-error-tail convergence claim now has its first two points. - **THE FINE-CADENCE CURVE (2026-07-06 night, 7 points, 12k->44k steps):** factor exact 0.753 -> 0.780 -> 0.805 -> 0.814 -> 0.827 -> 0.825 -> 0.829 (PLATEAU ~0.83 — the cheap-steps lever is exhausted; loss flat at ~3.34 with span at its entropy floor); SOLVE RATE 0 -> 1 -> 1 -> 2 -> 5 -> 5 -> 5 /60 (8.3%). **SEVEN consecutive 100%-detectable taxonomy points, zero silent**, while the mix rotated hard toward target digits (dominant cage-field error) — prediction #2 STILL unfalsified with its trigger zone now fully occupied: target errors dominate and no silent has appeared. DETECT_multi births at stage 3 (dropped factors -> under- constrained -> caught by the uniqueness probe — the third detection channel now live). NEW CLASS OBSERVED: CORRECT-with-field-errors (up to 2/5 solves carry a member mismatch) — wrong-but-EQUIVALENT parses that solve to gold: benign errors, worth excluding from exactness metrics later. **DECISION POINT: the 0.83 plateau means the next lever is architectural (second attention hop / iterative refinement) OR the loop itself (Brick-A/C — the design's actual answer to imperfect parsers). Brick-A is next per the queue, and now inherits a stable, characterized parser.** - **NAMED: the packing finding is NEURAL COLLAPSE** (Papyan/Han/Donoho — terminal-phase CE pulls class vectors toward a simplex equiangular tight frame; K=7 ideal cosine -1/6 ~= -0.167; measured -0.048 = collapse one-third realized, penalty OFF). **REGISTERED PREDICTION #3 (the mechanism test):** training with a simplex-ETF codebook penalty shifts confusion mass SPECIFICALLY off the thin walls (6-7 first, 1-2 second — the rho=0.53 geometry-predicts-confusions link is the mechanism). If accuracy rises but the confusion matrix reshuffles uniformly, the geometry story was correlation, not cause. (Deducer-side; fire on a free training slot.) - **SEMANTIC EXACTNESS (the wrong-but-equivalent consequence):** factor-exact systematically UNDERSELLS the parser — wrong-but-equivalent parses solve to gold. The honest per-parse metric is SOLUTION-SET EQUIVALENCE (solve the parsed graph, compare grids — the round-trip machinery computes it free). Long-term (Job B): real math admits many valid formalizations; grading on syntactic match to one canonical gold graph punishes legitimate readings. **The generator's unique-solution guarantee is what makes equivalence CHECKABLE — preserve it as a HARD REQUIREMENT when the domain widens.** - **LEVER ORDERING (settled by the plateau):** Brick-A next (wants the stable parser), Brick-C after (55 flagged failures are its specified input; a plateaued parser with perfect error visibility makes any loop gain UNCONFOUNDABLE with parser improvement), the architectural lever (second attention hop) stays PARKED unless the loop's ceiling proves too low. Information value over fun. - **Prediction #2 status note:** the trigger zone is fully occupied (target digits dominant) with no silent birth through 7 points. If the window closes without one, record THAT as the finding — KenKen's lattice polices even compensating-coincidence- rich errors — which sets the baseline for when sparser domains DO grow a silent class. Either resolution is informative. ## 9. Brick-A design (registered 2026-07-06, before build) 1. **Conditioning enters TRUNK-LEVEL** (prefix embeddings prepended to text, flowing through the frozen bidirectional Llama layers) — the honest zero-LoRA test: the frozen trunk must transform conditioning into different parse behavior. Head-level injection is the cheap comparison arm (if it matches trunk-level, the trunk adds no conditioning value — a finding to discover, not assume). Cost accepted: the banked precompute doesn't apply to prefixed passes; live trunk forwards (0.34s/batch). 2. **The zero-LoRA boundary:** trainable params in the EMISSION head and in ONE SHARED conditioning encoder are sanctioned (they are the interface); anything that varies PER CYCLE is forbidden. Cycles differ only in what the notebook contains, never in which parameters run. 3. **Conditioning corpus = the plateaued parser's ORGANIC failures** on train puzzles (real error distribution), supervised against gold fixes; synthetic corruptions are augmentation only. **Refinement (arm separation):** Brick-A conditions on ORACLE localization (gold-derived wrong-sentence flags) — it tests "can the frozen trunk USE a correct NACK," separated from "can the system GENERATE one gold-free" (Brick-C, where the add-back sweep + tier-0 replace the oracle). If Brick-A ran on the weak deployable localization (blame recall 0.067) and failed, conditioning-vs-localization would be unattributable. 4. **The measurement (field, pre-specified):** same puzzle, same weights, blank vs NACK-conditioned input THROUGH THE SAME GRAPH; read = does the parse delta (emission-change mass + attention-delta mass) CONCENTRATE on flagged slots/sentences vs a global reshuffle. Plus: flagged-slot FIX RATE vs a SHUFFLED-FLAGS control (same flag count, innocent sentences). **KILL: fix_rate(true) ~= fix_rate(shuffled) => conditioning isn't doing the work => the null fails toward the LoRA ladder.** - **Brick-A SUBSTRATE BLOCKAGE (2026-07-06 night, recorded honestly):** every backward-through-trunk training form hangs the AM driver — fused (real), eager multi-loss (real), three-JIT checkpointed VJP (real, on a RECOVERED device; jit1's input-grad extension through the cross-attention is a NEW graph class, not the hours-proven head step). Pre-wedge facts: trunk-bwd alone OK, JIT'd head-step OK. Meta-lesson (quirks): HANGS ACCUMULATE — recover (scripts/am_gpu_recover.py, self-service via caps) before trusting any bisect result. DECISION: run the HEAD-LEVEL arm first (registered §9.1 comparison arm; all-proven graph classes; banked states apply) — it measures fix-rate/concentration/kill TODAY. Trunk-level remains the registered honest null, PARKED behind the driver fight (smaller B/T, tinygrad update, or upstream fix are the candidate unlocks). - **BRICK-A HEAD-LEVEL ARM RESULT (2026-07-06): NULL FAILS at this arm — honest negative with TWO NAMED CONFOUNDS.** fix(true)=0.102 vs fix(shuffled)=0.083 (barely separated), concentration 0.47/0.54 (BOTH < 1: parse delta lands AWAY from flagged regions), preservation ~0.997 both, solves 7/8/6 (noise). Per the registered kill: this arm's conditioning is not doing the work. CONFOUNDS: (i) HEAD-LEVEL CANNOT CHANGE THE READING — trunk states are fixed; if the error lives in the states (not the aggregation), only trunk-level conditioning can fix it. The head-level failure is CONSISTENT WITH the trunk-level thesis, not a refutation of it — and notably falsifies the "head-level trivially works" prior. (ii) THE TRAINING OBJECTIVE MADE FLAGS REDUNDANT: v0 supervised ALL slots toward gold regardless of flags, so the model could improve flagged slots from gold alone — conditioning was never made LOAD-BEARING. A proper design forces flag-dependence (e.g., reproduce-previous-output when unflagged, emit-the-fix when flagged). ALSO: plateau-residual errors (survivors of 68k supervised steps) may be capacity errors no same-model re-parse can fix — the NACK fix-rate ceiling for an UNCHANGED model is itself an open measurement. NEXT: (a) flag-dependent training objective, (b) trunk-level arm when the driver fight is won, (c) if both fail: the §8.4 LoRA ladder, as specified. - **TEXT-RENDERED NACK (registered 2026-07-06 — the driver-fight dissolver):** the trunk is FROZEN, so backward-through-trunk is only needed if conditioning enters as LEARNED prefix params. Render the NACK as literal text instead ("NOTE: statement 7 may be wrong.") prepended to the problem: ordinary forward, head trains exactly as today (backward stops at the head input — the 68k-step-proven graph class), zero new gradient paths, zero driver exposure. This IS trunk-level conditioning (the flags change how the text is READ — what confound #1 demands) and arguably the PUREST plane-ride null: same weights, different INPUT, different behavior — input in the most literal sense. Cost: conditioned trunk states need live forwards / a small conditioned precompute (~6 min for the Brick-A corpus at 0.34s/batch). Token budget: note ~15-40 tokens, max 451+40 < 512, the truncation guard stands watch. COUNTER-PRIOR (register it): C1-A found TEACHING (auxiliary prediction) beat TELLING (hint concatenation), and text-rendered NACK is telling — but runtime verifier flags MUST enter as input somehow (no training-time auxiliary exists for an error that doesn't exist until inference), so C1-A predicts a MODEST effect size, not a wrong channel. Queue: flag-dependent head arm (running) -> text-rendered trunk arm -> LoRA ladder only if both fail. Plus the CEILING PROBE as its own measurement: deliberately overfit an unrestricted head on flagged positions from frozen states — slots where gold is NOT decodable are provably beyond ANY head-level conditioning; that fraction converts the capacity-error caveat into a DENOMINATOR (fix rates then read as fraction-of-the-fixable). - **BRICK-A FLAG-DEPENDENT ARM RESULT (2026-07-06): FLAG-USE PROVEN; localization real at the DECISION level; my registered field metric was MIS-SPECIFIED.** Numbers: fix(true)=0.438 (152/347) vs fix(shuffled)=0.360 — the flag-dependent objective lifted fix capability 4.3x over v0 (0.102->0.438), and the true-vs- shuffled gap (~2.9 sigma) is what LOCALIZATION buys; the 0.360 shuffled floor is what the shared GLOBAL-FAIL bit buys (generic revision energy). Preservation 0.998 both. SOLVES: blank 0 -> true-NACK 7 (12.3% of failures solved outright by ONE oracle-NACK round) vs shuffled 4. Blank solves dropped to 0 BY DESIGN (blank-mix trains copy-previous, errors included — fix energy activates only with the bit). THE METRIC LESSON: registered concentration (L1 prob-mass) reads 0.52<1 — but it is dominated by sub-threshold jitter; at the DECISION level, flagged slots flip at 43.8% vs unflagged at ~0.2% (~200:1). The prob-mass field was the wrong functional form; reported as registered (fails the letter), decision-level re-read passes the spirit. Named, not swapped silently. C1-A's modest-effect prediction for "telling" holds: real, significant, not crushing. VERDICT: head-level conditioning DOES work when made load-bearing — confound #2 was the v0 story. The remaining gap to 1.0 awaits the CEILING PROBE (denominator) and the TEXT-RENDERED trunk arm (reading-repair on top of flag-use). - **POST-ARM REGISTRATIONS (2026-07-06, before the next dataset):** (1) LOCALIZATION METRIC REPLACED, pre-registered: flagged-vs-unflagged FLIP-RATE RATIO at the decision level (the L1 prob-mass form aggregated sub-threshold jitter; promoted to CLAUDE.md §6 as the third sighting of the instrument lesson — the registered functional form must match the mechanism's decision structure). (2) CEILING-PROBE CAUTION: measure decodability from frozen TRUNK states (head-independent — a fresh unrestricted probe head, deliberately overfit), NOT from the current head's representations, which inherit the objective they were trained under. The probe denominates the INJECTION POINT, not the head. (3) BRICK-C'S BUDGET, handed by the decomposition: the shuffled floor (0.360) is what ANY global "something's wrong" signal buys; add-back blame quality only has to defend the +7.8pt LOCALIZATION increment. Imperfect blame degrades the increment, not the floor — the loop is robust to sloppy localization. (4) THE MARGINAL-SOLVES FRAMING: blank 0 -> true 7 is the loop's contribution OVER DOING NOTHING — seven puzzles unreachable by unconditioned same-weights re-parsing, solved in one SYN->NACK->retransmit round (oracle standing in for Brick-C). The core Alternator transaction has now executed end-to-end. - **CEILING PROBE v1 INVALIDATED BY ITS OWN BASELINE (2026-07-07):** fresh-init 6k steps read decodable=0.095 — BELOW the head arm's constructive 0.438. A ceiling cannot sit under a measured achievement: v1 measured OPTIMIZATION BUDGET, not decodability (loss 4.3 vs plateau 3.4; the original head needed 20k+ steps + the full corpus). Standing facts: decodability of test flagged slots >= 0.438 (constructive, from the head arm). v2 = partial warm-start from the plateaued gold-only head (pre-Brick-A, objective-independent) + fresh wide FFN + 12k steps. Lesson filed: an overfit-probe must DEMONSTRATE convergence before its residual can be called a ceiling. - **TEXT-NACK ARM RESULT (2026-07-07): FAILS — content-blind, diagnostically so.** fix(true) = fix(shuffled) = 0.295 EXACTLY (142/482 both): the note's CONTENT was never read — only its presence registered (generic revision). Flip-rate 0.921 vs 0.666 (~1.4:1 — global thrash, vs the head arm's 200:1); preserve degraded 0.998 -> 0.92 (the note shifts every token's position + steals attention); solves 1/0/0 (WORSE than blank); training loss never settled (3.7-4.0 vs 3.3). The drop-bias caveat does not rescue it: identical true/shuffled is content-blindness, independent of training distribution. DIAGNOSIS: referential binding ("statement 7" -> the 7th sentence) is deep-layer work; a FROZEN 4-LAYER prefix of a 1B LLM cannot compute it. The C1-A counter-prior ("telling is modest") was right and then some: telling through a shallow frozen reader is ~zero. The placebo arm is MOOT (content-blindness established at true-vs-shuffled; skipped, reason recorded). - **CEILING PROBE v2 (2026-07-07): CONVERGED, VALID — decodable = 0.533.** Warm- started, 12k steps, loss 3.47 ~= plateau (converged); ceiling ABOVE the constructive 0.438 (consistent, unlike v1). **46.7% of plateau errors are beyond ANY head-level channel; the head arm's 0.438 = 82% OF THE FIXABLE.** - **BRICK-A VERDICT (the plane-ride hypothesis, measured):** the zero-LoRA null PASSES — same frozen weights, different INPUT, appropriately different behavior — but the interface matters decisively: conditioning works as STRUCTURED FEATURES (position-aligned embeddings; 82% of fixable, 200:1 localization) and fails as NATURAL LANGUAGE through the shallow frozen trunk. This VINDICATES the original §8.2 interface design (notebook/NACK as feature channels, registry vocabulary) over the text-rendering shortcut. The remaining 46.7% needs actual reading-repair — candidates: deeper trunk prefix, the LoRA ladder, or the Alternator's real answer (structured ledger re-parse). Brick-C operates within the 53.3%. - **POST-BRICK-A REGISTRATIONS (2026-07-07):** (1) §6 law added: positional/referential structure enters AS STRUCTURE, not prose (2 sightings: sentence-counting, reference- binding). (2) FRONTIER RANKING for the 46.7%: LEDGER-CONDITIONED RE-PARSE FIRST — it changes the PROBLEM (parse against pinned verified context), not the injection point, and may move the 0.533 denominator itself; zero-LoRA-compatible; free once the loop exists. Deeper prefix second; LoRA ladder stays behind measured alternatives. (3) BRICK-C KILL CRITERION, pre-registered RELATIVE: add-back blame must RETAIN >= HALF the oracle's localization increment (the +7.8pt over the 0.36 floor) — the oracle arm is the ceiling condition; Brick-C's question is how much survives the real instrument. (4) JOB-B GATE IS DUE (pegged to Brick-A completion). - **UNKNOWNS-CORPUS DECISION (registered 2026-07-07, product track):** GENERATED ALGEBRA before any MATH-500 graph-ification — linear systems + small nonlinear compositions, NL via template families. The KenKen generator's deepest property transplants verbatim: the domain makes gold + equivalence checking FREE (unique-solution-by-construction, exact symbolic solve, round-trip gating), and decisions-per-problem becomes a GENERATOR DIAL (target the engine band directly). MATH-500 stays the EVALUATION target, never the training substrate. SEQUENCING: the registry extension (equals/linear relations) is the first NEURAL-side test of "new domain = predicate + bridge, never core" — the delta head's typed-slot + digit-spaced format meets its extensibility bill EARLY, during corpus work, not after Brick-C hardens KenKen-only assumptions. - **BRICK-C MEASURED (2026-07-07): THE LOOP CLOSES GOLD-FREE.** v0 KILLED at its pre-registered bar (retention 0.48 < 0.5) — the kill forced the instrument reading that surfaced a bug-vs-intent (malformed cages commented as blamed, never blamed). v0.1 (the labeled fix): fix oracle 0.438 | ADD-BACK 0.401 | shuffled 0.360 -> **retention 0.52, LIVES** — with the margin stated honestly: 0.48/0.52 straddle the bar, so the true retention sits AT ~0.5 within noise. The robust facts: the gold-free instrument (confidence-ordered add-back + attention slot->sentence map) recovers ~half the oracle's localization increment, and SOLVES — the end-to-end metric — ran 8 vs oracle 7 vs shuffled 4 in BOTH versions: full solve parity with the ceiling condition. THE ALTERNATOR LOOP HAS NOW RUN GOLD-FREE END TO END: parse -> symbolic self-diagnosis -> blame -> flags -> conditioned retransmit -> solve, no gold at any stage, 8/57 plateau failures recovered in one round. Margin upgrades (named, unbuilt): uniqueness-probe flags in the sweep, a real tier-0 confidence head for ordering, multi-round retransmission. ## 10. The math expansion (begun 2026-07-07) - **REGISTRY EXTENSION, SYMBOLIC HALF: DONE.** `arith3` — op(a,b)=r with unknowns on BOTH sides, L-ASYM ordered scope (a,b,r), integer-exact (div = exact-divisibility), constants as given variables (singleton domains, never factors). Predicate + pairwise-support propagator + bridge, ZERO csp_core edits — the 7th domain through the same seam. Smoke (hand systems, all asserted): triangular -> 0 decisions (calculator band); COUPLED (x+y=5, x-y=1) -> 1 decision (**engine band by construction** — AC keeps locally-supported values only joint reasoning kills); 4-pair chain -> 4 decisions (**the generator DIAL measured: coupling count = engine- band depth**); uniqueness checkable by ban-and-resolve (gold + equivalence stay FREE — the KenKen property transplanted). - **NEURAL-FORMAT FINDING (by analysis, ahead of any build — the answer the brick existed to surface):** the delta head's emission extends to arith3 EXCEPT the membership pointer: multi-hot over positions is ROLE-BLIND, and op(a,b)=r is order-sensitive for sub/div — the (a,b,r) roles cannot be recovered from a set. SMALLEST SURGERY, two options to decide at build time: (i) role-typed pointers (args-multihot + a separate RESULT pointer head — canonicalizing sub/div into add/mul form makes the two args genuinely unordered, so 2 heads suffice), or (ii) three categorical pointers (a, b, r). Everything else transplants: type menu extends by one entry, op vocab UNCHANGED (add/sub/mul/div already there), targets unused (constants are variables now — the digit heads idle or emit given-values). The typed-slot format's extensibility bill: ONE new pointer head, not surgery. - **RELATED WORK (verified 2026-07-07, published 07-06): Anthropic's "Verbalizable Representations Form a Global Workspace in Language Models"** (transformer-circuits + open-source J-lens). VERIFIED claims: an emergent small workspace ("a few dozen concepts, <1/10 of activity") that is reportable/controllable/load-bearing for deliberate reasoning; ablation kills multi-step reasoning while fluency/recall survive; unspoken intermediate concepts causally mediate answers (spider->8, swap ant->6); workspace monitoring surfaces hidden error/deception signals pre-output. CONVERGENT MOTIFS with Mycelium (analogies, NOT evidence — the over-claim guard applies): small-broadcast-channel over parallel substrate (waist/notebook/global latents, built explicitly); reasoning-in-residual-never-tokens (§6 law, observed in the wild); internal-state-more-honest-than-output (their safety program == our tier-3 late-JSD 0.687). The double edge, registered: emergence-at-scale cuts against NEEDING to build the workspace explicitly; our counter is the thesis itself — at 32M-87M params every joint is separately measurable. Closest neighboring result for the eventual paper's related work. - **BORROWED TECHNIQUE CANDIDATE — J-ORDERED MATRYOSHKA:** order waist dims by JACOBIAN SENSITIVITY of downstream decisions (|d output / d dim|), not variance. We already hold the measured motivation: variance ordering inherits the identity confound (the 0.755-vs-0.658 trap). Both waists are differentiable on existing ckpts — one backward per output per batch. The experiment: J-ordered prefix-width curve vs the current ordering; if the 128d survival sharpens further (or the cliff moves), sensitivity ordering wins the schedule. - **J-ORDERED MATRYOSHKA, DESIGN REFINEMENTS (registered before build):** (1) THE DISCRIMINATING REGION IS BELOW 128 — the measured curve is FLAT 128~=512, so both orderings read "fine" there; sweep 8/16/32/64/128/256/512 and REGISTER THE PREDICTION AS "J-ordering moves the CLIFF LEFT" (survival-at-128 is unfalsifiable on a saturated curve). (2) ESTIMATOR = DIAGONAL FISHER: mean of SQUARED per-instance gradients (signed averaging cancels opposing sensitivities and ranks live dims dead); per-sample backwards to avoid within-batch cancellation. (3) TWO TARGETS: decision-side (solve-relevant logit margins — the honest one; wrong-but-equivalent already proved loss and decision diverge) + loss-side as the cheap comparison arm (identical rankings would itself be a finding). (4) SCOPE OF A NULL: greedy diagonal ranking is not the optimal SUBSET per width (dims can be individually weak, jointly load-bearing — the codebook's pairwise geometry is exactly what a diagonal misses); a null reads "diagonal sensitivity doesn't beat variance," NOT "sensitivity ordering wrong in principle." (5) THE INCUMBENT IS NOT A STRAWMAN: Matryoshka training kept dims 0-127 always on — identity ordering carries the baked-in trained importance; Fisher matching it CONFIRMS the nested training, beating it below 128 is the win. - **J-ORDERED MATRYOSHKA RESULT (2026-07-07): PREDICTION REFUTED — the INCUMBENT WINS, and the refutation is structural.** Cliffs: identity (trained prefix) 128; variance/fisher_loss/fisher_decision all 256 — every post-hoc reordering moved the cliff RIGHT. The registration's item 5 named the mechanism in advance: nested training didn't just RANK dims, it built CO-ADAPTED FUNCTIONAL SUBSETS (dims 0-127 were trained to work AS A SET; a mask of the 128 most-sensitive dims scattered across the index space was never a training configuration, and the head cannot read it — despite 85-91% top-128 overlap with the incumbent, the 10-15% disagreement costs the whole cliff). SUB-PREDICTION CONFIRMED: decision-Fisher beats loss-Fisher at every width below 256 (0.120/0.139/0.255/0.627 vs 0.018/0.001/0.009/0.496) — the wrong-but-equivalent divergence is real in the Jacobian too. BELOW THE TRAINED FLOOR (<128, never sampled in training): the trained prefix DIES COMPLETELY (0.001-0.011) while variance/decision-Fisher degrade gracefully (0.306/0.255 at 64) — sensitivity orderings find dims with standalone signal; none is solve-capable. §8.6 CONSEQUENCE: the waist schedule's aggression limit is a TRAINING-TIME choice, not a post-hoc reordering — to survive at 64/32, SAMPLE those widths during nested training (optionally SEEDING the dim order by decision-Fisher before training: candidate, unregistered). NULL SCOPE honored: this reads "post-hoc diagonal reordering doesn't beat a trained incumbent," not "sensitivity ordering wrong in principle." - **POST-REFUTATION REGISTRATIONS (2026-07-07):** (1) composition-fragility promoted to §6 (third sighting). (2) FISHER-SEEDING registered WITH mechanism: seeding cannot help through ordering per se (training co-adapts whatever the mask exposes); its value proposition is BETTER RAW MATERIAL (dims with standalone signal — the thing finding #3 proved sensitivity tracks) for narrow sets to co-adapt from. Clean test: seeded vs unseeded nested training WITH width-64 in the schedule. HONEST MODEST PREDICTION: if co-adaptation dominates raw material they converge and the seed buys nothing — completing the story as "the schedule is everything, the ordering is nothing." (3) DESIGN RULE, forward-binding: everything that READS the silhouette (perceiver monitor, Brick-0 probes, tier-3 instruments) is built against the DECISION-relevant subspace, not the loss-relevant one — an order of magnitude at narrow widths is not a nuance. (4) Related-work line beyond the citation: the nested-trained incumbent is the ADVERSARIAL CASE for post-hoc dim selection — "co-adaptation defeats post-hoc sensitivity ranking" is a boundary condition of the J-lens found within 24h of its publication. - **CORPUS REGISTRATIONS (2026-07-07, per-band + mentions):** (1) EVAL IS PER-BAND from the first run: solve-rate and factor-exactness logged PER decisions-band (a smeared solve-rate would erase the one distinction Job-B established matters); curriculum-by-band available (calculator band first: parse errors are the only failure mode there). (2) THE FACTORIZATION QUESTION, answered free by that logging: is parse difficulty CORRELATED with solve difficulty? A flat parser error rate across bands = reading hardness and reasoning hardness on independent axes — the cleanest possible vindication of the two-phase design. (3) VARIABLE-MENTION ANNOTATIONS added to gold (generator emits every name occurrence as char spans, free): the registered pre-emption of referential binding — the text-NACK arm proved shallow layers don't bind references unaided; the result pointer gets name->slot binding AS STRUCTURE (§6 law applied prospectively), not as a hop to learn. Cheap now; expensive after a plateau gets misdiagnosed as capacity. ## 11. The algebra chapter's interpretive frame (registered 2026-07-07, BEFORE the head trains) - **THE DETECTABILITY INVERSION (on schedule, not a regression):** KenKen's seven- point 100%-detectable streak was a gift of constraint DENSITY. Linear systems are satisfiable for almost any constants: a mis-parsed literal yields a SAT, UNIQUE, cleanly-solving-but-WRONG system — UNSAT doesn't fire, the uniqueness probe doesn't fire. Surviving symbolic channels are STRUCTURAL: dangling variables / underdetermination (multi-solution probe) and overdetermined-inconsistency (UNSAT). **Do NOT grade the first algebra loop against KenKen's 100%.** - **REGISTERED PREDICTION #2-ALGEBRA (the KenKen limbo resolves here):** the silent class appears at SUBSTANTIAL rate, CONCENTRATED IN LITERAL-CONSTANT ERRORS, while membership/structural errors stay detectable. REFINEMENT from the corpus's own integer structure: coupled-pair constants are ~HALF-caught by integrality/parity (x+y=11, x-y=4 -> x=7.5 -> UNSAT over Z — a wrong constant flips parity ~50% of the time), while CHAIN constants go silent by default (a shifted k just shifts downstream values; SAT, unique, wrong). So predicted silent-rate ordering: chain-literals >> coupled-literals > structural (~0). The taxonomy tool runs unchanged; the corpus's gold answers grade it. - **THE NACK STACK'S COMPOSITION INVERTS:** tiers 1-2 (verifier, uniqueness) carried KenKen; algebra hands the load to TIER-0 (the confidence head — now with its motivating error class: low-confidence literal emissions) and TIER-3 (the deducer's soft-solve suspicion field — specced for exactly this soft-wrongness regime). Neither is built. The algebra chapter is not "new domain, same loop" — it is the domain that FORCES the two neural NACK tiers the design promised. - **HEAD FIELD LAYOUT (settled before build):** slots emit presence | type {rel, given} | op {add, mul} (canonical; sub/div die at the generator) | args = 2-hot over VARIABLE slots | RESULT = UNION TYPE: an is-literal MODE BIT gating result-POINTER (categorical over variable slots) vs the digit machinery (transplants verbatim) — both sides supervised from gold; given-factors use var-pointer + value digits. Variable slots anchor to the text through the MENTION spans (name->slot binding as structure, §6). PLUS the head nobody mentioned until now: the QUERY POINTER — one global supervised pointer over variable slots (gold free from the generator); without it the pipeline solves the system but cannot answer the question. ## 12. Tier-0 confidence (designed 2026-07-07, cross-channel; v0 = the incumbent, formalized) - **THE FRAME (post-J-lens):** the entropy null is the INCUMBENT WITH A TRACK RECORD (slot_confidence — presence-sigmoid x per-field top-prob product, zero params — ordered the add-back sweep to 0.52 retention). A trained head's bar is not "does supervised confidence work" but "does it beat a working incumbent." - **GRANULARITY:** per-field confidence is the primitive (a READ off the union layout's separate logits, not a build); factor-level is DERIVED. Open sub-question, a numpy afternoon: does per-field error correlate within-factor? Yes -> min() or a learned 4-weight combination beats the product; no -> the product stands. - **CALIBRATION vs DISCRIMINATION, split by consumer:** the add-back sweep consumes RANKING (metric: AUC — implicit in the 0.52, never read directly; MEASURE IT FIRST); phantom-gating and literal-flagging consume THRESHOLDS (metric: ECE, with the incumbent TEMPERATURE-SCALED first — one scalar fitted on train failures). Registered structure: incumbent -> per-field AUC + post-temperature ECE on banked artifacts -> a trained head must beat BOTH on the field where it claims value. - **THE WITHHOLDING-COST CURVE (pre-registered):** decode each banked KenKen test failure; withhold the k least-confident factors; solve the remainder; sweep k in 0..5. COLUMNS: (a) solve-to-GOLD rate — Code's addition: withholding a wrong factor from a dense graph can recover the EXACT grid with NO retransmission, so the curve doubles as the measurement of withhold-and-solve as a THIRD repair channel (vs retransmit, vs oracle); (b) taxonomy composition of the withheld graphs — UNSAT -> multi-solution conversions are DETECTION conversions, not just solve effects. REGISTERED PREDICTION (conditional — Code's pushback): the curve is non-monotone with a peak at k=1-2 on KenKen, CONDITIONAL ON incumbent per-factor AUC >= ~0.7 — at AUC ~0.65, k=1 withholds a CORRECT factor ~35-40% of the time and converts right-graphs to underdetermined, and the peak may never materialize. AUC lands first; the curve is interpreted against it, not alone. ALGEBRA FLIP (registered): sparse coupled systems starve immediately (withholding a given -> underdetermined by construction) — peak at k=0-1 or absent. DENSITY DECIDES THE GATING BUDGET: the cross-domain claim. - **THE NULL'S BLIND SPOT (the one principled edge for a trained head):** entropy measures decision SHARPNESS; confidently-wrong is structurally invisible to it — and plateau errors are exactly the sure-and-wrong class (the 6-7 codebook confusions are the deducer-side picture). Registered alternative form: does supervised correctness-prediction find errors ON THE HIGH-INCUMBENT-CONFIDENCE SUBSET? Zero edge there -> the null survives everywhere it matters. - **PRE-REGISTERED KILL for any trained tier-0 head:** must beat the calibrated incumbent's AUC on the high-confidence-error subset, or per-field ECE, by margins set AFTER the incumbent's numbers land. Zero GPU for all of v0 — measurement, not training. - **FIRST ALGEBRA RUN (2026-07-07 evening): 58/60 ANSWER, and the FACTORIZATION READ IS CLEAN.** Per-band: fac-exact 1.000/1.000/0.997/0.993 across decisions bands 0/1/2/3 — FLAT. Parse difficulty and solve difficulty are on INDEPENDENT AXES on first measured contact: the parser reads band-3 (4-decision) systems as easily as calculator-band ones, while the solve column does genuinely different work per band. The two-phase design's division of labor, vindicated by its first stratified table. ANSWER 58/60 = the full pipeline (parse -> graph -> GENUINE SEARCH (1-3 decisions) -> query-pointer -> answer) at near-ceiling. Noted oddity: band-2 ANSWER (29) > graph-solve (28) — a wrong-somewhere graph can still be right AT THE QUERY: wrong-but-right-where-asked, the equivalence class's algebra cousin. HONEST DEFLATOR: v0 templates are easy (105 tokens mean, low paraphrase diversity, mentions given as structure — the §6 laws compounding prospectively designed OUT the failure modes). Near-ceiling here means THE CORPUS NEEDS TEETH, not that parsing is solved: crank paraphrase depth + template families until failures exist, THEN prediction #2-algebra becomes testable (2 failures is an anecdote, not a taxonomy). The chapter's real tests still ahead; its plumbing is proven. - **THE CONVERGENCE EVAL (2026-07-07 night, teeth 0.8, n=300): PREDICTION #2-ALGEBRA RESOLVES BOTH WAYS.** The teeth bit: fac-exact 0.69-0.81 (v0 was ~1.0), ANSWER 121/300, a real failure population (179). **THE SILENT CLASS IS BORN — the inversion arrived as registered:** 14 SILENT (KenKen: zero across seven points); detectable fraction 0.92. **BUT THE ORDERING IS REFUTED:** silent literals are roughly uniform across roles (pair_diff 4 / pair_sum 2 / chain_k 2), and wrong literals overwhelmingly land DETECTABLE (unsat 195, multi 89 attributions) — far more caught than predicted. THE MECHANISM OF THE REFUTATION IS THE FAMILIAR ONE: the prediction was calibrated to the SINGLE-ERROR regime; at fac-exact ~0.73 parses carry ~5 wrong factors, and multiple wrong literals + structure JOINTLY over-constrain -> UNSAT catches what parity alone would have missed. (Delete-one blame died the same death — "single-error-regime assumptions break at multi-error density" is now a RECURRING prediction-failure mode; second sighting, watch for the third.) DETECT_multi=43: the uniqueness probe earns its keep at scale. ANSWER(121) > graph-solve(95): wrong-but-right-where-asked is now a LARGE class (26), not a curiosity. FACTORIZATION UNDER LOAD, honest read: fac-exact no longer flat (0.811 -> 0.685 by band 3) — but band correlates with SIZE (more pairs = more factors/vars/obliques), so the axes may correlate through LENGTH, not through solve-difficulty; the size-controlled read (fac-exact vs band at fixed n_vars) is the open analysis before the v0 flat claim is downgraded. TRAINING NOT CONVERGED (loss 5.3, falling): all numbers are floors. NET FOR THE LOOP: 92% detectable on algebra's thinner jaws — the NACK story survives the inversion far better than feared, and tier-0/tier-3's target class (the 14 silents + 43 multis) now exists. - **TIER-0 MEASUREMENT 1 (2026-07-07 nightcap): THE INCUMBENT HALF-SEES THE SILENTS, AND ITS BLIND SPOT IS FIELD-LOCALIZED.** Silents-vs-correct AUC (the registered separation number, n=14 caveat): product 0.727, min 0.680, digits 0.654, fields 0.601 — but **pointer 0.479 and query 0.398: AT-OR-BELOW CHANCE. Silent parses are AS-or-MORE confident in their pointers/query than correct parses are** — the entropy null's predicted blind spot (confidently-wrong), found alive and localized to exactly two fields. Meanwhile all-wrong-vs-correct product AUC = **0.905**: the incumbent is an EXCELLENT ranker of detectable errors (Brick-C's consumer is well served). Saturation note: median confidences 0.97-1.00 everywhere — thresholds are hopeless without recalibration; the AUCs are rank-based and unaffected. VERDICT (per the §12 frame): the null SURVIVES as the ordering signal and DIES as the silent-detector; the trained head's job is now precisely scoped — supervised correctness-prediction on the POINTER and QUERY fields, where entropy carries zero signal and input-conditional error patterns are the only hope. Neither clean ending; the design's completion now has an address. - **FIXED-n_vars FACTORIZATION READ (2026-07-07 morning): THE AXES-INDEPENDENCE CLAIM SURVIVES UNDER LOAD.** Raw corr(fac-exact, band) = -0.126 was ENTIRELY size-mediated (corr(band, n_vars)=+0.678, corr(fac-exact, n_vars)=-0.249): **PARTIAL corr(fac-exact, band | n_vars) = +0.061 ~= ZERO**, and within fixed-size buckets fac-exact is flat-or-RISING with band (12-15 vars: 0.665/0.676/0.732). The parser feels TEXT LENGTH, not solve difficulty — reading and reasoning on independent axes, now confirmed at teeth 0.8 with the confound removed. The v0 flat claim is UPGRADED, not asterisked. - **WITHHOLDING-COST CURVE (2026-07-07 morning, banked KenKen, n=57): THE THIRD REPAIR CHANNEL IS REAL — AND IT BEATS RETRANSMISSION.** Precondition first, per registration: incumbent per-factor AUC = **0.613 < 0.7** — the k=1-2 peak claim is formally VOIDED (not refuted). The curve: solve-to-GOLD 2/8/12/**15**/11/10 for k=0..5 — **non-monotone with the peak DISPLACED to k=3**, consistent with the precondition's own logic (imprecise ranking x ~5 errors/parse -> withhold more to cover the wrong set; peak position ~ errors-per-parse x ranking quality). QUALITATIVE SHAPE CONFIRMED, quantitative form shifted exactly as the conditional anticipated. THE HEADLINE: **15/57 (26%) of plateau failures solve to the EXACT gold grid by deleting the 3 least-confident factors — no NACK round, no retransmission — vs 8/57 for one-round retransmit.** Withhold-and-solve is the loop's zero-cost FIRST move on dense domains. SAFETY: wrong-grid = 0 at EVERY k — withholding never produced a silent-wrong solve; every non-gold outcome stayed detectable (UNSAT -> multi drift 1->23, the registered detection conversions). Dense domains keep the full safety net through withholding. (Algebra flip still registered: sparse systems should starve immediately — measure, don't assume.) COMPOSITION IMPLIED: withhold-and-solve -> if multi -> retransmit — the repair stack now has an ordering by cost. - **THE BIG SLICE (2026-07-07 midday, n=1500 fresh, 78 silents): THREE RESOLUTIONS.** (1) **The inversion number is STABLE:** detectable 0.91 on the second draw (0.92 first) — ~0.9 is algebra's detectability constant at this parser quality. (2) **Prediction #2-algebra's ordering: fully INVERTED at real n** — silent literals: pair_sum 16 / pair_diff 13 / **chain_k 2** — chains almost never silent (the registered chain>>coupled claim was exactly backwards). Mechanism: multi-error co-occurrence — chain-literal errors ride along with structural breaks (304 in UNSAT), while pair literals sometimes stand alone with parity's coin-flip. (3) **LAST NIGHT'S BLIND SPOT WAS A SMALL-n ARTIFACT:** at n=78, pointer AUC 0.479 -> 0.660, query 0.398 -> 0.685 — the "at-or-below chance" fields carry normal signal; flag-don't-model vindicated again (14 positives lied). **THE HEADLINE: min-combination AUC = 0.812 — AT the registered completion threshold.** The §12 combination-rule sub-question is answered empirically: MIN beats PRODUCT on silents (0.812 vs 0.734 — within-factor errors correlate; the weakest field is the signal). Per the registered read (>~0.8 = the stack completes with zero new params): **tier-0 v0 = the min-combined, temperature-calibrated entropy incumbent. The trained head's kill criterion is now "beat 0.812 at n=78" — it is NOT BUILT unless a future domain reopens the gap.** Cautions: 0.812 at n=78 is at-the-bar, not past it (CI ~±0.05); thresholding still needs calibration (medians 0.987/0.962); all-wrong product AUC 0.915 (the ranker stands). THE ZERO-LORA PHILOSOPHY'S FOURTH WIN THIS WEEK. - **POST-BIG-SLICE REGISTRATIONS (2026-07-07 midday):** (1) density-regime rule promoted to §6 (checklist form). (2) TRAINED-HEAD RE-ARM CONDITION: the head stays unbuilt at kill=0.812, RE-ARMED iff a future draw shows the pointer/query weakness DEEPENING on harder obliques (the teeth dial can force that question when it matters — they remain the weakest senses at real n: 0.66/0.69, weak-not-blind). (3) Withhold-and-solve NAMED: Law 7 (remove at READ, not from state) at the factor-graph level — parse keeps everything, solve drops suspects, the lattice fills in; it beats retransmission ON A 0.613-AUC RANKING because the LATTICE does the repair (neural proposes, symbolic disposes, again). (4) **THE COMPOSED REPAIR STACK, registered:** confidence-order -> withhold-and-solve (k at the measured peak) -> retransmit ONLY the survivors -> answer; one pass; end-to-end recovery on the banked failures is THE chapter-thesis number (components: withhold 15/57, retransmit 8/57, overlap unmeasured). ALGEBRA arm doubles as the registered SPARSE-DOMAIN FLIP test — standing prediction: the withhold peak collapses toward k=0-1 (coupled systems starve; every equation load-bearing). Note: algebra has NO trained retransmission head (Brick-A was KenKen) — its stack is withhold-only, stated honestly. - **THE COMPOSED STACK, MEASURED (2026-07-07 midday):** KenKen: **24/57 = 42% end-to-end recovery** — withhold-3 stage-1 recovers 7 free, then retransmit(+ withhold-again) recovers 17 of the 50 survivors. Composition beats both components (withhold-alone 15, retransmit-alone 8) and nearly TRIPLES retransmission. HONEST CAVEAT: stage-1's 7 vs the standalone 15 — different base parse (the stack runs the Brick-A ckpt's blank pass, trained to copy-previous; the standalone curve ran the plateaued head). 24/57 is the one-weight-set production number. - **THE SPARSE-DOMAIN FLIP: REFUTED — withhold-and-solve GENERALIZES.** Algebra (n=913 failures, FORCED-answer check — no luck counted): recovery 0/60/**77**/75/ 70/56 for k=0..5 — peak at k=2-3, NOT collapsed to k=0-1. 8.4% of failures yield their answer FREE at k=2 on a SPARSE domain. The starvation intuition was single-error thinking AGAIN: at ~5 errors/parse, withholding preferentially hits WRONG factors (even a 0.613-AUC ranking beats chance) and the query needs only ITS component forced — poison-removal beats starvation until k grows. **FOURTH sighting of the density-regime trap, the same morning the rule was promoted.** UNSAT drains 659->127 with k as underdetermination grows — the registered drift. CONSEQUENCE: stage-1 of the repair stack is DOMAIN-GENERAL; the Alternator's cheapest move works on both dense and sparse graphs. - **POST-STACK REGISTRATIONS (2026-07-07 afternoon):** (1) density rule enforcement = ARITHMETIC ("what does the k-th withheld factor actually hit at measured density x AUC") — §6 updated; 4th sighting happened with the rule in hand. (2) WITHHOLDING'S BOUNDARY PRE-DRAWN: it works where errors outnumber the ranking's mistakes AND the query's component keeps support — sparse domains DECOMPOSE (starvation is local; poison in other components costs nothing); the clause that bites is a query inside one large coupled component. (3) THE BLANK-PASS TAX: Brick-A's copy-previous training degraded unconditioned parsing (stack stage-1: 7 vs standalone 15) — fix = TWO-CHECKPOINT stack (plateaued head PARSES, conditioned head RETRANSMITS); the 42% has known headroom before the parser improves. (4) ALGEBRA RETRANSMIT TRANSPLANT = a training run (flag-dependent objective on algebra failures; flags via MENTION/FACTOR spans — char-exact, richer than KenKen's attention map); built two-checkpoint from day one. (5) MULTI-ROUND'S QUANTITATIVE FRAME, registered (density-regime stated: multi-error, survivor-selected): per-round recovery DECLINES; the asymptote is bounded by the decodable ceiling's share (0.533) of remaining errors. A round-2 recovery ABOVE that frame = retransmission moving decodability itself — the ledger-conditioning hypothesis arriving early, and the cheapest probe of the 46.7% frontier. - **TWO-CHECKPOINT STACK (2026-07-07 afternoon): 27/57 = 47%.** The blank-pass tax eliminated exactly as registered: stage-1 recovers the full standalone 15 (the plateaued head parses; copy-previous training no longer degrades the first parse), stage-2 retransmits 42 survivors and recovers 12. Composed recovery 42% -> 47% with zero new training — the known headroom collected. Division of roles now explicit in the stack: the PARSER checkpoint parses, the RETRANSMITTER checkpoint repairs; one trunk, two heads, each at its trained operating point. - **THE ALGEBRA TRANSPLANT (2026-07-07 evening): 222/920 = 24% composed recovery.** Stage-1 withhold-2: 77 (the known 8.4% floor); stage-2 retransmit(+withhold) on 843 survivors: **145 recovered (17% of survivors)** — the repair specialist nearly TRIPLES the stack on its first run (77 -> 222). Trained pure (no blank mix; the two-checkpoint architecture carried blank quality), dual-granularity flags live from birth. Cross-domain: KenKen 47%, algebra 24% — consistent with the weaker parser (0.72 vs 0.83) and thinner constraint surface. HONESTY CAVEAT (recorded, not hidden): the eval's SPAN-level suspect flags located suspects via GOLD factor spans (slot order makes this ~"the j-th statement", mildly oracle-ish); the FIELD-level channel is position-free and clean. Fully-deployable variant = spans from the model's own attention (the KenKen pattern); the dual-granularity ablation doubles as the leakage bound — if field-only ~= both, the caveat is moot. QUEUE: parser convergence (teeth headroom) -> multi-round on honest numbers, per the sequencing registration. ## 13. The Nazaré funnel: the silhouette library + Brick-0 + the parse-side render (registered 2026-07-07 night, before build) - **THE PHYSICS, STATED HONESTLY:** the canyon adds no energy — it REFRACTS a wide front into convergence. The waist cannot amplify signal absent from the trunk; it FOCUSES it (Matryoshka already proved the canyon exists: the parse signal survives at 128d). The focusing law is the J-lens lesson: concentrate the DECISION-relevant subspace — a canyon focuses noise too. - **THE THREE GAPS (the ledger's unfired registered work):** Brick-0 has never run (assumption #2 — matched filters segment — is the LAST unmeasured plane-ride assumption); the centroid LIBRARY was never built as an artifact (the George-Hotz good-drivers note: nowhere on disk is there a bank of learned silhouette signatures to match against); the parse-side silhouette has capture hooks and has never been rendered (the BirdNet re-run, where the priors favor the bird). - **BUILD (one script, banked data):** (a) PARSE-SIDE LIBRARY: per-factor-type token centroids in the waist space, built on TRAIN, evaluated on TEST as per-token SEGMENT-AND-CLASSIFY (cosine-to-centroid argmax vs gold span labels) at widths 512 AND 128 (the canyon check: does the narrow waist classify as well?); plus the first token x type-similarity RENDER. (b) DEDUCE-SIDE ANALYTIC ARM of Brick-0: prototype matching on the banked 4-variant trajectory capture — constituent classification despite REFUTED linearity (the sharpened question: does matched filtering survive nonlinear composition?). (c) THE LIBRARY ARTIFACT saved to disk (.cache/silhouette_library_v0.npz) — centroids become a matchable bank, the registry's learned twin. - **BARS:** parse-side per-token classification must beat the majority-class floor by a wide margin to claim segmentation (priors favor it: gold spans exist, text is banded); Brick-0's full form (learned latents vs this analytic library) follows — if the analytic library already recovers constituents, assumption #2 completes WITHOUT learned latents (zero-param win #5); if not, the learned arm gets its bar. - **§13 RESULTS (2026-07-07 night): ASSUMPTION #2 COMPLETES IN EXISTENCE FORM — THE LAST PLANE-RIDE ASSUMPTION, MEASURED.** (a) PARSE-SIDE SEGMENT-AND-CLASSIFY: per-token kind accuracy **0.863** vs 0.779 majority floor (per-kind recall: rowcol 1.00, none 0.93, cage 0.86, given 0.82 — all four kinds genuinely separated, no majority collapse); **the canyon holds: 0.843 at width 128** (2 points for 4x narrowing). The render is legible — kinds band exactly over their gold spans (the BirdNet picture, parse-side, as the priors predicted). (b) BRICK-0 ANALYTIC ARM: deduce-side 4-way variant classification **0.854 vs 0.25 chance** — matched filters SURVIVE nonlinear composition at the classification level (linearity refuted, classification intact: the amplitudes don't add but the DIRECTIONS discriminate). (c) THE LIBRARY EXISTS: .cache/silhouette_library_v0.npz — parse-kind centroids at both widths + deduce-variant prototypes, the registry's learned twin, matchable on disk. STATUS OF THE THREE ASSUMPTIONS (§8.7): #1 conditioning suffices — TRUE via structured features (Brick-A); #3 the NACK — exists in measured tiers (taxonomy + 0.812 min-confidence + stacks); #2 matched filters segment — TRUE in analytic form (this, zero learned params: **zero-param win #5**). The learned-latent arm of Brick-0 is now an UPGRADE question with measured bars (0.863/0.843 parse, 0.854 deduce), not an existence question. - **POST-§13 REGISTRATIONS (2026-07-08 early, the rolling session):** (1) THE DIRECTION PRINCIPLE, noted for promotion watch: two independent structures now say "the signal is a DIRECTION, not a magnitude" (codebook angular separation predicts confusions; silhouette directions discriminate where amplitudes don't add) — second sighting; a third promotes it to §6. (2) THE 0.863 TEETH-ROBUSTNESS CHECK, registered before hardening: parse-side kind classification partially rides the text's own lexical banding (that WAS the prior); the discriminating check is the same centroids on a HARDENED slice (max paraphrase/split-ref/distractors) — holds = the centroids encode STEPS; sags = partially WORDS. (KenKen's teeth are milder than algebra's — stated; the algebra-side library needs waist exposure and comes next.) (3) **THE LIBRARY AS RUNTIME CROSS-CHECK, registered:** the delta head and the centroid bank are two INDEPENDENT readers of the same waist; disagreement (head emits cage, silhouette matches given) is a gold-free inconsistency signal at PARSE time — upstream of withhold-and-solve in the cost ordering, a candidate NACK tier at zero marginal cost. MEASURABLE NOW: per-slot disagreement-vs-wrong AUC on banked failures; the interesting case is AUC near tier-0's while DECORRELATED from it — the combined ranker then attacks the stack's measured weak link (the 0.613 withholding order). (4) THE LEDGER LINE: five zero-param wins; the perceiver's mandate shrinks with each (monitoring ~= calib+JSD; segmentation ~= analytic 0.854/0.863); its remaining candidate job is hosting the global-broadcast latents in the multi-cycle loop. The design gets LIGHTER as it gets more measured. - **ROLLING-SESSION RESULTS (2026-07-08 early):** (1) TEETH-ROBUSTNESS: the margin HOLDS — hardened slice 0.835 vs floor 0.754 (8.1 points; original 8.4). Within KenKen's teeth range, the centroids encode STEPS, not words; canyon holds hard (0.821 @128). (2) THE CROSS-CHECK TIER, honestly sized after an instrument audit: my first AUC(disagreement)=0.678 was a TIE-ORDER ARTIFACT (binary score, 96% ties, unstable argsort — midrank fix applied; the decision-structure metric law's cousin: DISCRETE SCORES NEED MIDRANK or the AUC is fiction). TRUE numbers: disagreement standalone 0.551 as a ranker BUT a rare high-precision flag (rate 3.2%, precision ~0.64 vs 0.176 base = 3.6x enrichment), decorrelated from tier-0 (+0.024), and the combined ranker = **0.634 vs the 0.613 baseline** — a real, modest upgrade to the stack's measured weak link at zero parameters. Verdict: the library-as-cross-check is a USEFUL RARE FLAG + a +2.1-point ordering gain, not a second tier-0. Both recorded; the artifact catch kept in the ledger (the measurement program audited itself twice in one night). - **MORNING REGISTRATIONS (2026-07-08):** (1) §12 ADDENDUM — THE PORTFOLIO RULE: NACK signals are TWO TYPES with different combination laws. DENSE RANKERS (tier-0 confidence, belief-JSD) blend; RARE-PRECISE FLAGS (uniqueness probe, library cross-check — ~3% fire rate, 3.6x enrichment, near-chance as rankers) VETO or ESCALATE. Future tiers get classified on arrival, not force-fit into the ranker mold. (2) The algebra-side teeth check (hardened big slice vs saved algebra centroids) is the lexical-shortcut SETTLER — queued for the next gap; requires algebra waist exposure. (3) **STACK-AT-CONVERGENCE, the protocol:** the hygienic redo (cosine LR decay, periodic val, PICK-BEST-BY-VAL on the small test slice — bigtest stays untouched as the measurement set) is the board's highest-leverage item AND re-arms half the registered predictions under NEW DENSITY conditions (fewer errors/failure -> delete-one blame may start working, the withhold peak slides left, the silent composition may rotate). All re-runs BATCHED as one table — same measurements, same protocols, ONE variable moved (the KenKen 10x discipline applied to the whole pipeline). The batched stack includes the 24%-ASTERISK resolution as an arm: stage-2 with FIELD-ONLY flags (position-free, fully deployable) vs span+field (gold-located) — the ablation and the deployable number in one read; the delta IS the leakage bound. - **STACK-AT-CONVERGENCE (2026-07-08): THE TABLE.** Hygiene worked: clean val rise to 0.783 @36k, best-picked (the spike never happened). Bigtest: **ANSWER 802/1500 (53.5%)** vs 587 pre-convergence (+215); graph-solve 680; fac-exact 0.71-0.87 per band. Taxonomy: detectable ROSE to **0.95** (silents 78 -> 34 — fewer and harder); silent attribution still mixed (5/2/1 — no chain dominance at low n). REGIME ROTATIONS, two: (1) **the combination rule is REGIME-DEPENDENT** — product now beats min on silents (0.793 vs 0.731; at the old regime min won 0.812 vs 0.734). Per-field structure keeps rotating across regimes (n=14 -> 78 -> 34): the ROBUST facts are portfolio-level (product/min 0.73-0.93), the field-level claims are regime-local. (2) **query-confidence became a SILENT-SPECIFIC flag: silents AUC 0.927 while all-wrong AUC 0.442** — a textbook portfolio-rule rare flag (classify on arrival: FLAG, not ranker). Withhold curve at convergence: peak k=3 (86/703 = 12.2% — MORE recovery than pre-convergence 8.4%); the peak did NOT slide left — the re-armed prediction is refuted-as-stated (surviving failures are the harder tail; density fell but selection hardened — the two effects cancel-ish). - **THE COMPOSED-STACK COLLAPSE, DIAGNOSED (the good-news failure):** stage-2 recovered only 13/627 — because NACK re-prep found **14/2000 train failures**: the converged parser nearly memorizes ITS OWN TRAIN SPLIT, so the repair specialist had no training data. The better the parser, the fewer organic training failures — **the repair trainer needs a failure-mining slice held out from BOTH the parser's training and the measurement set.** Fix: fresh generated slice (seed 21), prep there, retrain, re-run both arms. The field-only-vs-both ablation (13 vs 12) is DEGENERATE under the broken retransmitter — no conclusion drawn; re-read after the fix. - **CONVERGENCE-TABLE READS (registered 2026-07-08):** (1) THE SELF-DEFEATING CURRICULUM LAW: a repair specialist's training data must come from the CURRENT model's failures on FRESH data, or it trains on ghosts — the parser memorizing its train split starved the retransmitter. Recurs at every capability jump; the failure-mining slice is the permanent plumbing (the old generator-and-specialist- co-evolve principle rediscovered by necessity). (2) THE EQUILIBRIUM CLAIM, stated as one: failures get rarer AND harder in compensating proportion as the parser improves, so the stack's operating regime (errors/failure, withhold peak k=3) is more stable than either trend predicts. If it holds at the NEXT convergence jump, the stack's constants are regime-invariant — no re-tuning per parser generation. (3) COMBINATION RULES ARE REGIME-LOCAL (product-beats-min rotated); only portfolio-level structure is robust — re-measure the blend at every regime shift, never inherit it. Query-confidence's three-regime arc (fiction -> weak -> 0.927 rare flag) is the instrument ledger in miniature. - **THE FINAL ROWS (2026-07-08): THE ASTERISK RESOLVES — FIELD-ONLY WINS.** The curriculum fix worked exactly as diagnosed: prep on the fresh mining slice found **1,622/3,000 failures** (vs 14/2,000 on the memorized train split); the retrained specialist recovers **148/627 survivors** (was 13 when starved). COMPOSED AT CONVERGENCE: **224/703 = 32%** (rate ROSE from 24% pre-convergence on a harder survivor pool — the repair loop improves WITH the parser, consistent with the equilibrium claim). **ARM=field_only: 226/703 — the fully-deployable arm matches and slightly beats gold-located spans.** The informal 3-5-point prediction lands at 0.3 points on the favorable side; the leakage bound is ZERO; the span channel adds nothing — "this factor's field is suspect" carries the entire repair signal. STRUCTURE BEATS LOCALIZATION, again, and the algebra number is now the same kind of number as KenKen's 47%: gold-free end to end. **THE PIPELINE HEADLINE: one-shot 802 + repaired 226 = 1,028/1,500 = 68.5% of teeth-hardened algebra problems answered, fully deployable, through genuine search, at 3.2M trained parameters per head over a frozen 4-layer trunk.** - **POST-FINAL-ROWS REGISTRATIONS (2026-07-08):** (1) SPAN-LEVEL FLAG MACHINERY RETIRED, measured-unnecessary (the best deletion: bought by clean ablation). Mechanism note: the specialist repairs in the currency it emits (typed factor slots) — field flags arrive pre-translated; text spans need slot-binding, the operation this architecture thrice showed it does poorly through shallow layers and now once showed it doesn't need. (2) EQUILIBRIUM LEDGER: two independent favorable reads in one day (survivor-hardening cancels density drop; repair rate RISES 24->32% on harder pools). Prior moved; the real test stays scheduled at the next convergence jump. (3) THE GAP TO MATH-500 IS ENUMERABLE: relation coverage (registry menu), phrasing wildness (template teeth vs prose), problem-shape diversity — corpus-and-registry questions, not architecture, exactly where the plane-ride design claimed the difficulty would live. (4) **THE EQUIVALENCE UPGRADE, cut ahead of multi-round:** the table's rows are graded INCONSISTENTLY — composed rows require FORCED answers; the one-shot 802 does not (lucky-unforced uncounted-for). The fix: ONE uniform metric — answer-at-query FORCED (solution-set equivalence at the query variable) — applied to every row; plus the correctness- boundary taxonomy (right-where-asked-with-wrong-graph vs lucky-unforced). Zero GPU-heavy; changes the baseline every future number reads against; runs BEFORE multi-round so the asymptote frame starts from the true floor. - **THE EQUIVALENCE GRADE (2026-07-08, the table's uniform metric):** raw 802 -> lucky-unforced only **5** (the old metric was 99.4% honest) -> FORCED-CORRECT **797**. THE REAL FINDING: **132 of 797 (16.6%) are right-asked-wrong-graph** — a sixth of all correct answers come from graphs that DIFFER from gold factor-wise yet FORCE the right answer. The equivalence class is large: graph-match metrics undersell the parser heavily, and the MATH-500 grading policy question now has a measured magnitude attached (16.6% of credit rides on it). CORRECTED END-TO-END BASELINE: 797 + 226 = **1,023/1,500 = 68.2%** — the true floor, uniform across all rows; multi-round's asymptote frame starts here. - **PRE-MULTI-ROUND REGISTRATIONS (2026-07-08):** (1) THE CURRICULUM-PURITY CHECK: prep labels failures by GRAPH-match — right-asked-wrong-graph parses (16.6% of correct!) may sit in the mining set as "failures," training the specialist to "fix" correct readings toward canonical gold (a bias against exactly the equivalence flexibility the 16.6% represents). Check the contamination, then filter: MINE FAILURES BY ANSWER-FORCING, not graph-match — the honest-metric principle applied one stage upstream. (2) GRADING-POLICY OPTIONS, registered while nothing is at stake: STRICT-GRAPH (undersells 16.6%, immune to equivalence errors); FORCED-ANSWER (tonight's metric — honest where uniqueness holds); ANSWER-MATCH (MATH-500's native grading — vulnerable to luck the forcing probe can't run on non-generated problems). The forced-vs-answer-match delta on generated corpora = the LUCK-INFLATION estimate to carry into any benchmark claim (measured tonight: 5/802 = 0.6% at this corpus's uniqueness density). (3) The 99.4%-honest audit note: audits expected to confirm are still worth running — they convert "presumably fine" into load-bearing. - **MULTI-ROUND (2026-07-08 night): THE REGISTERED FRAME CONFIRMED.** Per-round recovery: **123 -> 39 -> 5 -> 0** (19.6% -> 7.7% -> 1.1% -> 0%) — declining exactly as registered, hard asymptote by round 4 with 460/703 survivors unrecoverable by ANY depth of this loop. NO violation -> the ledger-conditioning hypothesis does NOT arrive early; the loop cannot move its own ceiling; the 46.7% frontier stays where the ranking put it (reading-repair: deeper prefix / ledger re-parse). Multi-round total: 76 + 167 = **243/703 = 34.6%**; END-TO-END: 797 + 243 = **1,040/1,500 = 69.3%**. THE PURITY RESULT: **279/1,622 = 17.2% of mined "failures" were right-asked-wrong-graph** — independently matching the 16.6% bigtest class rate (the equivalence class is a stable property of this parser, ~1/6 of its correct readings are non-canonical). HONEST TRADE-OFF: the pure specialist's round-1 recovery dipped (123 vs the impure 150) — fewer training samples and/or the removed parses carried useful canonicalization signal; multi-round total still exceeds the impure single round (243 vs 226). Round-2+ exists: +44 answers the single-round stack left on the table. - **THE DECAY-SHAPE READING (2026-07-08, registered):** 19.6% -> 7.7% -> 1.1% -> 0 is FASTER than geometric — a stochastic-retry world decays geometrically; this cliff is a HARD PARTITION: a fixable population pumped dry in two rounds + a different-in-kind remainder. Independent confirmation of the decodability boundary from a new instrument: the ceiling is a boundary between populations, not an average over a difficulty gradient. (2) EQUIVALENCE CLASS PROMOTED TO DESIGN PARAMETER: ~17% (16.6/17.2 across independent draws, within 0.6 points) — budgetable, not re-measured; every grading policy, curriculum filter, and benchmark claim inherits it. (3) PURITY-DIP DISAMBIGUATION, kept answerable: evicted-signal vs fewer-samples separates by a sample-matched impure retrain — not run (the trade was right regardless), recorded so it stays a question, not lore. (4) **THE 460 CHARACTERIZATION, registered prediction:** the reading-repair hypothesis says survivors are ENRICHED for referential-binding stress (oblique mentions, shuffled letters, size) relative to the recovered population — binding is the thrice-located shallow-layer weakness. UNIFORM across teeth = the reading-repair story needs rework BEFORE the ledger re-parse is built to serve it. Zero GPU; converts the boundary from a direction into a target profile. - **THE 460 CHARACTERIZATION (2026-07-08): PREDICTION REFUTED — SURVIVORS ARE UNIFORM ACROSS TEETH.** Enrichment (survivors vs recovered, n=460 vs 243): oblique **1.01x** (0.850 vs 0.840), shuffled **0.92x** (0.687 vs 0.749), irrelevant **0.85x** (0.463 vs 0.547), n_vars **0.99x** (13.80 vs 13.91), band **1.07x** (2.44 vs 2.29). Not one feature above 1.07x; shuffled and irrelevant are slightly DE-enriched. The registered rule fires: the reading-repair story REWORKS before the ledger re-parse is built. WHAT THIS SEPARATES: binding stress explains the parser's ERRORS (thrice-located, real) but NOT which errors are UNREPAIRABLE — the teeth dials are orthogonal to survivorship. The hard partition is drawn along an axis the input-feature profile cannot see; the live candidates are PARSE-SIDE properties: error MULTIPLICITY per parse (the §6 density-regime law predicts exactly this — at ~5 errors/failure, withhold-2 + single-round fixes exhaust at low multiplicity and strand the high-multiplicity tail) and wrong-FIELD mix (which heads are wrong, not which inputs are hard). Next cheap probe, registered: profile survivors vs recovered on errors-per-parse and per-field error distribution (gold is available; zero GPU). PREDICTION: survivors are enriched for error multiplicity ≥3; if THAT also comes back uniform, the remainder is plausibly decode-degenerate (belief never concentrates) rather than mis-read, and the frontier reranks toward the deducer-side suspicion transplant instead of any re-parse. (The ledger re-parse is NOT killed — it is unjustified-as-designed; its premise must be re-earned.) Script: `scripts/characterize_survivors.py` (commit 4e19697). - **THE CONDITIONAL MULTIPLICITY PROBE, registered (2026-07-08, the relay's sharpening):** the naive prediction ("survivors enriched for multiplicity") has a TAUTOLOGY RISK — the loop fixes ~1-2 errors/round, so dense parses surviving is arithmetic, not discovery. The informative cuts are CONDITIONAL: (1) S(m) — survivorship as a FUNCTION of initial errors-per-parse + midrank AUC; (2) the mechanical model — recover-by-round ceil(m/f), grid f: does multiplicity alone reproduce 123->39->5->0? (3) the residual — within m-bins, do survivors differ in FIELD MIX or teeth? THREE FUTURES, thresholds pinned BEFORE measuring: (A) AUC>=0.75 AND decay reproduced (each round within ~2x) -> the loop is ROUND-BUDGET-LIMITED on the dense tail; answer = more rounds + better suspect-ranking, NOT re-parse or transplant. (B) AUC 0.6-0.75 or per-bin residual structure -> multiplicity real but unsaturated; the residual axis is the frontier. (C) AUC<0.6 -> plausibly DECODE-DEGENERATE; the deducer-side suspicion transplant reranks up. Side-glance registered: the shuffled/irrelevant DE-enrichment (0.92x/0.85x) predicts loud teeth produce LOW-multiplicity detectable errors — checkable in the same run. Script: `scripts/survivor_multiplicity.py`. - **MULTIPLICITY PROBE RESULT (2026-07-08): FUTURE C FIRES MECHANICALLY — AUC(m->survival) = 0.524 < 0.6 — BUT WITH A POST-HOC SIGNATURE NONE OF THE THREE FUTURES ANTICIPATED.** Multiplicity is uniform too (means 8.13 vs 7.32); Future A doubly dead: NO fixes-per-round capacity reproduces the front-loaded 123->39->5->0 decay (all mechanical models predict flat/rising per-round recovery) — iteration does not compound; third-instrument confirmation of the hard partition. THE SIGNATURE: S(m) is INVERTED at the low end — m=1 failures survive at **0.929** (n=42), m=2 at 0.837, vs ~0.71 for m>=4. The most unrecoverable population is the parses with the FEWEST errors. Candidate mechanism, named post-hoc: OMISSION BLINDNESS — `missing` is the largest error kind in both populations (~29-36% of error mass), and a missing factor is structurally invisible to the entire stack (withhold only REMOVES — worsening an underconstrained parse; the specialist's flags attach to EMITTED slots and its unflagged->copy objective actively teaches it never to ADD). Loud-teeth glance confirmed the relay's read: shuffled -> MORE errors (m 8.65 vs 6.15) yet DE-enriched among survivors — loud teeth make in-jurisdiction errors. REGISTERED FOLLOW-UP (CUT 4, prediction pinned before measuring): decompose m = m_add (missing) + m_corr (wrong-field/phantom/query). OMISSION-BLINDNESS predicts survival tracks m_add and collapses at m_add=0; flat-in-both = genuine decode-degeneracy and the transplant rerank stands. If omission- blindness confirms, the frontier is an ADDITIVE repair mechanism — the second look must be allowed to say MORE, not just different (the ledger re-parse premise re-earned in a narrower, reshaped form: omission-repair, not reading-repair). Script: `scripts/survivor_multiplicity.py` CUT 4; profile persisted to `.cache/survivor_profile_bigtest.npz` (future cuts zero-GPU). - **CUT 4 RESULT (2026-07-08): OMISSION-BLINDNESS REFUTED TOO.** AUC(m_add) = 0.525, AUC(m_corr) = 0.522 — survival tracks NEITHER decomposition. The cell table actively inverts the prediction: (m_add=0, m_corr=1) — one in-jurisdiction emitted-slot error, the stack's bread and butter — survives at **0.914** (n=35), the HIGHEST cell; (m_add=2+, m_corr=1) survives at 0.500, the lowest. m<=2 survivor error kinds are broad (rel_args 0.27, missing 0.27, given_value 0.21), not missing-dominated. TWO POST-HOC SIGNATURES: (1) query_wrong = 0.14 of m<=2 survivor error mass vs 0.03 population — ~5x enriched; a wrong query has NO slot: unwithholdable, unflaggable. (2) The (0,1)-cell inversion names SUSPECT-RANKING BLINDNESS: withhold-2 strips the two LEAST-confident factors; a confidently-wrong factor escapes, the flags hand the specialist the WRONG suspects, and unflagged->copy propagates the true error forever — which also explains the front-loaded decay (same mis-pointed flags every round). CONVERGENCE: this mechanism and the mechanical Future-C rerank agree — the binding constraint is SUSPICION QUALITY, not repair capacity; the deducer-side suspicion transplant is exactly a better suspect-ranker. NEXT PROBE registered (`scripts/survivor_suspicion_rank.py`, blank-parse-only — identity from the npz, no 4-round replay): P1 AUC(min wrong-slot confidence-rank -> survival) >= 0.65; P2 withhold-2 coverage several-fold higher among stage-1-recovered than low-m survivors; P3 m<=2 survivors dominated by rank-escapes + unflaggable query errors. FLAT ranks = the suspicion story dies too; decode-degeneracy stands. - **SUSPICION-RANK RESULT (2026-07-08): THE SUSPICION STORY DIES TOO — THIRD CONSECUTIVE REFUTATION, AND IT FLIPS THE FRONTIER.** P1 FLAT: AUC(min wrong-slot rank -> survival) = **0.518** (bar 0.65); wrong slots sit near the bottom of the confidence ranking in EVERY population (min-rank-norm 0.026-0.044; frac-in-bottom2 survivors 0.356 vs round-recovered 0.346). Localization is NOT the bottleneck. P3 INVERTED: only 26.6% of m<=2 survivors are rank-escapes/query — **~73% had their single error correctly flagged in the bottom-2 and the specialist STILL failed to fix it, 4 rounds running.** REPAIR GENERATION is the wall: told exactly where it is wrong, the parser cannot produce the correct replacement, and being deterministic it re-emits the same wrong content every round — the true explanation of the front-loaded decay. STRATEGIC CONSEQUENCE: the transplant rerank LOSES its rationale at the moment decode-degeneracy is confirmed — the transplant is a better RANKER and ranking is already adequate. Decode-degeneracy now stands on direct evidence, not elimination, and points at the frozen L0-L3 trunk: the information to reconstruct these factors is plausibly not in the trunk states at those positions (§6: a frozen 4-layer prefix can't BIND references) — binding weakness finally connects to survivorship at its correct jurisdiction: it explains WHICH CONTENT is unreconstructable, not which inputs are hard. CUT 2 registered (same script): flagged-but-unfixed survivor errors ENRICHED for rel_args vs flagged recovered errors, bar >=1.5x -> binding-is-the-wall; uniform-across-kinds -> head-capacity story, escalate to a trunk-information probe. Script: `scripts/survivor_suspicion_rank.py`. - **CUTS 2/3/3b (2026-07-08): THE ANATOMY BOTTOMS OUT — THE PARTITION IS ENCODE-SIDE vs DECODE-SIDE.** CUT 2: the binding prediction refuted AGAIN at the relation level — rel_args **0.78x** (DE-enriched; flagged binding errors are relatively fixable); the enriched kind is **given_value 1.41x** (0.56 of m<=2 survivor flagged-error mass). CUT 3 anatomy: flagged given_value errors are predominantly VALUE-ELSEWHERE (misbinding — right number, wrong variable) in every population (survivors 0.755, round-recovered 0.900); hallucinated values (not-in-gold) are the minority but **2.4x enriched** among survivors (0.245 vs 0.100). CUT 3b: SWAPS pass the ratio bar (0.043 vs 0.014 = 3x > 1.5x) but FAIL ON MASS — 10/233 cases; the joint-decode conclusion pinned to the bar does NOT fire. LEDGER LESSON: **an enrichment bar without a mass bar is a trap** — register the mass the mechanism must explain, not just the ratio (the §6 enforcement-is-arithmetic rule applied to one's own bars). THE SYNTHESIS: dominant unfixable errors are ONE-DIRECTIONAL misbindings — no coordination needed, a single-slot edit suffices — yet 4 correctly-flagged rounds fail. The chain closes on the substrate: every repair round re-decodes the SAME precomputed frozen trunk states; only flag features change. If the binding/value was mis-committed AT ENCODING TIME, no head-side conditioning recovers it. THE HARD PARTITION, NAMED: the repair stack drains DECODE-side errors (front-loaded 123->39->5->0) and cannot touch ENCODE-side casualties — which explains every refutation at once (uniform teeth: encoding failures happen at some rate on all inputs; flat multiplicity; adequate localization; flagged-but-unfixed). ORACLE-CEILING ARM registered (`scripts/survivor_oracle_ceiling.py`): perfect gold-derived per-field flags (= the specialist's TRAINING regime; deployed withhold-flags were doubly OOD: mis-pointed + all-fields pattern) re-derived each round, 4 rounds, on the 460. Upper-bounds ALL flag-quality improvements in one number. REGISTERED: <10% recovery -> encode-side wall MEASURED (frontier = change the ENCODING: second-view re-render with position-aligned suspect marks per the §6 structure law, and/or deeper prefix — NOT any suspicion/repair improvement; the transplant rerank dies too). >30% -> the deployed flag deriver was the constraint (fix it — cheapest win of the week). 10-30% -> partition and re-profile the oracle-recovered. - **ORACLE-CEILING RESULT (2026-07-08 night): 64/460 = 13.9% — the mixed band, near the low end; THE ENCODE-SIDE WALL IS MEASURED.** Perfect gold-derived per-field flags (the specialist's own training regime), re-derived each round, 4 rounds: 44 -> 16 -> 4 -> 0 (front-loaded AGAIN — 4th independent sighting today). **396/460 = 86% of survivors — 26.4% of the corpus — are unrecoverable even when told exactly which slot and which field is wrong.** Flag quality was never the lever: the ceiling bounds EVERY suspicion/ranking/repair improvement (tier-0 trained head, transplant, better derivers) at <=64 cases (<= +4.3 pts end-to-end, and only partially capturable gold-free). The frontier is a BUILD decision now — change the ENCODING (second-view re-render with position- aligned suspect marks per the §6 structure law, and/or deeper prefix), options + dead-ends tabled in `docs/NEXT_SESSION.md` for the relay seam. Registered follow-up left open per the 10-30% rule: partition + re-profile the 64 oracle-recovered vs the 396 (zero GPU, profile npz on disk). - **THE DECISION + PRE-BUILD DISCRIMINATORS (2026-07-08 night, relay call):** OPTION 1 (second-view re-render) is the chosen frontier — the multi-cycle Alternator's first empirical mandate (396 encode-side casualties = the measured need for re-reading under deductive feedback; the speculative component's customers arrived before the component, same as tier-0/tier-3). HAZARD FLAGGED: learned mark embeddings backprop through the trunk = the documented AM-driver hang. v0 = ZERO-NEW-PARAM marks: reserved vocabulary tokens inserted at suspect spans (token-shift/gold realignment machinery exists), forward-only re-encode; learned marks are the gated upgrade arm. WHY text-NACK's death doesn't transfer: it died on REFERENTIAL binding ("statement 7" -> sentence 7 through 4 shallow layers); a position-aligned mark carries its information BY BEING AT THE LOCATION — nothing to bind (the §6 structural-entry law at the trunk's front door). TWO DISCRIMINATORS FIRST: (a) option-4 re-profile of oracle-64 vs hard-396 (zero GPU, folded into the oracle script rerun); (b) THE DEPTH PROBE (`scripts/survivor_depth_probe.py`) — a fresh value-probe reads gold given values from mean-pooled gold-span states at L4 vs L8, evaluated on clean baseline vs wrong-givens-recovered vs wrong-givens-396. THREE-WAY VERDICT, bars pinned: instrument bar base-L4 > 0.70 else no verdict; ROUTING = 396 within 10pts of baseline at L4 (info present, head mis-routes — the §6 attention-bootstrap ghost; marker = attention beacon; deeper prefix retired); DEPTH = >=20pt gap at L4 AND L8 closes >=50% (deeper prefix wins; re-render overkill); CONTENT = >=20pt gap, L8 closes <50% (never written; re-render mandated). v0 FRAME registered (density regime: multi-error encode-side- selected population): ANY meaningful recovery on the 396 moves a measured ceiling; relay's directional prediction — recovery CONCENTRATED in one-directional given-misbindings (mark = "re-examine this given" at its location) confirms directed attention; UNIFORM recovery = generic re-rolling. - **DEPTH-PROBE VERDICT (2026-07-08 night): ROUTING WALL — DECISIVE.** Fresh value-probe, digit-exact at gold given spans: base **L4 0.998 / L8 0.953**; wrong-recovered 1.000/0.946; **the 396's wrong givens 0.996/0.942**. The gold value is FULLY PRESENT in the current L4 encoding at its location, at baseline fidelity, on exactly the givens the parser misbinds and cannot repair. DEEPER PREFIX RETIRED with a number (L8 <= L4 everywhere — depth smears local literals). HONEST CORRECTION to the partition's name: "encode-side" was operationally right (no conditioning fixes it) but mechanistically WRONG — the partition is decode-side vs **ROUTING-side**: the trained pointer circuit deterministically reads the wrong location (§6 attention-bootstrap law: pointers don't move without direct supervision), and this happens DESPITE span-supervised training. Option-4 re-profile came back FLAT (oracle-64 vs 396: mult 0.89x, kinds ~1x, teeth ~1x) — the 64 are flag-quality stragglers of the same population, one mass. THE FORK OPENED BY THE VERDICT (relay to adjudicate): **(A) span-restricted structural read** — the probe IS a repair head (reads only the span, 0.996 on the hard population): flagged given slot -> pool states over the suspect var's predicted mention span -> probe-decode value -> substitute -> re-solve. Zero re-render, zero retrain; deployability = gold-free span prediction quality (mention head exists, unmeasured on survivors); scope = given_value class (~0.36 of survivor error mass). **(B) marker-token re-render v0** — the general mechanism (now understood as attention BEACON, not re-encoding); design wrinkle: deployment places the marker via the model's own routing (marks where it LOOKS, not where it should look) — the train/deploy placement gap needs its own registered measure. **(C) both, sequenced** — A as the immediate bite, B as the Alternator build with A's recovery as the baseline to beat. Probe cached: L8 states `.cache/algebra_bigtest_L8_states.npy`; script `scripts/survivor_depth_probe.py`. - **ARM C REGISTERED + FIRED (2026-07-09): REPLACE-AND-SOLVE — candidate- restriction repair, the sibling of withhold-and-solve.** The relay's adjudication: C is not merely cheapest-first — the routing verdict makes it the MECHANISTICALLY CORRECT fix (don't steer a learned pointer; shrink its candidate space — the same move as every §6 win), and it demotes marker-v0 from thesis-test to comparison arm. Law 7's read-side discipline applied to attention itself: remove/restrict at READ, from the candidate set, never from state — the state was fine all along. DESIGN SHARPENING at build time: for GIVEN VALUES the right reader is not even the probe — values are digit literals in the input text; the inventory lexes symbolically at 1.0 reliability (the probe earns its place in v1 on args/ops where lexing can't reach). Moves: REPLACE (one given's value <- inventory; withhold could never fix a misbound given — removal loses the constraint; replace keeps it, corrected) + SWAP (the coordinated two-slot exchange a parallel marginal decoder can't emit). Solver disposes; ACCEPTANCE pinned: all passing moves must agree on the answer, else ambiguous-reject. Fully gold-free. DEPLOYMENT-HONESTY: fires only on VISIBLE failures; forced-WRONG originals (gold-only failures) counted separately — they are a grading-policy problem, not a repair problem. REGISTERED: mine — 60-120/460, concentrated low-m given_value, ambiguity <10% of fixable; relay (polarity-flipped) — this arm takes the BULK of the convertible fraction and the beacon adds ~nothing on top (pointers don't re-aim under conditioning; a beacon is conditioning via input) — beacon-beats-floor would be the interesting outcome; soundness — accepted-but-wrong ~0 (forced-unique acceptance admits no luck; measured). RELAY PRE-READ on the mention-quality unknown: teeth-uniformity implies survivors aren't adversarial text, so mention F1 should hold; degradation would CONTRADICT the uniformity table and be a finding itself. §6 PROMOTION QUEUED (fifth sighting, first at inference): **pointer errors are never fixed by conditioning — only by candidate restriction (structure) or retraining (supervision).** Two remedies, zero exceptions so far. Script: `scripts/repair_replace_swap.py`. - **ARM C v0 RESULT (2026-07-09): SOUNDNESS GATE FAILED — AND THE FAILURE IS THE FINDING.** 60 accepted, **55 WRONG** (luck gate predicted ~0, measured 0.92): FORCED-UNIQUE IS NOT A CORRECTNESS CERTIFICATE when the graph itself is suspect — substitute into a multi-error parse and you can force a consistent, unique, WRONG answer. Accepted set: mean m 8.12, m<=2 share 0.13 — imposters on high-error parses, not repairs. Honest yield +5/460; as designed the stage ships 55 indistinguishable wrong answers — NOT DEPLOYABLE. Both registrations refuted (mine 60-120 concentrated-low-m: off 10x with the wrong shape; relay's bulk-of-convertible: the single-move convertible fraction is ~5). THE DEEPER DISCOVERY: **70/460 survivors are forced-WRONG originals — deployment-INVISIBLE.** A single misbound given usually keeps the system fully constrained -> forces cleanly to a wrong answer -> looks like a SUCCESS gold-free. The m=1 survivors at 0.93 were hiding here all along: the routing-wall population mostly doesn't present as failure. THE BUG-CLASS THIS EXPOSES: every acceptance test in the measured stack compared against GOLD; deployment accepts any forced answer at every stage. Three contaminations: forced-wrong one-shots accepted wrong (never reaching repair); phantom recoveries (measured recoveries whose original was forced — deployment never fires); withhold/round imposters. AUDIT REGISTERED + FIRED (`scripts/deployment_honest_audit.py`): full stack replayed under gold-free acceptance; P1 one-shot forced-wrong 100-180; P2 phantom recoveries >0; P3 per-stage precision declines down the stack; P4 deployment-honest end-to-end < 0.693. NOTHING IS QUOTED WITHOUT THIS NUMBER AGAIN. Law candidate (with arm C as first sighting): **acceptance criteria must be measured for imposter rate at the deployed error density** — "forced-unique" was pinned as sound from a clean-graph intuition and failed at m~8. - **DEPLOYMENT-HONEST AUDIT RESULT (2026-07-09): P4 REFUTED UPWARD — the honest number is 1051/1500 = 70.1% (> the gold-checked 69.3%), and the audit caught a SECOND acceptance bug on the way.** Per-stage gold-free: one-shot 887 accepted / 797 correct (precision 0.899; 90 forced-wrong committed — P1 near-miss vs 100-180); withhold 127/74 (**0.583 — the LEAKIEST stage**: removal-based acceptance weakens forcing, arm C's law quantified in the deployed stack); rounds 203/140, 56/38, 3/2, 1/0. Answered 1277 (abstained 223), answered-precision 0.823. P2 CONFIRMED: ~20 of the measured 243 recoveries were PHANTOM (their originals forced wrong at one-shot; deployment commits them wrong and never repairs). P3: pattern yes, monotone no. WHY P4 INVERTED: the measured pipeline's round acceptance ONLY evaluated the WITHHELD variant (solve_check always ran k_wh=2) — a fully-correct specialist re-parse, minus its two least-confident (correct) factors, often un-forces and was REJECTED. The audit's accept-plain-first ordering recovers ~+11 net correct (round-1 140 vs measured 123). SECOND SIGHTING of the acceptance- criteria law in two days: acceptance rules must be AUDITED as mechanisms, not assumed — both bugs (gold-checked accepts; withheld-only round accepts) were invisible to every headline number until replayed under deployment rules. **THE QUOTABLE NUMBER GOING FORWARD: 70.1% deployment-honest end-to-end, answered-precision 0.823, abstention 14.9%.** The 226 committed-wrong answers are the measured customer for a calibrated abstention signal (the waist-space/OOD thread — probe queued behind this audit). - **WAIST PROBE REGISTERED + FIRED (2026-07-09, the autoencoder thread lands as instrumentation):** relay + Code corrections adopted — the probe space is **fst**, the algebra head's slot-vector bank (the one decoder-backed waist space; the TAP has no decoder — category error corrected; pointer heads excluded as problem-relative). HALF 1, interpolation coherence, registered **50/50** (the coordinate-swap evidence does not bear on convex combination; 128d-lossless hints tame): same-kind cross-problem pairs, alpha=0.5, decoded through the linear field heads; pinned coherent = sharpness ratio >=0.80 AND midpoint-decodes-an-endpoint >=0.50. Coherent -> KL machinery buys little; garbage -> measured deficiency for the per-kind-prior VAE arm (single-prior KL stays parked — cousin of the attract dud). HALF 2, the paying customer: per-kind fst centroids from TRAIN (deployable labels), score = worst slot cosine to claimed-kind centroid, vs the audit's 226 committed-wrong / 1051 correct. Registered: dense AUC 0.55-0.65 (misbindings look locally normal); USABLE-FLAG bar precision@top-10% >= 2x base (0.354). Classified on arrival per the portfolio rule: dense ranker / rare-flag / dead. Audit script now persists per-sample outcomes (`.cache/deploy_audit_bigtest.npz`). - **WAIST PROBE RESULTS (2026-07-09): BOTH HALVES LAND.** HALF 1 **COHERENT**, decisively: sharpness ratio 0.940 (bar 0.80), midpoint-decodes-an-endpoint 0.843 (bar 0.50), n=561 same-length cross-problem pairs. The parse-side waist is SMOOTH WITHIN KIND — convex combinations decode cleanly; the coordinate-swap evidence indeed did not transfer to interpolation. Per the pinned rule: **KL/VAE machinery buys little — parked, no deficiency** (the per-kind-prior note stays parked alongside). HALF 2: dense AUC **0.728** — my registered prior (0.55-0.65) REFUTED UPWARD; clears BOTH portfolio bars: dense ranker (>=0.70) AND rare-flag (precision@10% 0.417 vs bar 0.354; top-20% = 45% recall of committed-wrong at 0.40 precision). First instrument that consults neither solver nor emission-head confidence; the two halves cohere (centroid distance is meaningful BECAUSE the space is smooth). POLICY NOTE pinned before use: blind abstention LOSES accuracy on MATH-500-style grading (drop ~153 correct to avoid ~102 wrong at top-20%); the paying use is **flag-as-NACK-on-ACCEPTED-answers** — route flagged accepts through a second look instead of committing. This is the §8.5 session-monitor role arriving from measurement (retransmission decisions on accepted traffic), not from spec. NEXT CUT (cheap, registered direction not prediction): stage-split of flagged wrongs — concentration in the 90 one-shot invisibles would close the arc (the routing-wall population, first invisible to repair, now visible to the waist monitor). Script: `scripts/waist_abstention_probe.py`. - **RATCHET-NACK REGISTERED + FIRED (2026-07-09, relay adjudication: ratchet, not re-roll):** flagged accepted answers KEEP their original by default; the NACK round's revision replaces only if it STRICTLY DOMINATES. The asymmetric hazard: most flags land on correct answers (~0.4 precision), and the 0.998 unflagged-preservation number was measured on FAILURES, not re-opened successes — a population the specialist was never trained to leave alone. v0 SCOPE: stage-0 accepts only (the detector scores plateaued-parser space; later-stage parses live in specialist space, centroids uncalibrated — the per-stage score table logged this run decides v1); v0 DOMINANCE = revision forced + unflagged slots factor-exact-agree (anomaly-decrease criterion deferred to v1 for the same cross-space reason). REGISTERED BARS: break rate 0 by construction (any break = dominance leak -> tighten); fix rate on flagged stage-0 wrongs > 0.346 (relay, density regime stated: committed-wrong parses are NEAR-MISSES by selection — the regime where per-round repair is strongest); deploy iff answered-precision rises without net answers falling (floor 0.823 / 70.1%). TWO DIALS from one detector: recovery mode (ratcheted replacement) and precision mode (flag->abstain on undominated). Stage-split logged (direction: flagged wrongs concentrate in the 90 one-shot invisibles); per-stage calibration column logged (if repaired parses skew anomalous even when correct, v1 needs per-stage calibration before trusting the ratchet beyond stage 0). **THE MONITOR SENTENCE (relay): the session monitor — the perceiver's last unclaimed job but hosting global latents — arrived as a centroid library plus a distance threshold, watching accepted outputs and routing suspicion back into the loop. Zero parameters, eighteen months early, measured on arrival.** Script: `scripts/ratchet_nack.py`. - **RATCHET-NACK v0 RESULT (2026-07-09): KILL BAR FIRED + FIX-RATE REFUTED — DETECTION AND REPAIR ARE SEPARATE CAPABILITIES.** BREAK=1 at top-20% (0 at top-10%): a correct parse's flagged slot revised to a different value that still forced with unflagged agreement — formally dominant, actually wrong. The dominance test LEAKS; the deferred anomaly-decrease criterion is LOAD-BEARING (v1 needs per-space centroids). FIX RATE 1/17 / 2/37 (~6%) vs the 0.346 bar — relay prediction refuted WITH MECHANISM: the flagged stage-0 wrongs ARE the routing-wall invisibles — near-misses BECAUSE they are single confident misbindings, the population the oracle ceiling already measured 86% unfixable with perfect flags. The selection argument never asked WHY the parses were near-misses. **SELECTION EFFECTS HAVE JURISDICTIONS TOO** (the density-law's cousin, 5th sighting of unexamined-population predictions). PER-STAGE COLUMN: stage-0 AUC 0.745; stage-1 DEAD (0.532); correct parses drift anomalous with stage (0.703->0.748) — per-stage calibration required before v1 trusts anything beyond stage 0. Wrong-recall roughly uniform across stages (~0.41-0.51); false alarms concentrate late. DIALS AS MEASURED: recovery mode FLAT (+1 net, 70.1%); precision mode REAL — **0.880 answered-precision at 0.615 end-to-end** (top-20%) — the instrument for wrong-costs-more-than-missing deployments (not MATH-500). THE HONEST ROLE: the monitor SEES (0.728) but this specialist cannot FIX what it sees; the flagged population's only untried repair is the parked BEACON arm (polarity-flipped prediction standing: input marks can't re-aim trained pointers) — beacon failure there = the population is DETECT-AND-ABSTAIN ONLY under current machinery. Script: `scripts/ratchet_nack.py`. - **THE BEACON, FIRED AS THE CLOSING MEASUREMENT (2026-07-09, relay adjudication):** the 396's story is complete except one sentence and both endings close it. Population: the 460 (states 99.6% correct, pointer mis-aimed, 86% unfixable under perfect flags; every conditioning repair dead by measurement). The beacon is the ONLY untried arm — INPUT-level saliency, mechanistically distinct from all head conditioning. v0: bracket the suspect sentence (flagged slot's attention-argmax sentence — marking where the pointer LOOKS) with reserved token 128002, forward-only L0-L3 re-encode, UNCHANGED heads re-parse, gold-free acceptance with right/wrong split. PINNED ENDINGS: <=2% recovery -> flipped prediction confirmed, pointers don't re-aim under input conditioning either, the population is DETECT-AND-ABSTAIN ONLY, chapter closes zero-loose-ends; >=10% -> input marks move what conditioning can't (the week's most interesting result; beacon graduates). COMPOSABILITY COLUMN (relay): monitor score on beacon-accepted parses — drops-on-repairs = detect->beacon->re-score composes into a self-contained final tier; no-movement = the monitor can't certify its own fixes (ratchet lesson one level up). v1 RATCHET explicitly DEFERRED (relay): building better replacement machinery for a population measured 86%-unreplaceable is the infrastructure-before-customer trap; waits for a population both flaggable AND fixable. §6 LINE (relay's named form, third sighting): **a selection criterion's jurisdiction is which property it selects on — "survived filter X" is evidence about detectability, not repairability.** NEXT CHAPTER after the verdict: the registry expansion (MATH-500 relations), carrying the design constraint forward — every new relation's pointer gets candidate restriction and span supervision FROM BIRTH, or it grows its own 396. Script: `scripts/beacon_closing_arm.py`. - **BEACON VERDICT — THE CHAPTER CLOSES (2026-07-09): 14/460 = 3.0%,** middle band (bars were <=2% / >=10%), honest call: the flipped prediction is CONFIRMED IN SUBSTANCE — input marks do not meaningfully re-aim trained pointers; the population is **DETECT-AND-ABSTAIN ONLY under current machinery**; the beacon does NOT graduate. THE FOOTNOTE, recorded: 11 of 14 recoveries are in the hard-396 — parses PERFECT oracle flags could not fix. Input perturbation occasionally moves what head conditioning provably cannot — an EXISTENCE result, not a mechanism: 71 accepted-WRONG vs 14 right (precision 0.165; the marks shake the table, they don't aim the pointer). Third consecutive confirmation of the acceptance law. Composability: the monitor ranks marked-state accepts in the correct direction (right 0.736 < wrong 0.782) — a score-gated beacon is possible in principle but is v1 machinery for a 3% mechanism (infrastructure-before-customer; deferred with the v1 ratchet). **THE SURVIVOR ARC, FINAL LEDGER:** teeth uniform -> multiplicity flat -> omission dead -> suspicion flat -> binding de-enriched -> routing wall (states 99.6% correct, pointer mis-aimed) -> oracle ceiling 13.9% -> invisibles found (70 forced-wrong) -> acceptance bugs x2 -> honest 70.1%/0.823 -> monitor arrives (0.728, zero params) -> ratchet leak + fix 6% -> beacon 3%. Nine registered refutations, four §6-grade laws, two retired builds, one working instrument, one closed population. NEXT CHAPTER: the registry expansion toward MATH-500, carrying the constraint — every new relation's pointer gets candidate restriction + span supervision from birth, or it grows its own 396. - **THE MATH-500 BAND-SWEEP (2026-07-09 — the registry chapter's first move, measurement before build):** MATH-500 acquired (`.cache/math500_test.jsonl`, 500 problems, subject/level/answer labels — MEASURED, never trained on). Transparent regex classifier + n=20 hand-audit. EXPECTATIONS SCORED: plain-integer answers **62.2%** (registered 50-60 — slightly above; +11.2% fractions = rationals near-term; ~26% tail of expressions/tuples/radicals/ intervals); algebra+prealgebra 41.2% (reg ~40 ✓); geometry 26% (reg ~20, under-called); INEQUALITY **14.0% measured with a KNOWN UPWARD BIAS** (the audit caught "for n >= 1" domain qualifiers firing the tag; reasoning-core share <10%) -> **ARCHITECTURE VERDICT AS REGISTERED: interval reasoning is a LATER chapter; csp_core's predicate interface stays closed.** AUDIT CAVEATS: "factor" inflates quadratic/poly; the linear-arith-only residual (17.2%, mean level 2.8 — the easiest band) is OPTIMISTIC (radical- simplification problems hide in it) — current-registry-reachable < 17%. GREEDY SET-COVER LIST (marginal): geometry +52, quadratic/poly +49, trig/precalc +54, modular/divis +44, inequality +41, combinatorics +41... (100% at 13 categories). ENGINE-FIT ANNOTATION (for the relay's tranche call): raw coverage ranks geometry/trig first, but they need different FRONTS (diagrams, continuous identities — far from factor-graph CSP); the CSP-NATIVE tranche is **quadratic/poly (integer polynomial roots — still integer-domain, search-tier-able), modular/divis (GAC-native — the search tier eats these), ratio/percent, sequence/series, base-repr, abs-floor** — factor-graph-friendly relations entering through predicate + bridge as always. Mean-level column: linear-arith 2.8, inequality 4.1 (difficulty tracks the shopping list's tail). Script: `scripts/math500_band_sweep.py`. - **TTA + PROGRESSIVE RESIZING REGISTERED (2026-07-09, Bryce's fastai transplant; TTA FIRED):** the MC-pi lens pins the design constraint — views must be solution-preserving with DECORRELATED failure modes (correlated darts estimate nothing). DEPLOYMENT-HONESTY FLAW NAMED BEFORE FIRING: re-rendering needs GOLD factors — oracle machinery at MATH-500 time (the graph is what parsing is FOR; re-rendering the parser's own graph correlates with its errors). TWO ARMS, mirroring oracle-flag->deployed-flag: ARM O = K=4 gold re-renders (letters/templates/surfaces/order shuffled, teeth-easy — the mechanism ceiling); ARM D = K=4 sentence permutations (pure text transform, graph-free, deployable; sentence-index features genuinely shift). THE MC-PI GATE measured FIRST per arm: same-wrong-answer rate across views on wrong-forced originals < 0.30, else voting is VOID. REGISTERED: relay — ARM O voting recovers a NONZERO routing-wall slice (the only mechanism class that changes what the pointer keys on; sidesteps, not steers) + agreement-AUC near/above 0.728 (portfolio- combinable: behavioral stability vs representation geometry); mine — O decorrelates (<0.30), D weaker (0.30-0.60, misbinding plausibly local to sentence content), voting net-positive in O / ~flat in D; imposter split standard; vote rule majority >=3 of 5. PROGRESSIVE RESIZING (registered, rides with the registry tranche — no run now): the resolution axis is RENDERING COMPLEXITY (teeth + n_vars/band), not waist width (that is capacity, already measured); each new relation ships a coarse->fine curriculum knob + the ablation (curriculum vs all-teeth-from-birth, same budget). 50/50-leaning-modest; the bet is POINTER-CIRCUIT PREVENTION (candidate-restricted pointers learn binding on unambiguous mentions before obliqueness arrives — prevention of the next 396 is worth more than a point of fac-exact). One machinery, two clocks: TTA is the generator's symmetries at inference; the curriculum is the same dials at training. Script: `scripts/tta_views.py`. - **TTA RESULTS (2026-07-09): THE DEPLOYABLE ARM DOMINATES THE ORACLE ARM, AND COMPOSITION SETS A NEW HEADLINE — 1072/1500 = 71.5% / precision 0.833.** ARM D (sentence permutation, graph-free): MC-pi gate PASSED at 0.106 same-wrong (my 0.30-0.60 registration REFUTED upward — even pure order-shuffle moves the misbindings; 20% of wrong originals' views come back RIGHT); voting 763R/13W/724A = answered-precision **0.983**; routing-wall recoveries **33/460 survivors + 8/90 invisibles** (more than beacon+ratchet combined, deployable, zero training); agreement-AUC **0.840** — the board's best anomaly signal (beats the waist monitor's 0.728; behavioral stability > representation geometry). ARM O (gold re-renders): most decorrelated (0.031) but per-view forced collapses to ~21% (always-shuffled letters are far off the parser's operating point) — 69 accepts (all right), agreement-AUC 0.637. **THE MC-PI SECOND CLAUSE, measured: darts must be independent AND land on the board** — view design trades decorrelation against per-view competence, and gold access was not the binding resource; parse competence was. The oracle arm is the floor, the deployable arm the ceiling — a happy inversion. COMPOSITIONS (zero-GPU from persisted outcomes): COMPOSE 1 (vote->stack) **1072/1500 = 71.5%, precision 0.833, answered 1287** — all three dials beat the floor simultaneously; deploy-iff met; THE NEW QUOTABLE. COMPOSE 2 (stack->vote) 70.7%. Portfolio note: low-agreement top-128 catches 31/226 stack committed-wrongs. Relay's nonzero routing-wall prediction CONFIRMED on both arms. npz: `.cache/tta_arm_{O,D}_bigtest.npz`. Scripts: `scripts/tta_views.py` + inline composition. - **PORTFOLIO + THRESHOLD RESULTS (2026-07-09 night):** READ 1 bars REFUTED AS REGISTERED — Spearman(disagreement, waist) = 0.464 (bar <0.4), combo AUC 0.833 < agreement-alone 0.840 (bar >0.86) — BUT the tail tells the opposite story: the combo WINS at every abstention operating point (top-10%: flag-precision 0.528 vs 0.386, kept-precision 0.862 vs 0.846; +1.6pt). MY BAR WAS MIS-REGISTERED per the existing §6 law — AUC is a whole-ranking summary, abstention is a TAIL decision (4th sighting of metric-must-match-decision-structure, this time in my own registration). HONEST VERDICT: the portfolio PAYS where it is operated; the waist signal's rare-flag character complements behavioral agreement in the tail while diluting the mid-ranking. READ 2 CONFIRMED (relay): the certification channel exists — **unanimity 5/5 = 0.9982 precision at 38.1% coverage (570R/1W)**; t=4/5 = 0.9925 @ 44.3%; t=3/5 = 0.9832 @ 51.7%. LEDGER SENTENCES adopted from the relay: (1) THE THIRD CATEGORY — nine mechanisms tried to FIX THE ESTIMATOR (same input, better behavior); TTA changed what the input looks like and AVERAGED — randomizing away bias instead of repairing the instrument; deterministic surface-keyed failure is TTA's favorite food. (2) THE INDEPENDENCE-COMPETENCE LAW (§6 candidate, both clauses): darts must be independent AND drawn from the distribution the board was calibrated on — decorrelation buys nothing past the point where per-view competence falls faster. (3) CURRICULUM METRIC REFRAME: progressive resizing chases VIEW-ROBUSTNESS (competence under re-rendering, measurable per-relation from birth via the TTA harness), not raw fac-exact — a voting system needs per-view competence above the vote's break-even, after which independence does the rest. Scripts: `scripts/portfolio_and_threshold.py`; waist scores persisted (`.cache/waist_scores_bigtest.npz`). - **THE DECISION LATTICE, FROZEN (2026-07-09, relay structural note): the deployment stack's four rungs and their interfaces — the productized Alternator, whatever the multi-cycle loop becomes.** Every rung zero-parameter and gold-free: 1. **CERTIFY** — TTA-D K=5 unanimity of forced answers. Dial: **0.9982 precision @ 38.1% coverage.** 2. **ANSWER** — majority 3/5 vote; on vote-abstain, the deployed stack (one-shot -> withhold-2 -> 4 specialist rounds), gold-free forced acceptance. Composite dial: **71.5% end-to-end / 0.833 precision.** 3. **FLAG** — rank-sum(view-disagreement, waist-centroid distance) read at the tail; downgrade or abstain per deployment mode. Dial: kept-precision **0.862 @ 10% abstention.** 4. **ABSTAIN** — no forced answer anywhere. INTERFACES: views = solution-preserving TEXT transforms (sentence permutation v0; any future view generator is priced by the independence-competence curve); votes = forced answers only; anomaly signals = agreement + per-kind waist centroids (train split). **THE EXPANSION ACCEPTANCE TEST: a new relation passes when the lattice HOLDS ITS DIALS on the expanded domain — one table, all four rungs — not on fac-exact alone.** A relation that lifts fac-exact but degrades a certification dial fails acceptance. (The cheapest insurance against the tranche quietly breaking a channel nobody re-measured.) - **THE TRANCHE CHARTER (inherited, one breath):** relations enter as predicate + bridge, zero core edits — the generality law's next test, now on the neural side too; corpora solution-first, gate-checked, band-labeled, mention-spanned; every pointer born candidate-restricted and span-supervised (the 396's rule as prevention); every relation ships its curriculum knob chasing VIEW-ROBUSTNESS; the five-seat audience grades from the first checkpoint. **DECISION PENDING (Bryce + relay, BEFORE the generator is written): MULTI-ROOT ANSWERS.** Quadratics break the single-forced-value frame; the policy determines gold format, forcing-probe semantics, and what unanimity MEANS on a set. Options tabled: (1) ANSWER-SET gold (forcing probe generalizes via ban-and-resolve enumeration; every lattice rung's semantics changes; multi-value answers are only 2.0% of MATH-500); (2) QUERY-CONSTRAINED single root (the NL carries a selector — "the positive solution", "the larger root" — a new mention type under the pointer law; forcing probe and ALL lattice semantics unchanged; matches MATH's dominant convention); (3) hybrid: selector-graded with the full set in gold metadata as a diagnostic column. CODE'S RECOMMENDATION: (2) as v0 with (3)'s metadata — quadratics enter as just-another-relation plus one supervised mention type; the set-frame waits for the 2% it serves. - **MULTI-ROOT POLICY RATIFIED (2026-07-09, relay + Code): option 2 with option 3's metadata.** Option 1 failed on jurisdiction grounds the ledger owns — re-defining every lattice rung a day after the freeze to serve a measured 2% is the mass-bar law applied to design. THE SELECTOR SPLIT (the load-bearing detail — pointer law + division of labor at the same door): the parser's selector head classifies the selector TYPE from a supervised mention span (positive / larger / smaller / in-range — a small CLOSED vocabulary, the <=32-way codebook-selection regime that bootstraps from task gradient); the SOLVER enumerates roots and applies the comparison symbolically. The parser never computes a value to compare — it reads which comparison was requested. GENERATOR REQUIREMENTS folded in: (1) the selector gets its own TEETH (oblique phrasings — "the solution that isn't negative") and its own DECORRELATION CHECK: the selector must be INVARIANT under view re-rendering (permutation may move the phrase; the referent may not change). (2) REGISTERED PREDICTION (relay, jurisdiction stated): selector errors will be RARE but disproportionately SILENT — right graph, right roots, wrong pick; undetectable by UNSAT and by uniqueness (the selected root is forced GIVEN the selector) — their natural detector is behavioral: view disagreement on the ANSWER despite agreement on the GRAPH, logged as its own diagnostic column. If confirmed, the five-seat audience becomes load-bearing for quadratics in a way it never was for linear systems. (3) NO-REAL-ROOTS POLICY: gated OUT of the training corpus (discriminant sign stamped into metadata; banked not built) — "the correct answer is that there is no answer" is a semantically different abstention than "I'm not sure," and the abstain rung doesn't take on that meaning until a benchmark category demands it. THE TRANCHE HAS NO OPEN DESIGN QUESTIONS — only builds, on Bryce's word. Second time this week a thrice-measured REAL phenomenon turned out not to govern the question it was assumed to govern (binding weakness -> survivorship; before that, the density trap). The instrument lesson's causal cousin: locating a real cause of X does not license it as the cause of adjacent-X. One sighting from a §6 promotion. - **THE TRANCHE, BUILT + FIRED (2026-07-09, Bryce's word):** four seams, each committed green before the next. (1) REGISTRY: LTYPE_MOD (a mod k = r, params=k) + LTYPE_SEL (x = sel(a,b), closed vocab larger/smaller/even/odd; ties and not-exactly-one-even are VIOLATED — ill-defined selectors SELF-GATE through uniqueness) via predicate + bridge, **zero csp_core edits** (8th/9th ltype; empty git diff is the proof). mul(x,x) square forms EXCLUDED v0 — repeated scope vars would make the pairwise propagator unsound (the general-regime law applied preemptively). Soundness gates: exhaustive predicates, 500-trial propagator support checks, hole-monotonicity, and end-to-end (Vieta+sel forced; UNSELECTED pair provably symmetric; both-even self-gates; CRT chains solve — two mods force a=17 through propagation). (2) GENERATOR (`algebra2_nl_gen.py`): Vieta pairs (sum+product = the integer-domain quadratic; discriminant always a perfect square by solution-first construction) + selector factors with their own teeth (oblique phrasings at teeth*0.3) and VIEW-INVARIANCE by construction (whole-sentence templates) + MOD in two roles (derived = calculator band; CRT with lcm > m = engine band — the lcm>m uniqueness requirement was caught by the roundtrip gate on first fire). SYMMETRY-AWARE gate: Vieta root pairs are symmetric BY DESIGN (the text cannot bind letters to roots); the gate requires MULTISET match on pairs + exact/unique everywhere else; the query pool never draws raw roots; the mod-base pool excludes roots (teeth orthogonality). Corpora: 2500 train + 800 test at teeth 0.8, token-budget gated; mixed train = 2000 old + 2500 new = 4500. Bands 1-8, ~1.3 sel + ~1.1 mod factors/sample. (3) HEAD: 4-way ftype + h_sel behind **ALG2=1** (legacy build BYTE-COMPATIBLE — every lattice script still loads the old ckpt); explicit per-kind loss masks (the old rel mask (1-is_lit) is wrong once mod/sel exist); mod modulus rides the digit head; sel args ride the bilinear pointers (span supervision via fspan/vspan = POINTER LAW AT BIRTH); WARM_FROM loader with printed skips (train-side allowance; eval loads still hard-error); env-able corpus/ckpt/split names (no clobbering of legacy artifacts). (4) TRAINING FIRED: warm from the legacy ckpt (h_ftype/h_sel fresh), 14k steps cosine, pick-best-by-val, mixed corpus -> `.cache/phase1_algebra2_head.safetensors`. ACCEPTANCE NEXT: per-band eval on alg2test + the LATTICE TABLE (all four dials, old bigtest regression) — a relation that lifts fac-exact but degrades a dial FAILS. Curriculum ablation (all-teeth-from-birth vs coarse->fine) = the one-extra-run arm, after the v0 baseline lands. - **TRANCHE ACCEPTANCE ROW 1 (2026-07-09): OLD-CORPUS REGRESSION PASSED — UPWARD.** The tranche head on old bigtest: **888/1500 = 59.2% one-shot ANSWER vs legacy 802/1500 = 53.5%** (+86 answers, +5.7pt), graph-solve 741, query 0.98-1.00 across bands, fac 0.75-0.88. NOT mere no-forgetting: the mixed-corpus warm-start IMPROVED the old domain — weight-side generality at the head level (new relations helped the old ones; the multitask grail's parser-side sighting). New-corpus row: 480/800 = 60% one-shot at teeth 0.8 with sel/mod/CRT in play. Remaining lattice rows queued in NEXT_SESSION: TTA dials on alg2test, monitor centroids in the new fst space, specialist retrain (composed-stack number gates on it), curriculum ablation (view-robustness graded), the relay's selector silent-error column. - **THE +5.7pt DIFFERENTIAL, REGISTERED + FIRED (2026-07-09, relay):** "compound interest" is a description, not an explanation. Three candidates: (a) regularization-by-variety (progressive-resizing logic arriving through the data mix); (b) representational pressure (nine kinds organize the slot space more cleanly than seven — neural-collapse frame; old-kind centroid DRIFT during the monitor rebuild is the free mechanism diagnostic); (c) PLUMBING — the loss-mask bug fix + fresh 14k cosine would lift the old corpus with zero new relations. THE CONTROL: legacy corpus only, tranche code path, same schedule/warm-start. BARS: ctrl >= 870 -> (c) dominates, compound-interest dies honestly; <= 820 -> the generality thesis has its cleanest parser-side evidence; between -> mixed, attributed proportionally. Relay registration: 60/40 toward real-but-smaller generality with (c) contributing (a fixed bug the same day as a surprise gain is what the audit-that-confirms principle exists for). SEQUENCING adopted: the selector silent-error column RIDES the TTA run (shortest shelf life — retraining after the curriculum ablation would muddy whose errors got measured). Design appreciation on record: discriminant-perfect-square-by-construction DISSOLVED the no-real-roots policy (edge case made unrepresentable, not handled); the roundtrip gate refusing symmetric pairs until taught multiset-match, and ill-defined selectors self-gating as VIOLATED — three edge policies, zero new mechanisms. - **DIFFERENTIAL VERDICT (2026-07-09): 820/1500 — ON the pinned bar; the GENERALITY CLAUSE FIRES.** Decomposition of the +86: **+18 plumbing** (mask fix + fresh cosine — real, minor, correctly priced), **+68 GENERALITY** (the new relations themselves — 79% of the gain). The relay's 60/40 registration confirmed in shape: real-but-smaller, (c) contributing. Mechanism hint in the per-band contrast: the mixed head beats the control most on QUERY pointing (0.98 vs 0.89 at band 3) — new relation kinds sharpened the referential machinery the old corpus had plateaued (consistent with hypothesis (b); the centroid-drift diagnostic will say more). THE SPEC SENTENCE, EARNED: **expanding the registry improved the core — the weight-side generality grail, now measured on the parser side with its plumbing honestly deducted.** Every future expansion-improves-core claim cites this control. Ctrl ckpt: `.cache/phase1_algebra_ctrl_head.safetensors`. - **LATTICE ACCEPTANCE ROW 2 (2026-07-09): CERTIFICATION SURVIVES THE EXPANSION — unanimity 5/5 = 1.0000 precision (411/411) at 51.4% coverage on alg2test** (legacy head on old bigtest was 0.9982 @ 38.1% — the tranche head is MORE view-robust on a HARDER domain). t=3/5 = 0.9979 @ 59.4%. MC-pi gate 0.104. Only 12/800 one-shot invisibles (bigtest: 90/1500) — the new corpus produces fewer forced-wrongs. CAVEATS, honest: (1) per-kind coverage column DEGENERATE by generator design (n_vieta = randint(0,2) or 1 -> every sample has a selector; no within-corpus linear/mod contrast) — the per-kind baseline moves to the cross-corpus comparison / a kind-varied slice; (2) the selector silent-error column as implemented catches QUERY-pointer flips (4/4108 graph-agree pairs, 0.10%), not selector-PICK flips (a wrong pick changes the graph hash) — signature refined: pick errors live in pairs whose graphs differ ONLY at the sel factor; rare-and- silent confirmed in spirit, precise enrichment queued with the refinement. REMAINING ROWS: old-domain TTA regression (doubles as the per-kind linear baseline), monitor rebuild + centroid drift, specialist retrain, curriculum ablation. Script: `scripts/tta_alg2_dials.py`. - **ROW-2 ASTERISKS + THE DIVIDENDS REGISTRATION (2026-07-09, relay):** (1) ZERO-NUMERATOR discipline: 1.0000 @ 411 certified is "no errors observed, error rate upper-bounded ~0.25%" — NOT "the channel is perfect." The original one-wrong-in-571 is the correct prior; this draw is consistent with it, not better. (2) The selector silent-error prediction stays FORMALLY OPEN until the old-domain TTA regression supplies the linear-kind baseline — it resolves against its full comparison, not half of one. (3) REGISTERED WHILE CHEAP — THE DIMINISHING-DIVIDENDS PREDICTION: expansion-improves-everything has a natural expiration. Representational- pressure gains are sublinear in class count (nine-to-fifteen buys less than seven-to-nine); at some tranche the mixed-corpus dilution cost overtakes the organization dividend. PREDICTION: dividends diminish; the crossover shows FIRST in per-kind coverage on the OLDEST relations; the remedy when it arrives is CURRICULUM WEIGHTING, not tranche reversal. The lattice table is already the instrument that catches it — that is what the freeze grades every tranche for. If tranche 3 still improves everything, this registration dies happily and the generality thesis strengthens further. NEXT FIRE when the rhythm resumes: the SPECIALIST RETRAIN (gates the composed-stack headline on the expanded domain). - **THE ATLAS, NAMED + PARKED (2026-07-09, Bryce + relay): two hyperbolic spaces, two jurisdictions, zero shared coordinates.** THE BALL (topology, the original §7 object): one problem's factor-graph wiring, hierarchical by construction — behind its flag awaiting the relaxation research. THE ATLAS (semantics, NEW name, conditional): the problem POPULATION's subject taxonomy — relation kinds + domain centroids embedded by family, distinct from the ball per the two-channel spine (hanging the taxonomy on the ball would be the §8.2 channel conflation). FIRST CUSTOMER: the monitor's centroid library under expansion — flat libraries grade novelty gracelessly ("far from everything"); a hierarchical library separates "new leaf of a known family" (parse cautiously, flag for expansion planning) from "genuinely OOD" (abstain hard) — a certification-relevant distinction. Rhymes: radius-as-resolution (coarse cycles read families near origin, fine cycles read leaves near boundary) + per-family curriculum weighting (the pre-committed dividends remedy). TWO GATES, both armed, NOTHING BUILT: (1) the flat library's OOD gradation actually degrading under expansion (watched by the per-tranche lattice/monitor rows); (2) the delta-probe — Gromov delta-hyperbolicity / cophenetic correlation of the learned kind-centroid distance matrix, runnable at tranche 2-3 when the tree has depth (nine near-sibling kinds today = a bush). PRIOR ON RECORD: the radial-depth prize was refuted once (rho 0.13) — hyperbolic structure must be a measured property of the data, never an aesthetic; the representations say whether they know the taxonomy, first. Naming plea adopted: "the ball" and "the atlas," never "the ball" ambiguously. - **THE COMPOSED STACK ON THE EXPANDED DOMAIN (2026-07-09): 533/800 = 66.6% deployment-honest, answered-precision 0.896.** Chain debugged en route: forward_cond predated the sel head (None-grad -> optimizer refused, fixed with the same conditional-emit guard) + int8 overflow in the purity rebuild. Specialist v2: mined 1268 organic failures from FRESH data (self-defeating- curriculum law honored; 180 purity-removed), 6k steps. PER-STAGE: one-shot 489/477 = **0.975 precision** (vs bigtest 0.899 — the tranche head emits only 12 forced-wrongs in 800 vs 90 in 1500: the invisible-wrong class SHRANK by ~4x per capita on the harder corpus); withhold 58/40 (0.690); rounds WEAK: 48 accepts, 16 correct (0.40 -> 0.17 -> 0) — the fresh specialist underperforms its legacy sibling's round precisions (0.69/0.68). KNOWN LEVERS, not yet pulled: the NACK trainer lacks cosine+pick-best (loss rose 5.0 -> 6.8 late — the SAME hygiene gap that bit the parser overnight once); mining pool is small (1268). QUEUED: TTA COMPOSE-1 on alg2test (the vote channel printed 475 accepts @ 0.9979 but per-sample outcomes weren't dumped — one flag added next session) — the expected composed+vote headline sits above the stack-only 66.6%. Cosmetic: the audit's saved-path print is hardcoded (the save honors AUDIT_NPZ; bigtest artifact verified untouched). Audit npz: `.cache/deploy_audit_alg2test.npz`; specialist: `.cache/phase1_algebra2_nack.safetensors`. - **BRYCE'S GUT + THE DIAGNOSIS (2026-07-09): "we're not breathing right" — CONFIRMED BY ASSEMBLY OF THE WEEK'S OWN FINDINGS. The parser does not breathe: it is a ONE-SHOT PARALLEL DECODER (every slot argmaxes its marginal once; no slot sees the others' decisions; nothing settles) inside a project whose validated engine works BY settling. Re-read in this light: SWAPS = the textbook joint-vs-marginals failure (v98's own diagnostic); value-MISBINDINGS (75% of the wall) = relational errors two slots make about the same region with no collision mechanism; the ROUTING WALL = a pointer committed once with no second breath to contest it; the REPAIR DECAY refuted re-DECODING (changed conditioning, same state) — it does NOT touch breathing (evolving slot state is a different mechanism); TTA = five independent inhales, no exhale between. §8.1 spec'd cycles; the skeleton built steps 1-3 and the loop never came. REGISTERED PROPOSAL — BRICK-P, THE PARSER'S BREATH: transplant the v98 recipe to the slot banks (K=2-4 passes, slot queries attend waist + previous slot states, per-breath ladder CE, delta-gate, zero new mechanism kinds). BARS pinned on the week's signatures at fixed budget vs the one-breath incumbent: swap rate, misbinding-collision rate, invisible-wrong count — if breathing is the missing thing THOSE move; fac-exact-only = capacity, dies honestly. Cheap probe first: K=2 warm-started. - **THREE RELAY REGISTRATIONS (2026-07-09):** (1) THE PREVENTION LAW, registered pending tranche-2's second sighting: **confident wrongness yields to representational pressure, not to repair** — nine decode-side mechanisms got single digits; two relation kinds got 4x per-capita prevention as a side effect; if the invisible rate drops again as kinds grow, the law has two sightings + a mechanism (sharper class boundaries leave less room for decisive misbinding) and expansion becomes the ONLY measured lever on the class the abstention stack exists to contain. (2) SPECIALIST WEAKNESS: THREE causes pre-registered before the hygiene fix claims credit — trainer hygiene (known), thin curriculum (1268), and SELECTION-HARDENING (better parsers produce more survivor-like failures by construction — the selection-jurisdiction law aimed at the repair curriculum). Post-hygiene read: recovery to ~0.68 = trainer; partial = the equilibrium claim extends to the repair stack. (3) The 0.9979 gets zero-numerator discipline: "error rate bounded near a quarter percent across both domains" — three consistent draws, none independent enough to tighten the bound. - **BRICK-P FIRST LOOK (2026-07-09 night): K=2 breathing beats the one-shot incumbent on BOTH domains** — val 0.8109 (incumbent 0.8091, still climbing at step 8000); alg2test ANSWER **497/800 = 62.1%** (+17); old bigtest **920/1500 = 61.3%** (+32; legacy started the week at 802). Fac up across nearly every band both corpora; ~1M added params; 0.07s/step. HONEST FRAME, as registered: this is CAPACITY-COMPATIBLE — the verdict belongs to the SPLIT BARS (collision rate vs lone-misbinding rate vs invisible count, breath vs incumbent) — next session's first job. Micro-signal logged: alg2test query accuracy dipped at bands 2-4 while fac rose — watch in the split-bar read. Ckpt: `.cache/phase1_breath_head.safetensors` (ALG_BREATH=2 + ALG2=1 to load). - **BRICK-P SPLIT-BAR VERDICT (2026-07-10): MIXED — and the fourth outcome none of the three sentences anticipated: THE GAINS LIVE OUTSIDE THE COUNTED POPULATION.** alg2test: all given-error classes drop (wrongG -15%, swaps -28% vs lone -11% — collision-selective in proportion, weak verdict-A, but n=18->13 underpowered). bigtest: given errors FLAT (565->561) yet ANSWER +32 on the same corpus — the breathing gain lands in REL args/results/query (uncounted by bars aimed at given slots). INVISIBLES FLAT both domains — breathing does not touch confident wrongness; the PREVENTION LAW's territory stays expansion-only (3rd consistent observation). RIDER: belief-movement AUC 0.601/0.608 — real, weak, not a portfolio member on arrival. STANDING: breathing survives (it beat the incumbent both domains, gates-closed) but has NOT proven its distinctive relational mechanism at K=2; the registered next cuts are (a) REL-side error counters (the bars' blind spot, where the +32 apparently lives), (b) the K sweep (one settle step may simply be too shallow — the deducer needed 16), (c) swap-count power via a bigger eval draw. Ledger lesson: BARS INHERIT THE JURISDICTION OF THEIR COUNTERS — a verdict frame aimed at given slots cannot see a rel-slot mechanism. Script: `scripts/brickp_split_bars.py`. - **REL-SIDE COUNTERS, REGISTERED BEFORE BUILT (2026-07-10, relay):** the facts already constrain the finding. PREDICTION: if breathing's gain lives relationally, the thinned rel-side errors should be disproportionately COLLISION-TYPE (mutually inconsistent args/result/query claims — two slots claiming the same mention; result and query pointers disagreeing on one var) rather than LONE rel misbindings. Lone-thinning instead = the mechanism is NOT negotiation (candidates: per-breath supervision as regularizer; the settle step as implicit TTA on slot states) and the thesis needs its third formulation. THE COUNTERS CARRY THE SPLIT FROM BIRTH — "bars inherit the jurisdiction of their counters," applied prospectively one day after minting. K-SWEEP PRIOR calibrated: the parser's slot graph is shallow (factors touch 2-4 slots; chains short) — honest prior is saturation at K=3-4, NOT the deducer's 16 (49-cell lattice, ~4 hops/breath). Monotone gains past K=4 would itself be a finding: settling propagating beyond adjacency. RIDER re-measured at the sweep-selected K (one settle step barely lets the movement field mean anything; the deducer's 0.687 took sixteen breaths of dynamics). THE THREE-JURISDICTION DIVISION, now three-ways measured: PREVENTION (representational pressure — the only lever on confident wrongness), NEGOTIATION (breathing — the apparent lever on relational coherence, pending the rel counters), DETECTION (the abstention portfolio — for what neither prevents nor negotiates away). No overlaps claimed without a counter to witness. - **THE WEEK'S CLOSING FRAME (2026-07-10, relay — for the paper's discussion):** the three-jurisdiction row is the Alternator's spec rewritten by measurement. The plane-ride design assigned repair to a notebook, monitoring to a perceiver, cycles to a six-breath loop; what the month built — prevention through representational pressure, negotiation through settling, detection through a zero-parameter portfolio — is the same functional architecture with every component replaced by whatever the measurements ratified. **The spec's nouns died; its verbs all survived.** The honest answer to "did the design work": wrong in every particular, right in every jurisdiction — and the method (registration, kill bars, cheap disconfirmation, jurisdiction discipline) is what converted one into the other without an unexamined premise surviving the trip. - **REL-SIDE VERDICT (2026-07-10): NEGOTIATION REFUTED — THE THIRD FORMULATION IS RE-READING.** alg2test: LONE thinned 2x more than COLL (-14.3% vs -7.4%), query REGRESSED +14.3% (flagged); bigtest: COLL -3.4%, LONE flat, QUERY -45%. The dominant consistent signature is in the column nobody predicted: **MISSING factors dropped both domains (-5.4%/-9.6%)** — the breath head FINDS rels the one-shot never emitted. That picks between the pre-listed candidates: not ladder-regularizer, not slot-negotiation — the h_tok pathway (text re-attention CONDITIONED ON BELIEFS) recovering missed factors on a second conditioned read. THIRD FORMULATION: **the parser's breath is a second look at the page, not a negotiation among the readers.** Coverage, not coherence. DISCRIMINATING CUT queued (cheap, one arm each): ablate h_slot vs h_tok — which term carries the gain; if h_tok alone suffices, the slot-slot machinery (and its mask) simplifies away and K-sweep becomes a re-read sweep. Query regression on alg2test rides the ablation as a watch column. Script: `scripts/brickp_rel_bars.py`. - **BRICK-R REGISTERED, NOT FIRED (2026-07-10, Bryce's packet instinct + relay mapping): THE SELECTIVE-REPEAT LOOP.** Today's stack is STOP-AND-WAIT ARQ (send whole parse, await verdict, retransmit whole, x4) — networking retired it fifty years ago; the three upgrades map onto built machinery: (1) SELECTIVE REPEAT — sequence-number-stable factor identity: VERIFIED factors pinned as DELIVERED (never re-decoded, never re-risked — the ratchet's zero-break criterion PER-FACTOR), rounds spent only on the NACKed window; the mechanism that lets round counts grow without preservation risk (the cap that held multi-round at 4). The verifier's field-level flags are the ACK stream; the accumulate-ledger was always the receiver's buffer. (2) SENDER-SIDE CRC — tier-0 confidence vetoing phantom emissions BEFORE transmit (registered months ago, never deployed; dropping a known-bad frame at the sender costs nothing vs a full round-trip at the receiver). (3) CONGESTION CONTROL — per-problem adaptive round budgets read from the LIVE recovery decay (the ack stream as channel state; two silent rounds -> back off to abstention) — Dopri5 stepping at the session level, free from numbers the audit already logs. BAR: equal-or-better recovery at strictly lower round cost and ZERO delivered-factor breaks. SEQUENCING: does NOT jump the queue — the re-read finding reshapes what a round IS (if h_tok carries the architecture, a retransmission round is a conditioned re-read and selective repeat becomes "re-read only the NACKed spans," composing with masked attention). Waits on the h_tok/h_slot ablation verdict. - **ABLATION VERDICT (2026-07-10): THE ARMS ARE REDUNDANT, NOT COMPLEMENTARY.** tok-only: val 0.8155 (best), 495/924; slot-only: 0.8127, 501/910; both: 0.8109, 497/920; incumbent 480/888. EACH arm alone reproduces the full gain; the combined head exceeds neither anywhere — NO SYNERGY. If re-read and negotiation were distinct levers, both-terms should win somewhere; interchangeable channels point (Occam) at what both provide identically: AN EXTRA GATED, LADDER-SUPERVISED TRANSFORMATION STEP. The leading explanation is now DEPTH-WITH-SUPERVISION, not either named mechanism; the MISS-recovery signature may be what any second pass buys. REGISTERED CONTROL (the decider, one run): the DEPTH-ONLY arm — same gate, same ladder, second pass = plain per-slot FFN, no cross-attention, no mask, no re-read. Matches ~920 = breathing dies honestly as named depth (keep the simplest form); falls short = the attention second-look is load-bearing and tok-only (best val, zero mask machinery) is the keeper. Brick-R and the K-sweep wait on this verdict — their nouns change with it. Ckpts: `.cache/phase1_breath_{tok,slot}.safetensors`. - **THE DECIDER (2026-07-10): BREATHING DIES HONESTLY AS NAMED DEPTH.** Depth-only (blind per-slot MLP second pass — no text, no neighbors, no mask): val 0.8149, 497/800, 917/1500 — matching tok (495/924), slot (501/910), both (497/920) within a ~6-answer band. ALL second-pass variants are interchangeable; attention on the second pass is NOT load-bearing. Brick-P's kill criterion fires in refined form: the +2pt gain is real and earned (gates-closed) but the mechanism is ONE MORE GATED, LADDER-SUPERVISED STEP. CONSEQUENCES: (1) production head = the DEPTH form (simplest — zero mask machinery, no second bank pass; ties best val); (2) the K-sweep re-prices as ordinary depth scaling (not the frontier); (3) Brick-R stands ON ITS OWN — its rounds were never going to be conditioned re-reads; selective repeat/CRC/congestion control are protocol-level, orthogonal to head internals; (4) the parser-breathing thesis is REFUTED at K=2 in all attention forms — the deducer's breathing remains what it always was (validated, on graphs); any future parser- settling claim now carries the burden of beating the depth control. FOUR formulations in 48 hours: not-breathing -> negotiation -> re-reading -> depth-with-supervision. The gut found a real +2pt; the ledger found its true name. Ckpt: `.cache/phase1_breath_depth.safetensors`. - **THE ARC'S CLOSING SENTENCES (2026-07-10, relay):** the depth control was the ablation nobody wanted to be true, and it was built anyway. Not-breathing -> negotiation -> re-reading -> depth-with-supervision: each renaming SHRANK the claim — most projects' stories grow in the telling; this one's got smaller and truer at every instrument. The honest residue, undeflated: the gut found a real +2 both domains that three weeks of repair mechanisms never touched, and the production head is SIMPLER than the story implied. THE RE-PRICED LEDGER ENTRY (belongs beside the factorization result as the two-phase design's sharpest characterization): the parser's task, unlike the deducer's, has NO joint structure a single pass can't see — **the solver settles because constraints interact; the reader deepens because text doesn't.** QUEUE CORRECTIONS: (1) the K-SWEEP formally CONVERTS — its registered story (settling dynamics) died with the negotiation arm; re-registered small as DEPTH SCALING under deep supervision (prior: diminishing returns after +1 layer; the ladder now reads as deep supervision with its own literature and expected shape). (2) BRICK-R survives the renaming CLEANER: a retransmission round is unambiguously a conditioned re-decode; selective repeat's value — pin delivered factors, spend rounds on the NACKed window, budget by the decay signal — stands on pure protocol economics, no architecture story required. Registered on its own merits; runs on them. - **NACK HYGIENE VERDICT (2026-07-10): 66.6% -> 70.3% / 0.909 on alg2test.** Cosine + loss-EMA pick-best (save-after-restore) recovered round-1 precision 0.400 -> 0.648, round-2 0.167 -> 0.600; rounds now 45R/26W (was 16/32). THE PRE-REGISTERED 3-CAUSE SPLIT RESOLVES: recovery to 0.648 vs legacy ~0.68 = the trainer was MOST of it; the ~4pt residual is selection-hardening's share (the equilibrium claim extends weakly to the repair stack, as registered — the confound protection worked; hygiene could not steal it). The expanded-domain composed stack now EXCEEDS the old domain's 70.1% pre-TTA, on the harder corpus. PAPER FLAGS both closed same morning: 68.2 floor cited; census = **40.7M trained total, 9.1M in the deployed algebra lattice** (5.1M parser + 4.0M specialist) on 506M frozen-leveraged — the 90M title corrected 2x in our own disfavor-turned- favor. Script: `scripts/param_census.py`. - **THE EXPANDED DOMAIN'S FULL-LATTICE HEADLINE (2026-07-10): 567/800 = 70.9% / 0.910 answered-precision** (COMPOSE-1: 3/5 vote -> hygiene stack) — above stack-only 70.3%, above the ORIGINAL domain's 70.1%, on the harder corpus, from 9.1M trained parameters. The tranche is now CARRYING the dials, not holding them. PER-KIND RIDER (composed level, the curriculum's pre-intervention baseline): unanimity coverage sel-only 0.547 / sel+crt 0.519 / **sel+mod 0.473** — modular samples certify ~7pt lower: the view-robustness deficit, measured before the curriculum exists to move it. Per-view answers persisted (`.cache/tta_alg2_views.npz` — any-threshold re-votes now zero-GPU). QUEUE: monitor rebuild + drift, curriculum ablation (target: close the mod gap), tranche 2 vs its banked list (the prevention law's 3rd sighting + diminishing-dividends both come due). - **TWO NOTES FOR THE LEDGER (2026-07-10, relay):** (1) §9 HALF-SENTENCE — the "instruments arrive with their customers pre-measured" pattern, third sighting (tier-0 got the silents; the ledger re-parse got its population probe; the curriculum knob now gets a named 7-point gap instead of a vibe): registered measurement doesn't just prevent false claims, it PRE-POSITIONS every intervention with a target and a baseline — the method's compounding dividend. (2) TRANCHE-2 DESIGN DECISION, flagged BEFORE the generator is written: ratio/percent is the first relation kind whose answers flirt with RATIONALS — the integrality-jaw expiration's registered arrival condition. The generator must CHOOSE: integer-forced (keep the jaw, defer the expiration) vs rationals-in (pay the detectability cost early, taxonomy watching). Either defensible; chosen, not inherited. DECISION PENDING (Bryce + relay) before tranche 2 fires. Queue order held: monitor rebuild (drift feeds §3's +68 mechanism story) -> curriculum ablation (target: the 0.473 mod gap, graded composed-level) -> tranche 2 (prevention law 3rd sighting + diminishing-dividends due). - **INTEGRALITY DECISION RATIFIED (2026-07-10, relay + Code concur): INTEGER-FORCED TRANCHE 2; RATIONALS AS TRANCHE 3'S HEADLINE VARIABLE.** The deciding principle: ONE VARIABLE PER MEASUREMENT — tranche 2 carries two standing predictions (prevention law 3rd sighting; diminishing-dividends) whose attribution dies if the jaw retires in the same tranche. Tranche 2 ships ratio/percent/sequences/base-repr/abs-floor over Z (solution-first makes integer-forcing natural — the perfect-square move again). TRANCHE 3 = the integrality-expiration EXPERIMENT: same relation kinds, rationals admitted, one variable moved, detectability measured before/after — "detection power = constraint density" gets its cleanest demonstration; a figure, not a regression. Calendar aligns: the §8 external anchor needs rational-experienced parsing exactly then. CODE'S NOTE for tranche 3's registration: bounded-denominator rationals SCALE TO INTEGERS over a common denominator (LCM move) — Q-valued problems can enter as scaled-Z CSPs with exact predicates and zero core edits; the jaw's retirement is PARTIAL AND TUNABLE, not binary. Counter- argument on record: integer-forced ratio/percent is a slightly unnatural subspecies; risk bounded (generator controls difficulty; mentions are number-type-blind; digit heads extend as format). **Keep the jaw one more tranche, then retire it on purpose, with instruments watching.** - **DRIFT + MONITOR V2 (2026-07-10): MECHANISM (b) CONFIRMED IN DIRECTION; THE GEOMETRIC MONITOR DEGRADES BY SELECTION.** Drift (same old-corpus slots, per-space geometry): all three old-kind pairwise centroid cosines DROPPED in the tranche space (0.172->0.126, 0.092->0.088, 0.445->0.410) and within-kind coherence mostly rose (0.358->0.415, 0.401->0.454; rel_mul -0.02) — tighter clusters, farther apart: the neural-collapse signature; the +68's geometry, 5/6 stats in the registered direction (modest magnitudes, honest label). MONITOR V2: AUC **0.543** on the hygiene stack's 56 committed-wrongs (v1: 0.728 on the legacy 226) — SELECTION-HARDENING EXTENDS TO DETECTION (3rd bite): a better pipeline's residual errors look geometrically normal; detectors calibrated on a weaker stack's wrongs degrade as the stack improves. The abstention portfolio's geometric member weakens with pipeline quality — the behavioral member (agreement, 0.840 on the old population) gets its test on the 56 next (zero-GPU from tta_alg2_views.npz + the audit). Library rebuilt in tranche space (5 kinds): `.cache/monitor_centroids_alg2.npz`. - **AGREEMENT ON THE 56 (2026-07-10): AUC 0.925 — THE PORTFOLIO'S TWO MEMBERS SCALE IN OPPOSITE DIRECTIONS WITH PIPELINE QUALITY.** Behavioral (view disagreement): 0.840 on the legacy 226 -> **0.925** on the hygiene stack's 56; geometric (waist centroids): 0.728 -> 0.543 on the same populations. Mechanism: selection-hardening — a better stack's residuals are selected to look representation-normal, but remain BEHAVIORALLY unstable under re-rendering. THE DURABLE DETECTOR IS BEHAVIORAL, and it rides free on the votes already computed. Paper's abstention story updated: geometry is the weak-stack instrument; agreement is the strong-stack instrument; the portfolio's composition should re-weight toward behavior as the pipeline improves — measured, both directions. - **THE GOODHART COROLLARY, REGISTERED BEFORE CONTACT (2026-07-10, relay — selection-hardening's FOURTH face, its deepest: it applies to INSTRUMENTS, not just populations).** WHY behavior stayed sharp: selection only shapes errors against filters they actually FACE — survivors were selected past tier-0, verifier, uniqueness, monitor, but never against re-rendering; TTA was the HELD-OUT examiner. PREDICTION: the vote joined the acceptance path in the composed headline, so agreement entered the selection pressure — the NEXT generation of committed-wrongs will be selected to hold their story across five retellings; agreement-AUC on committed-wrongs will decline MONOTONELY across future stack generations (measure at each; the instrument doesn't weaken — its population hardens against it). THE DEPLOYMENT LAW: **any signal promoted to gate becomes selected-against; the portfolio must always hold one examiner out of the acceptance path.** Instrument rotation as design principle — today behavior polices geometry's blind spot; tomorrow something must police behavior's (bench candidates: the library cross-check, which never joined acceptance; genuinely new view families — paraphrase re-renders when the independence-competence curve prices them — unselected-against by construction). A law that began as a confound registration now explains why detectors AGE — a sentence the abstention literature doesn't have and §7 now does. - **CURRICULUM ABLATION REFUTED (2026-07-10): coarse->fine is STRICTLY WORSE at equal budget.** val 0.7698 vs 0.8091; one-shot 445 vs 480 (alg2test), 825 vs 888 (bigtest); unanimity coverage 0.474 vs 0.514. The 50/50- leaning-modest registration resolves past modest to NEGATIVE. MECHANISM: the fastai resolution analogy breaks — image resolution is the same distribution at lower fidelity; TEETH ARE A DISTRIBUTION SHIFT (the easy pool excludes patterns the test carries at 0.8), so 2/3 of the budget trained partly off-distribution and the decayed-LR final third couldn't recover. Pointer-circuit prevention never materialized. VERDICTS: all-teeth-from-birth is the KEEPER; progressive resizing dies for the parser (transfer condition failed: the axis must be fidelity, not distribution); the 0.473 mod-certification gap needs a DIFFERENT lever (candidates: more mod training mass in tranche-2's mixed corpus; a mod-targeted view family). Ckpt kept for forensics: `.cache/phase1_curriculum_head.safetensors`. - **TWO TRANCHE-2 REGISTRATIONS (2026-07-10, relay):** (1) COMPOSITIONAL CLOSURE — the tranche's first finding, before any training: new-ltypes per category covered is IMPROVING (T1: 2/2 = 1:1; T2: 2/4 = 1:2, sequences/ abs/ratio assembled from existing parts). START THE TABLE (one line per tranche); if T3's rationals cost <=1 primitive, the paper gains the claim nobody else can make: the relation menu converges toward a BASIS — coverage growth decouples from vocabulary growth, the strongest form of the generality thesis. Atlas implication: a compositional basis is FLAT by construction at the primitive level — hierarchy lives in COMPOSITIONS; the delta-probe should target PROBLEM representations, not relation embeddings. (2) HIDDEN VARIABLES ARE A NEW GOLD-FORMAT SPECIES (the ratio's product var; sequences' enumerated terms): variables with NO mention span. Pinned before the first template: empty mention-set is a TYPE, not a degenerate case; pointers are NEVER asked to bind them (generator-enumerated, solver-walked — the ratified division of labor); the round-trip gate verifies hidden-var plumbing survives reconstruction. Same class as week one's span-set contiguity catch — the gold decision everything downstream inherits, cheap now, an eval anomaly later. - **TRANCHE-2 FIRST FIRE BROKE — DIAGNOSED IN TWO CUTS, FIXED BY PAD-WARM (2026-07-10):** run 1 collapsed mod/sel domains (alg2test 480 -> 98, graph-solve 0) while IMPROVING pure-rel bigtest (907) — the discriminator (tranche-1 head through the current code: 480/202 EXACT) exonerated the eval bridge in one run; emission inspection showed degradation-everywhere, not plumbing. ROOT CAUSE: the warm-start's shape-mismatch skip DISCARDED THE TRAINED 4-WAY FTYPE ROUTER (4->6 widening) — the one head gating every per-kind loss mask relearned from scratch inside a converged circuit: the bootstrap-trap family, self-inflicted (new §6-family sighting: **never discard a trained router to widen it — pad-warm the prefix, fresh-init only the new rows**). Loader upgraded: prefix-shaped params copy their trained slice with a printed PAD-WARM. Retrain in flight, three-table verdict pending against bars 480/888 and the broken run's 41/98/907. - **TRANCHE-2 TABLES, HONEST (2026-07-10, after the one-character post-mortem: decode's ftype guard read ==4, sending every 6-wide slot down the legacy branch — the 'collapse' was a comparison operator; pad-warm was real hygiene but not the cause; the discriminator's exoneration of the 4-wide path was TRUE AND INCOMPLETE — a guard that dispatches on width is only exercised by the width you test):** alg2test regression **505/800** (bar 480, +25); bigtest **915/1500** (bar 888, +27) — EXPANSION-DIVIDENDS' THIRD INSTANCE (plumbing deduction pending; the differential precedent is the citation). alg3test DEBUT: **233/800 one-shot, fac 0.80-0.86 flat across bands** — reading strong, forcing sparse (the KenKen-g10 shape); QUERY dropped to 0.71-0.81 and the new binding surface is ORDINAL mentions (the registered suspect for the gap). Closure table row 2 confirmed shipped: 2 ltypes / 3 categories. NEXT (the five-prediction table needs the composed layer): specialist remine+retrain on mixed3, TTA dials + per-kind certification on alg3test, invisible-per-capita (prevention 3rd sighting), oldest-relation coverage (dividends), ordinal-query column. - **COMPOSED LAYER ON ALG3TEST (2026-07-10): THE DEBUT GAP IS A FORCING GAP, AND THE ORDINAL SUSPECT IS CONFIRMED WITH A GENERATOR ROOT CAUSE.** Rider: ordinal-term queries one-shot-fail **0.870 vs 0.656 direct** (+21pt). Audit: only 43/800 FORCE at one-shot (233 raw answers were largely under-constrained); composed 83/800; certification coverage 0.033; per-kind unanimity: **fdiv 0.008 / pct 0.026** vs linear 0.231. ROOT CAUSE (my own render3 comment flagged it, unimplemented): TERM VARS ARE LETTER-STARVED — seq sentences use ordinals only; the term's letter appears twice (preamble + query sentence, both low-content) and the query pointer binds on starvation rations. FIX (mechanical, next fire): seq sentences carry letter+ordinal APPOSITION ("the second term, e, is ...") and/or ordinal-phrase queries with recorded mentions; regenerate corpus, retrain, THEN read the five-prediction table — reading it now would measure the flaw, not the tranche. MC-pi gate incidentally PASSED at 0.025 (most decorrelated arm yet). Specialist v3 trained (1602 mined, purity 192); machinery all pct/fdiv-aware and banked. - **THE SYNC DIAGNOSIS (2026-07-10, Bryce's gut + relay walk + Code concur):** "out of sync" resolved to candidates 1+4 with this week's incident log as evidence — the specialist training one generation behind the parser (cross-generation curriculum lag, structural; the self- defeating law caught only the within-generation form) and ARTIFACT DRIFT (warm-start shape mismatch, audit-npz near-clobber, per-generation gold keys, env-ckpt coupling — pairwise-agreement burden growing quadratically, nothing enforcing it). Candidate 3 (the deducer static through three parser vocabularies — Phase 2 has never seen SEL/PCT/FDIV neurally) acknowledged as the ARCHITECTURAL desync — a chapter, not a fix; the Alternator's unpaid debt, December-scale. SHIPPED: generation manifest v0 (`scripts/generation_manifest.py`, `.cache/GENERATION.json`) — artifacts pinned by hash + env + regression bars, KNOWN-STALE as a tracked field (centroids in tranche-1 space; thresholds gen-1). REGISTERED v1: the atomic GENERATION BUMP — one script: remine -> specialist retrain -> centroid rebuild -> threshold refit -> manifest write; loaders refuse cross-generation mixes unless overridden. Synchronization converted from discipline (decays) to mechanism (doesn't) — the no-silent-fallbacks law applied to TIME. Gen-4 = the apposition corpus fix + the first full bump. - **TWO BUMP-DESIGN REGISTRATIONS (2026-07-10, relay):** (1) THE BUMP IS A TRANSACTION — atomicity has its own failure mode: five fallible stages must never leave the system in an undeclared N-and-a-half. The clean form: gen-N+1 builds entirely ALONGSIDE gen-N (new artifact paths throughout, nothing overwritten — the house pattern already lives this way); the MANIFEST WRITE is the single atomic commit point. The manifest is a transaction log, not a registry — generations as sequenced, acknowledged, retransmittable deliveries (the packet instinct one level up). Cheap to specify now; miserable after a half-bumped generation prints a plausible-looking table. (2) CANDIDATE 3'S PAYOFF SENTENCE, registered with the debt: when the deducer-meets-new-kinds chapter opens, the QUESTION is whether Phase 2 seeing PCT/SEL/FDIV neurally buys anything the symbolic tier doesn't — and the answer-shaped thread is SOFT GRAPHS: the parser's confidence outputs ARE the uncertain graph the original design promised the deducer; parser emits uncertain factors, deducer settles them NEURALLY, symbolic tier disposes what settles hard. The manifest names the debt; this names the payoff that would justify paying it. GEN-4 = apposition corpus fix + the first full transactional bump; the five-prediction table reads only after it. - **GEN-4 COMMITTED (2026-07-10, the first transactional bump — all stages green, manifest-last):** THE APPOSITION CURE WORKED — ordinal-q fail 0.870 -> **0.670** (gap to direct 21pt -> 11pt); graph-solve 95 -> 194; composed 83 -> **143/800**; TTA t=3 coverage 0.133 -> 0.297 @ 0.987; certification 38 @ 1.0000. **DIVIDENDS' FOURTH INSTANCE, BIGGEST YET:** alg2test 505 -> **541**, bigtest 915 -> **959** — four expansions, zero regressions, both prior domains improved every time, best val ever (0.8343). HONEST RESIDUAL: the new kinds certify at fdiv 0.030 / pct 0.050 — chains+params multiply the exactness forcing needs; real domain difficulty now, the specialist/lattice's territory. Gen-4 manifest written + checked; gen-3 untouched alongside. NEXT: the FIVE-PREDICTION TABLE reads against gen-4 (invisibles per capita across generations for prevention's verdict; oldest-kind coverage for dividends' crossover; the mod-gap re-read; freeze table #3; dividends attribution with the differential citation). - **TWO FRAMINGS BEFORE THE TABLE (2026-07-10, relay):** (1) THE APPOSITION CURE'S MECHANISM: not capacity, not architecture — a SUPERVISION-SURFACE problem. "The third term, l," gives the ordinal a letter to anchor: the pointer law's oldest clause (binding enters as structure) collecting its SIXTH sighting, in generator clothing, at the cheapest remedy on record — a comma and a letter. The fix wasn't in the model; it was giving the text something bindable. (2) THE SLOPE PRE-REGISTRATION: the dividends series reads as dividends-PER-EXPANSION; diminishing-dividends predicts the increment shrinks while the sign holds. If confirmed, BOTH standing claims win simultaneously and the reviewer-proof sentence is "expansion pays, at a declining rate, with the crossover instrumented and not yet arrived." ATTRIBUTION CAVEAT pinned: gen-3->4 bundles the corpus fix with fresh data — its increment (+36/+44) is NOT a pure expansion read; the differential-control citation covers tranche boundaries, not intra-tranche bumps. Bigtest series so far: 802 -> 888 (+86, T1) -> 915 (+27, T2) -> 959 (+44, gen-4 bundled); alg2test: 480 -> 505 -> 541. - **THE FIVE-PREDICTION TABLE, READ (2026-07-10, against gen-4):** (1) PREVENTION: **MIXED — two-sighting limbo continues.** Bigtest halved again (6.0% -> 3.33% invisibles, its second drop; forced-precision 0.950) but alg2test ROSE (1.5% -> 2.6%) — the law holds on the oldest domain and is counter-sighted on the middle one; not promotable, honestly split. alg4test debuts at 6.25% (new domains start high, as bigtest once did). (2) DIVIDENDS SLOPE: **both standing claims win** — direction 4-for-4 (541/959 all-time highs), PURE-expansion increments shrank (+86 T1 -> +27 T2; gen-4's +44 is bundled, unattributable by the pinned caveat). The reviewer-proof sentence stands: expansion pays, at a declining rate, crossover instrumented and not yet arrived. (3) MOD GAP: the mass lever PARTIALLY works — sel+mod 0.473 -> 0.500 under tripled neighbors (gap 7.4 -> 6.1pt); real, unclosed; the view-family lever stays on the bench. (4) FREEZE TABLE #3: **PASSED** — alg2test under gen-4: certification 0.516 @ 1.0000 (held from 0.514/0.9982), t=3 0.672 @ 0.9981 (improved); alg4test debut dials on record (1.0000 @ 0.048). (5) ATTRIBUTION: T1 differential cited; bundling caveat governs. INCIDENT LOGGED: the TTA views npz is a SHARED PATH overwritten between domains — artifact drift inside the same day the manifest shipped; views join the manifest's coverage at the next bump. The table waited three weeks and cost one afternoon to read — against a system that knows what time it is. - **THE TABLE'S CLOSING FRAMINGS (2026-07-10, relay):** (1) PREVENTION'S SPLIT IS THE LAW DISCOVERING ITS JURISDICTION — narrower and truer than its registration: representational pressure suppresses confident wrongness IN MATURE VOCABULARIES (class geometry sharpened over generations), while DEBUT vocabularies generate fresh invisibles faster than pressure can police. Not limbo — growth. The three-jurisdiction row holds at finer grain: prevention owns the old kinds' confident wrongness; detection owns the debut kinds'. (2) DIVIDENDS' DOUBLE CONFIRMATION is the paper's §6-meets-§9 exhibit: direction and diminishment both true, the bundling caveat pinned BEFORE the numbers could flatter it — a sentence that holds under push from either side because both sides were registered before contact. The certification channel has now survived TWO vocabulary expansions and a generation protocol WITHOUT ONCE BENDING (1.0000 on 413, zero-numerator discipline attached) — the artifact the paper leads with. (3) The views-file catch: the manifest shipped in the morning and recruited its next artifact by evening — mechanism finding its own customers. STANDING: the EXTERNAL ANCHOR (§8) is the one build between here and the arXiv draft; the evidence chapters are essentially written in banked measurements. - **HOUSE CLEANED (2026-07-10, Sonnet-scanned, gates-verified):** scripts: 53 concluded/broken scripts -> `scripts/archive/` via git mv (44 remain live: 22 pipeline roots + utilities + the doc-referenced active cluster); three v1XX trainers found BROKEN since the 751c56f deep-clean (imports deleted modules — archived). .cache: **535GB -> 63GB (472GB freed)** — deleted the two retired-era hoards (gsm8k_steps 130GB, v200_perceiver 118GB), the fg_v100-v121 orphan tail (~65GB), superseded kenken_ckpts, regenerable trunk/text-nack/L8/stale-generation state caches, unused HF downloads, and the dead breath-arm ckpts (depth + curriculum + ctrl kept per spec notes). PINNED SURVIVED UNTOUCHED: fg_ckpts, sudoku_ckpts, llama/pythia weights, gen-3/gen-4 artifact sets, all corpora jsonl. POST-GATES: manifest --check consistent (gen 4); live-pipeline imports OK; algebra2 soundness ALL PASSED; five-prediction numbers reproduced identically. Git history holds everything tracked, before and after. - **THE EXTERNAL ANCHOR (2026-07-10): HONESTY DOES NOT SURVIVE FOREIGN TEXT — the month's most important refutation, and §8's real content.** P1: ~as coverage predicted (answered slice small). P2 REFUTED: certified precision **2/97** on integer answers, 63 certifications on non-integer answers (0 possible) — the 1.0000-in-distribution channel signs foreign garbage confidently. P3 REFUTED: abstention FLAT across strata (67.5% vs 66.1%) — the lattice does not know what it doesn't know. MECHANISM (visible in the flat ~164/view forced counts): the parser mis-reads foreign text STABLY; sentence permutation decorrelates template variation, not distributional confusion — **unanimity certifies reading STABILITY, which coincides with truth only in-distribution.** Every portfolio signal (agreement included) is distribution-calibrated; OOD breaks the seal silently. THE PRE-REGISTERED CUSTOMER ARRIVES: the atlas's gate-1 (flat library's OOD gradation) is now OPEN — the missing organ is a TEXTUAL OOD DETECTOR firing before any parse is trusted (trunk-state distance from the training distribution; the "far from every family = abstain hard" read). THE HONEST §8 SENTENCE: on foreign text the lattice certifies stability, not correctness — the certification claim is DISTRIBUTION-BOUNDED, and the anchor measured exactly where the bound lies. More valuable than a soft yes: the December roadmap's true starting point is OOD-awareness, not coverage. Outcomes banked: `.cache/math500_anchor_outcomes.json`; script `scripts/math500_anchor.py`. - **THE FUNNEL (2026-07-10, Bryce's sixth instinct — arriving the same night as the measurement proving which stage it was missing):** the system IS a form-into-database funnel: parser = the form, registry = the schema, solver = the database (GAC = referential integrity, MRV/LCV = the query planner). Mycelium built every stage EXCEPT the one every production form has: **input validation at the mouth.** The anchor's 2/97 is a funnel with referential integrity but no schema validation on input — five hundred strangers' submissions poured downstream, every check below the mouth passing, because the checks validate the GRAPH's consistency, not whether the form was filled in a language the reader speaks. RECOGNITION (the fourth jurisdiction) is the funnel's mouth. ANCHOR MECHANISM, final form: all five darts thrown by the SAME ARM — on foreign text the arm's bias is systematic, not random; permutation decorrelates template-family errors, not dialect-never-learned errors. The anchor was DESIGNED as the held-out examiner — which is why this is the paper's most valuable measurement, not a deployment incident. RECOGNITION-ORGAN DESIGN, registered: (1) input-space OOD is SELECTION-SAFE (no pressure shapes errors against it unless the corpus chases the detector — which the generation manifest would make explicit); (2) zero-parameter first instrument: trunk-state centroids over the training family, threshold calibrated so MATH-500 reads foreign and held-out generator text reads native — BOTH populations already banked; (3) the §8 constructive close: **recognition buys honesty now; coverage buys capability later** — the funnel gets its mouth first, then the mouth learns more languages (paraphrase-teeth, tranche-3 rationals, real-text fine-tuning with the anchor as standing examiner). - **THE MOUTH WORKS (2026-07-10): P1 AUC 1.0000 both scores — foreign refused 100.0% at 1% native false-refusal; P2 all 160/160 of the anchor's false certifications REFUSED at the kNN threshold. The funnel has its mouth: zero parameters, one afternoon, calibration populations already banked. The lattice + mouth now signs NOTHING it cannot read — §8 closes constructively: recognition bought honesty tonight; coverage buys capability next. P3 HONEST READ: the gradation is FAINT (whole MATH-500 spread 0.236-0.273 vs native threshold 0.044 — everything is "different forest"; leaf-vs-forest unanswerable at this distance) and its ordering inverts intuition: Intermediate Algebra NEAREST, Prealgebra FARTHEST — our dialect is terse symbol-dense fact-sentences, nearer LaTeX-heavy text than natural prose. HYPOTHESIS LOGGED for the coverage roadmap: the language gap is PROSE STYLE before relation vocabulary — paraphrase-teeth toward natural prose may close more mouth-distance than new ltypes. Artifact `.cache/recognition_mouth.npz` joins the manifest at the next bump (with the TTA views file). Cosmetic: NaN-divide warning on skipped overflow rows (filtered, harmless — tidy at next touch). THE FOUR JURISDICTIONS COMPLETE: prevention, negotiation(->depth), detection, RECOGNITION — each with a measured instrument and a bounded claim. - **THE IR QUESTION, REGISTERED (2026-07-10, Bryce's seventh instinct + relay + Code):** the funnel already has THREE IRs — the registry (symbolic), the mention/span structure (annotation), and THE NATIVE DIALECT ITSELF (text-level: terse symbolic fact-sentences are a DISCOVERED canonical IR — the generator compiles graphs into it; the parser inverts it). What's missing is the compiler's FRONT HALF (verbose prose -> dialect) — never built because the corpus never contained prose. THE FREE LUNCH: solution-first generation renders the SAME graph in two registers -> paired (prose, dialect, graph) triples, gold at every layer — no designed logical form (discovered beats designed; C2's tombstone). THE THREE-OUTCOME PROBE (the fork, machinery = survivor_depth_probe transplanted): ship a VERBOSE teeth family; run the head + state-probe on verbose renders. (a) head parses fine after mixed training -> IR stays implicit, December = data; (b) states decodable, head fails -> head-side fix; (c) states not decodable -> ONLY THEN the explicit prose->dialect translation stage earns its build (output re-enters the funnel unchanged; layered funnels get layered mouths — dialect-conformance is easier than open-prose OOD; but generation machinery + a new silent-error species is the real cost, pre-registered). PRIORS: relay 70/20/10, Code 65/25/10. TWO PINS (Code): (1) T_ALG=256 will select verbose samples toward SMALL problems — match band/size across registers or read size-controlled slices (register-size confound, cheap at generation); (2) verbose training MOVES the native family — the mouth's threshold recalibrates per generation (joins the manifest's calibration constants at next bump). Even if (a) wins on generated prose, real MATH narrative re-asks the question at the boundary — where the mouth is standing. - **THE IR FORK RESOLVES: OUTCOME (a), OVERWHELMINGLY (2026-07-10).** Zero-shot register gap on MATCHED GRAPHS: terse 581/600 vs verbose 10/600 — near-total blindness. (b)/(c) discriminator: verbose given-value decodability **1.000** — (c) DEAD; four frozen layers compile narrative prose perfectly; the trunk was always bilingual. Mouth column: verbose read 0.093 pre-training — foreign but BETWEEN home (0.044) and MATH-500 (0.25), the learns-languages thesis' predicted geometry. AFTER 2000 pairs + 10k steps warm from gen-4: verbose **600/600 ANSWER**, terse twin 589/600, val 0.9752 — AND bigtest 926 > the 915 bar: **THE FIFTH DIVIDENDS INSTANCE — a new REGISTER pays like a new relation kind.** The translation stage dies unbuilt (10% priors, correctly); the IR stays implicit; December = MORE BOOKS, with mouth-distance-closed-per-corpus as the unit of progress. Caveats pinned: vtest pairs are budget-biased small (1137/3137 rejected — the size note); paired-register val runs hot; the post-training mouth re-read (does verbose now read native?) is the recalibration item riding the next manifest bump. Bilingual ckpt: `.cache/phase1_bilingual_head.safetensors`. Priors scored: relay 70(a) / Code 65(a) — (a) won; the discovered IR needed no translator, only literature. - **THE FORK'S HONEST NUANCE (2026-07-10, relay):** the outcome landed BETWEEN the sentences: 1.000 decodability + a blind zero-shot head is **(b)'s diagnostic signature**, cured by **(a)'s remedy** — the ledger line is *(b)-diagnosed, (a)-cured*. The relay's tiebreak (union head vs narrative structure) partially collected: the head WAS the bottleneck, just one that two thousand pairs dissolved. THE PRECEDENT THAT MATTERS: if real MATH-500 prose someday shows the same signature (decodable states, blind head) but RESISTS paired training, that is the residual (b/c) world announcing itself — and tonight names the probe to point at it. DISCIPLINE: 600/600 = zero-numerator ("error rate bounded below ~0.5%", not "perfect"); the diminishing-dividends clock TRANSFERS to the register axis (increments-per-register expected to shrink as registers accumulate — the watch is standing). NEXT MOVE: the recalibration bump (verbose training moved the native family — the point; threshold refits, joins the manifest) + LOG MATH-500's distance under the NEW calibration as the roadmap's first official gradient datapoint. The month in one sentence: the dancer could always hear the second language; someone just needed to read to her. - **GEN-5 COMMITTED (2026-07-10, the first SCRIPTED transactional bump — commit path + FOUR abort witnesses: archived-import, user kill, OOM, and the staged injection, all holding clean at the prior generation):** bilingual parser promoted; specialist v5 (1140 mined across FOUR registers); 7-kind centroids; mouth recalibrated. **THE GRADIENT READ: 3% CLOSED — below even the mostly-local band.** The staircase is STRICTLY LOCAL: generator-verbose moved essentially nothing toward MATH-500 (0.209 -> 0.204 over-threshold; refusal still 100%). Verdict: December's books must be drawn from or imitate REAL math prose — widening the generator does not walk toward the target register. **THE FIRST MIXED-SIGN EXPANSION:** bars bigtest 926 (+11), alg4test 336 (+5), vtest 600 (new capability) — but **alg2test 541 -> 507 (-34)**: five expansions paid uniformly, the sixth paid unevenly — plausibly the crossover watch's FIRST SIGHTING, on the register axis, at the middle domain (where the registration said to look). PROMOTION CAVEAT: the freeze's full acceptance (four lattice dials on alg2test under gen-5) is the standing next read — if certification held, the dip is one-shot-only and the promotion stands; if a dial bent, gen-5's parser choice gets revisited (gen-4 intact alongside, one manifest edit away — the transaction's whole point). 8 artifacts pinned incl. mouth threshold + views paths. - **THE MORNING'S LEDGER LINES (2026-07-10, relay):** (1) THE COSTUME MECHANISM: generator-verbose taught the head to parse OUR SKELETONS WEARING NARRATIVE CLOTHES, and the mouth correctly refused to count costumes as a language — the style gap lives in distributional properties (sentence rhythm, referential habits, framing conventions) no template dressing imitates; the mouth's first lesson, now with a slope attached (3%, noise-adjacent). December's unit of work: **mouth-distance closed per book.** OPEN DESIGN QUESTION for the first book: harvest-and-annotate real problems vs LLM-imitated register — the second needs an §8 honesty note if used (imitated style is itself a distribution). The paired- register machinery transfers whole (real prose paraphrased INTO the dialect = the same free triples, authentic style on the left). (2) THE PROTOCOL'S SENTENCE: four aborts from four different directions (archived import, user kill, OOM, injection), four clean holds, one commit — not a script that worked; A MECHANISM WITH AN EVIDENCE FILE, graduated in one morning. (3) FORK REORDERED by the gradient's mandate: real-prose books = the critical path to the anchor's re-examination; tranche 3 orthogonal, interleaves; the paper's §8 gains its final sentence when the first real book moves the mouth's needle. The promotion-caveat dials fire FIRST (the crossover's registered signature location); pre-committed remedy if sighted: curriculum weighting toward old kinds, never tranche reversal. - **THE PROMOTION-CAVEAT VERDICT (2026-07-10): DIALS HELD — GEN-5 STANDS CLEAN; CROSSOVER UNSIGHTED.** alg2test under the bilingual parser: certification 0.511 @ **1.0000** (gen-4: 0.516 @ 1.0000 — held); t=3 0.645 @ 0.9942 (softened, above bar); per-kind on the oldest relations 0.558/0.496/0.477 vs 0.561/0.500/0.485 — ALL within noise, no old-kind bend. The registered crossover signature did not arrive: the -34 one-shot dip is ABSORBED AT THE LATTICE LEVEL (the composed layer doing precisely its job), banks as variance with an honest asterisk on the sixth expansion, and the watch resets, still armed. The freeze's FOURTH acceptance table passes; gen-5's promotion is clean, not annotated. THE BOARD: the fork is Bryce's — real-prose books (critical path; mouth-distance per book; sourcing question open), tranche 3 (orthogonal, interleaves), the paper (§8 awaiting its final sentence from the first real book). - **THE EXPLICITATION FORK, REGISTERED (2026-07-10, Bryce's revisit + relay + Code):** the IR fork answered STYLE (no layer needed); the reopened question is EXPLICITATION — real prose withholds facts the reader must MANUFACTURE ("a dozen split among her three children, keeping twice as many" = producing 12, a fourth share, an unwritten multiplication before anything binds). THE KEY REFRAME: the discovered IR is precisely the ALL-FACTS-EXPLICIT fixed point — prose->dialect = explicitate then bind; the fork cleared binding; explicitation is GENERATION-shaped, the boundary where (c) could genuinely fire. CODE'S CLASS SPLIT (changes the probe design): LEXICAL implicits (dozen=12, twice=x2) have evoking-phrase SPANS — the existing decodability probe transplants; STRUCTURAL implicits (unstated shares, conservation relations) have NO anchor — need pooled/ query-style reads, a different instrument. THE PROBE: ~20 hand-annotated MATH-500 problems (annotation = dialect rewrites, not graphs — cheap gold), implicit-fact decodability at L0-L3 vs L0-L7. JURISDICTION CAUTION (relay): the old L8 refutation measured ROUTING on our dialect — world-knowledge inference is a FRESH depth question. OUTCOMES: shallow -> pairs cure it; deep-only -> the deeper-prefix conversation reopens with a real customer; nowhere -> the explicitation stage earns its build as the funnel's first GENERATIVE layer (structural-facts organ if the class split holds — smaller than feared). PRIORS: relay 40/35/25; Code 45/25/30 with MIXED (lexical-shallow, structural-stage) as the tiebreak. THE HARVEST GATE (the design's gem): real answer keys make solve-to-official-answer the round-trip gate for harvested books — prose -> dialect -> graph -> solve -> match key; dialect checkable TWICE (mouth-v2 conformance + end-to-end) — the layered-mouths architecture arriving with its validation story written. First work item of the real-prose chapter; its verdict decides whether December's books teach a reader or train a translator. - **EXPLICITATION PROBE READING FRAME (2026-07-10, relay, pre-print):** (1) n=9 STRUCTURAL BANDS pinned: 8-9/9 = manufacturing plausibly done; 4-5/9 = mixed, the class split becomes the finding; 0-2/9 = stage question live. Nothing subtler is readable at this n; zero-numerator discipline applies BOTH directions; the probe decides the FORK, not the magnitude (magnitude waits for the harvest chapter's larger set). (2) POOLED-READ CONFOUND status: Code's probe targets COMPOSED VALUES (digits of 96/28/360), not relation-presence — a linear probe cannot multiply ingredients, so a decoded composition is evidence of computed-by-trunk rather than aggregated-by-probe; the caveat softens but n=9 noise dominates — bands govern. (3) Whichever lands, §8's architecture paragraph gets its cost word: DATA (pairs cure), DEPTH (the frozen slice's first real customer post-L8-fencing), or ORGAN (the funnel's first generative layer, pre-sized structural). Either answer flatters the method: she hears it = the frozen-trunk bet's biggest dividend; she doesn't = the fork caught it for twenty problems before a book was harvested against the wrong architecture. - **EXPLICITATION PROBE VERDICT (2026-07-10): INSTRUMENT-LIMITED, DIRECTIONAL STAGE-WARD.** Lexical 0.25 (L0-3) / 0.31 (L0-7); structural **0/9 both depths — the 0-2 band fired, stage question formally LIVE** — but the TRANSFER CAVEAT is load-bearing: the probe reads written-digits- at-spans; lexical implicits are semantic ("octagon" carries 8 unwritten). WHAT IS CLEANLY ESTABLISHED: implicit values, if present, are ENCODED DIFFERENTLY than written ones — the same linear map does not transfer, so explicitation is not representation-free; the pairs-cure-it world, if it exists, is not the trivial version. Depth gain (one fact) = noise; the scripted LOO sanity was noted but NOT implemented (honest flag). THE FOLLOW-UP INSTRUMENT, specified by the failure: a SYNTHETIC LEXICAL-IMPLICIT corpus (the generator emitting dozen/twice/number-words with gold — free by construction) to train a probe whose task matches the question, then re-read MATH's lexical set. The fork holds its verdict until that instrument reports; December's budget word (data / depth / organ) waits with it. Twenty problems, one afternoon, and the question sharpened twice — the method's economics intact. - **THREE REGISTRATIONS BEFORE THE SYNTHETIC RE-RUN (2026-07-10, relay):** (1) THE TRANSFER FAILURE PROMOTED FROM CAVEAT TO FINDING: implicit values live in a DIFFERENT ENCODING than written ones — the trunk does not hallucinate the token "12" onto "dozen" in written-value geometry. The strongest decodable-shallow form ("the fact is just there, same shape as written") is DEAD; the surviving question is finer: different coordinate system (probe-trainable — what the synthetic corpus tests) vs uncomposed ingredients (stage-ward). The re-run DISCRIMINATES; it is not a do-over. (2) THE PHRASE-SPLIT PIN: train and test implicits must not share evoking phrases (dozen/score/fortnight train; twice-as-many/split-evenly/ days-of-month test) — a passing probe then reads ENCODING GEOMETRY, not vocabulary trivia. Within-phrase-pass/cross-phrase-fail = the mixed verdict in sharper clothes (lexical implicits are dictionary lookups the head learns from pairs; the general question stays open for structural). Enforced at mint time, free. (3) §9 META-NOTE — the day's cleanest exhibit: a probe returned readable-looking stage-ward numbers and the instrument-validity check caught the untrained examiner BEFORE the verdict banked. Most projects ship "implicit facts aren't decodable"; the ledger shipped "our probe can't yet distinguish" — smaller, true, one afternoon to fix. THE PRINCIPLE: **the verdict that flatters your architecture hypothesis needs the same instrument scrutiny as the one that flatters your hopes.** QUEUE: synthetic corpus (phrase-split pinned) -> probe retrained -> fork verdict -> December's budget word. - **THE EXPLICITATION FORK RESOLVES: STAGE-WARD, WITH THE DICTIONARY NUANCE (2026-07-10).** Retrained examiner (phrase-split): within 1.00 (instrument VALID — the negative is trustworthy), CROSS-PHRASE **0.00/0.06** at both depths, MATH-lex 0.00. **There is no shared evoked-quantity geometry in the frozen trunk** — "a dozen" does not light a magnitude direction that "a baker's dozen" also lights; implicit values are not precomputed anywhere probe-readable, shallow or deep (the deeper-prefix customer never materialized — fenced L8 verdict extends to inference). Combined with structural 0/9: EXPLICITATION IS REAL WORK THE TRUNK HASN'T DONE. THE NUANCE (the within column's gift): lexical implicits are DICTIONARY-LEARNABLE per phrase — a finite lexicon the generator enumerates into pairs (the selector-vocabulary pattern); novel evokers and ALL structural facts need the GENERATIVE ORGAN. **DECEMBER'S BUDGET WORD: ORGAN — sized structural, with a lexicon appendix.** The funnel's first generative layer earns its build honestly: prose -> explicit dialect, validated twice (mouth-v2 conformance + solve-to-official-answer on harvested books). Three probes, two afternoons, twenty hand-annotated problems — and the architecture question that opened as a metaphor closed as a measured build order. - **THE ORGAN'S CHAPTER CHARTER (2026-07-10, relay — banked before the word):** (1) THE CAVEAT AT CORRECT WIDTH: what died is LINEAR decodability — "not probe-readable," not "not present"; but the operational claim survives via THE CONTRAST CLASS: the same probe family read verbose states at 1.000 and evoked values at 0.00 — the bilingual precedent is what makes the negative meaningful. The organ builds against the strongest available instrument, limits stated. (2) SCOPE RESIZED FINAL: lexical implicits DON'T NEED THE ORGAN — head-side dictionary via pairs (the selector-vocabulary pattern); the organ's true scope is STRUCTURAL manufacturing only (the unstated share, the unwritten conservation) — boundary now measured, not suspected. (3) THE CHARTER DECISION everything inherits: **the organ WRITES DIALECT TEXT, not graph deltas** — appended explicit sentences re-entering the funnel unchanged: costs a re-encode, buys mouth-v2 conformance on the intermediate, byte-identical parser/lattice below, and the harvest gate's double-check. Graph-delta output = faster and unverifiable, a silent-error species upstream of every jaw with no examiner between it and the solve. The funnel's first WRITER is its most-audited citizen. Full birthright by standard practice: pointer law for bindings, two-checkpoint if it repairs, taxonomy tier for its error species, manifest citizenship from checkpoint one. (4) SEQUENCING: **THE HARVEST COMES FIRST regardless** — real prose with official answers is simultaneously the organ's training substrate (paired triples, the n=20 annotation pattern proven), its examiner (solve-to-answer-key), and the mouth's odometer corpus. December opens with books whichever way the organ's details settle. Eight instincts, the funnel counted twice: it named the missing mouth, then the missing writer. - **THE HARVEST OPENS (2026-07-10): 1,743 in-reach problems from the MATH TRAIN split (disjoint from the examiner). ODOMETER ZERO-POINT: 0.2488 — statistically identical to MATH-500's 0.2480: the harvest is a VALID PROXY (closing distance to these books closes distance to the benchmark). The level gradient REPLICATES the anchor's inversion (L1 prosiest/most foreign 0.270 -> L5 most symbolic 0.231) — the prose-style mechanism confirmed on a second corpus. Corpus: `.cache/math_harvest_v0.jsonl`; states banked. NEXT: the seed annotation — dialect rewrites gated by solve-to-official-answer, the harvest gate live from annotation one. - **THE HARVEST GATE'S FIRST DAY (2026-07-10): 0/5 banked — AND THE ZERO IS THE SYSTEM WORKING.** Seed dialect rewrites of real MATH-train problems, gated by solve-to-official-answer: all rejected, zero false banks. The rejections are diagnostic: seed[1] (sum/diff/lesser — the wild Vieta+sel) parsed near-perfectly with ONE selector-arg pointer off; seed[2] exposed the real lesson — MY DIALECT WAS OUT OF SHAPE (3-var 5-sentence miniatures; the training distribution starts ~10 vars) and the parser's bindings wobble on miniatures + far-OOD values (900/841 vs corpus m=60). TWO NAMED FIXES for seed round 2: (i) annotate IN-SHAPE (corpus-sized preambles/lengths, values in-range where possible); (ii) THE GATE SHOULD BE THE LATTICE, not one-shot — seed[1]'s single unstable pointer is what the 5-view vote exists to fix. The harvest stands: 1,743 in-reach problems, odometer zeroed at 0.2488 (== the benchmark — valid proxy), level-inversion replicated, seed machinery live and correctly strict. Scripts: `harvest_v0.py`, `harvest_seed_gate.py`. - **DAISY CHAINS + MATH KNOTS (2026-07-11, Bryce's ninth — two metaphors, both machinery):** (1) THE CHAIN-PRECISION BUDGET, registered BEFORE the organ trains: with the writer live, the pipeline is a true chain (organ writes -> parser reads -> solver carries) and chain precision is a PRODUCT — a silent error at link one looks native to every instrument below. The organ's certified-write bar is therefore DERIVED, not chosen: **organ certified-error budget = (end-to-end 0.25% bound) minus the parser link's measured certified error at integration time** — pinned as a FORMULA (current estimate: parser consumes ~0.15-0.2%, leaving the organ ~0.05-0.1%, i.e., certified-write precision ~99.9%). The layered mouths are the per-link circuit breakers (mouth-v2 on the write, lattice on the parse, answer key on the solve): three examiners, three links, no silence propagates — the daisy chain stays floral, not electrical. (2) THE KNOT FRAMING, adopted into the certification section: **reading is isotopy; manufacturing is surgery.** The 3% style test was an unknot (ambient-isotopic to the dialect — pairs taught the deformation); structural implicits are true knots (no smooth rearrangement yields the unstated share — cut and re-glue = generation; the organ is the funnel's first surgeon, hence the operating-room protocols). TTA permutations are REIDEMEISTER MOVES — diagrams of the same knot — so agreement is a KNOT INVARIANT, and the anchor's failure has its theorem-shaped sentence: on foreign text the parser computed a diagram-dependent quantity; five diagrams voted unanimously for the wrong knot. **CERTIFY INVARIANTS, NOT DIAGRAMS** — the mouth guards the language, the vote guards invariance, the answer key guards the ultimate invariant; each mouth in the chain guards a different one. §8's framing upgraded from wound to theorem. - **FIRST HARVEST GOLD (2026-07-11): n=1 BANKED.** Seed[0] (MATH-train: "sum 45, difference 3, lesser number") passed the lattice gate — 3 forced views, unanimous 21, == the official answer key. A real problem, hand-explicitated, machine-verified end to end: **the organ's training substrate exists**, and its first entry is the corpus's own Vieta+selector pattern found in the wild. THE BOUNDARY the rejections drew: all three-digit-given seeds (900/841/289/225) parsed to nothing across all views — the digit head's trained range (values <=60) is a hard wall; FIX (one line): mint larger given-values into the next training mix, widening the harvestable slice. Seed[1]'s in-range rejection: hand-dialect still drifts from template phrasing — round 3 goes template-exact. Substrate: `.cache/harvest_seed.jsonl` (gate=lattice-vote+answer-key). - **THREE REGISTRATIONS AT THE LIBRARY DOOR (2026-07-11, relay):** (1) THE TEACHER-DEMONSTRATION FRAMING: the annotation flow is behavioral-cloning substrate in the strictest sense — every banked entry is a worked example of explicitation-as-surgery, machine-verified by the same gate the organ will face; the surgeon trains inside an already-certified operating room (no prior component got that inheritance). (2) GENERATION-INDEXED GOLD: "write in the generator's voice" means THE CURRENT GENERATION'S voice — the dialect's boundaries move as mixes widen, so every banked entry cites the manifest generation it was written against (sync law applied to prose style; a field now, a vintage-mismatch mystery prevented later). (3) THE BOOTSTRAP DESIGN, contingent on round-3 economics: the bilingual head's own parses of NEAR-NATIVE harvest problems can propose dialect rewrites, THE GATE DISPOSES — propose/dispose eating its own tail; hand surgery reserved for the knotted cases. The gate makes the old self-improvement dream safe: nothing banks that doesn't carry to the answer key, no matter who — or what — wrote it. - **ROUND THREE'S FINDING (2026-07-11): THE SHAPE BOUNDARY.** Garden seed rejected AGAIN in template-exact voice (votes empty across all views) — reproducible, so structural: the dialect sentences are in-distribution but the GRAPH SHAPE is not (unknown-first mul chain, a in three factors, implicit a+2a=30) — our generator mints chains-from-knowns and Vieta pairs; the parser learned those SHAPES, not free composition of its relations. **Sentence-level native, graph-level foreign** — the third boundary the gate has measured (after value-range and voice). Midpoint (FDIV meets prose): one view forced wrong — same shape story on an 11-var double-fdiv chain. THE FIX IS THE STANDING LEVER: SHAPE DIVERSITY in the next generation's mix (random DAG compositions, unknown-first chains, reused intermediates) — a corpus change, and the dividends law predicts it pays everywhere, not just at the gate. Substrate holds n=1; the gate's boundary map now has three walls (values, voice, shapes), each with a one-generation fix. The harvest is teaching the generator what the wild actually looks like — which was always the point of books. - **GEN-6 CHARTER: TEACH THE MOVES, NOT MORE DIAGRAMS (2026-07-11, the knots talking — banked before the word):** the parser learned to recognize the specific knot diagrams the generator printed, not knot theory; it memorized diagrams, not Reidemeister moves. GEN-6's objective at the right abstraction: RANDOM DAG COMPOSITION — sample the WIRING, not fixed architectures; no finite diagram set can be memorized, so the parser is forced into compositional binding. Discovered-beats-designed pointed at the corpus itself: the graph shapes were THE LAST DESIGNED DECOMPOSITION hiding in the pipeline, invisible from inside because every internal eval sampled the same phrasebook — found only by real text refusing to fit. SHARPENED DIVIDENDS REGISTRATION (a frame-change, not a content-add): prediction — dividends hold and run LARGER than the register expansion's (composition is what the circuits compute; old shapes become easy special cases of a general skill). FALSIFIER: if capacity was sized for diagram-memorization, shape-mix COSTS old-shape fac-exact while buying wild-shape generality — the lattice table catches it in one read; that outcome prices "head grows before corpus does." DESIGN PIN: shape diversity extends to QUERY AND MENTION STRUCTURE — unknown-first chains put references BEFORE definitions (a genuinely new binding pattern; the pointer law's history says the risk concentrates exactly there); the generator samples query position + mention ordering as part of the shape. ACCEPTANCE PROBES FOR FREE: the three rejected harvest seeds — the wild sentences that named the walls bank when the walls come down. The month compressed: anchor found the language boundary, probes found the inference boundary, harvest found the composition boundary — each wall named by real text, each fix one generation of existing machinery. The phrasebook becomes a grammar. - **GEN-6 VERDICT TREE + BOOTSTRAP RE-PRICING (2026-07-11, relay, banked while the grammar lessons burn):** (1) THE INTERPRETATION HAZARD PINNED: the garden banking is n=1 CLOSURE (demonstration), not the class-level measurement — that lives in the wild-shape test row. Three sentences ready: garden banks AND wild-shapes strong AND old shapes hold -> GRAMMAR (dividends at larger-than-registers, registration confirmed); garden banks but wild-shapes soft -> the wall thinned where probed (phrasebook grew a page; gen-7 widens wiring variance); old shapes regress -> the CAPACITY TRADE (head-growth conversation opens with a measured invoice). The house pattern applied to its own celebration. (2) THE BOOTSTRAP RE-PRICES the moment the wall thins, regardless of verdict: post-commit probe = 20 cheapest un-annotated harvest problems (registry-wearing-prose, in-range), THE HEAD PROPOSES dialect, THE GATE DISPOSES. >=1/3 banked machine-proposed -> substrate accumulation goes machine-priced overnight; annotation hours redirect to the knotted cases — the machine drafts the isotopies, the human performs the surgeries. (3) The loop named: wild text -> rejection -> named wall -> generation charter -> re-examination — the oldest self-improvement design, running at generation cadence with the gate as incorruptible oracle. - **GEN-6: THE GRAMMAR SENTENCE FIRES (2026-07-11).** All three verdict-tree branches land on the best caption: WILD SHAPES **563/700 = 80.4%** one-shot (the highest debut ever; graph-solve 77%); OLD SHAPES SURGE — bigtest **1000/1500** (+74; legacy started at 802), alg2test **551** (+44 — the gen-5 register dip ERASED retroactively: transient, cured by moves), alg4test 371 (+35), vtest 600/600 held; val **0.8860** (prior best 0.8343). THE SHARPENED REGISTRATION COLLECTS IN FULL: dividends larger than the register expansion (+74 vs +11) — composition IS what the circuits compute; the old fixed shapes became easy special cases of a general skill, as predicted. THE GARDEN BANKS — 4/4 unanimous at 200: the wild sentence that named the wall walked through it one generation later; substrate n=2. Midpoint honestly still out (double-FDIV chain; fdiv absent from the DAG rotation — gen-7's one-line addition). The loop's first full cycle is complete: wild text -> rejection -> named wall -> charter -> re-examination -> BANKED. Ckpt: `.cache/phase1_gen6_head.safetensors`. NEXT: the bootstrap re-pricing probe (registered) — the wall is thin; let the machine draft. - **THE SIXTH EXPANSION'S ATTRIBUTION CLOSES (2026-07-11, relay):** the gen-5 alg2test dip (-34) was never diminishment — it was THE DIAGRAM-MEMORIZATION TAX being collected (register training crowding memorized shapes), refunded by gen-6 with interest (+44, past the pre-dip bar). The dividends streak reads SIX-FOR-SIX with one payment a generation late; the crossover watch's first candidate closes fully attributed, and the watch stands armed with a cleaner prior: mixed signs under frame-changes may be taxes, not crossovers — distinguish by whether the NEXT frame-expansion refunds them. - **THE BOOTSTRAP CENSUS (2026-07-11): banked 2 / near-miss 26 / knotted 72 (n=100, L1-3 in-range prose).** The bootstrap does NOT re-price yet — and the honest read of the 72%: it CONFLATES TWO WALLS (the mouth reads the whole pool as foreign at 0.254, so structural refusal may be style-unreadable rather than explicitation-needed). The census bounds the organ's kingdom FROM ABOVE; its true survey waits on the style wall. SEQUENCING VERDICT, the funnel's original order reasserted with numbers: **BOOKS BEFORE SURGEON** — the machine cannot draft what it cannot read; December's real-prose corpus precedes organ economics; the census re-runs after each book and its knotted column's SHRINKAGE measures what reading alone recovers vs what only surgery can. THE MILESTONE IN THE SMALL COLUMN: two machine-made triples banked — the first substrate entries produced end-to-end without a human hand (raw prose, parsed, voted, verified to the author's answer; substrate n=4: 2 hand + 2 machine). Diversity guard: no narrowing signal (banked prose reads MORE foreign than pool — n=2, informational). The loop touched its tail tonight — barely, and genuinely. - **REGISTERED (2026-07-11, relay): THE MOUTH-DISTANCE JOIN — first partial attribution of the 72.** The census's knotted column conflates two walls; a refusal can be style (mouth's territory, books cure) or structure (organ's territory). The join: cross every census outcome with its prose's mouth kNN-distance (one deterministic census re-run capturing both columns — labels weren't persisted; same pool slice, same view seeds, same gen-6 ckpt). PRE-REGISTERED DECISION RULES, pinned before measurement: (1) **AUC** (rank-sum) of mouth distance separating KNOTTED vs PARSE-CARRIED (banked+near). AUC ≥ 0.60 → style-wall attribution holds (refusals concentrate where the prose reads foreign; the relay's prediction: books recover the high-distance tier). AUC ~0.5 → the 72 stays UNATTRIBUTED — an honest negative that also weakens the books-will-recover prediction, since refusal would then be independent of readability as the mouth measures it. (2) **THE PATIENT LIST**: knotted items at mouth distance ≤ the carried group's MEDIAN are the early knotted candidates — the organ's first genuinely visible patients, named and counted tonight rather than after three books. (3) **BOOK-1 FALSIFIABLE PREDICTION**: knotted items ABOVE the carried median are claimed style-recoverable; after the first book, their recovery rate must exceed the below-median tier's or the attribution was wrong. Threshold note: the calibrated native line (0.0443) is unusable here — the entire pool reads foreign (mean 0.254) — so the split is relative (carried-median), pinned now. Zero new training; one eval-cost re-run. - **VERDICT (2026-07-11): THE JOIN RETURNS THE HONEST NEGATIVE — AUC 0.535, the 72 stays UNATTRIBUTED.** Census replay exact (2/26/72; deterministic seeds held). Knotted mouth distance mean 0.2560 / median 0.2499 vs carried 0.2491 / 0.2354 — indistinguishable. MOUTH DISTANCE DOES NOT PREDICT REFUSAL: the mouth measures REGISTER (surface style, corpus-level), not per-item parseability. Its odometer role survives untouched (it was chartered corpus-level); its per-item attribution ambition dies tonight, pre-registered. THE SENTINEL IN THE PATIENT LIST: idx-21 ("sum 45, diff 3" — the problem seed-2 BANKED in dialect, structure PROVEN in-reach) sits at d=0.2065 INSIDE the low-distance tier as a raw-prose refusal — a certified style-only casualty at near-carried distance, demonstrating the tiers mix in both directions, which is exactly what AUC 0.535 says. Rule-2's list (32 items) is therefore NOT a patient roster — reading it confirms the mix: quadratic factoring / completing-the-square / geometric sequences (true organ patients, moves outside the grammar) interleaved with in-grammar prose like idx-21. Rule-3's book-1 prediction is WITHDRAWN with its premise (registered falsifiable, falsified at the instrument stage). THE CONSTRUCTIVE RESIDUE: the per-item attribution instrument already exists and is the BOOK ITSELF — idx-21 proves the protocol (raw refuses + hand dialect banks = style casualty; dialect also refuses = organ patient). Every book annotation doubles as an attribution measurement; the census's knotted column will be attributed item-by-item as the books ship, not by any cheaper proxy. BOOKS BEFORE SURGEON stands, now with its own attribution built in. Data: `.cache/census_mouth_join.json` (n=100, census + mouth_d columns); script `scripts/census_mouth_join.py`. - **THE JOIN'S THREE PERMANENT READINGS (2026-07-11, relay — binding on book 1):** (1) **THE MOUTH'S JURISDICTION IS MEASURED, NOT ASSUMED** — it is a CORPUS-REGISTER instrument (odometer charter intact: corpus distance is what it was calibrated on); its per-item ambition died at a pre-pinned bar because refusal has at least two causes and the mouth sees one axis. The jurisdiction law applied to the project's own newest organ within a week of its birth: instruments don't inherit resolution they weren't calibrated for. The mouth recognizes languages; it doesn't diagnose readers. (2) **IDX-21 IS THE ANCHOR SENTINEL** — structure certified in-reach by its banked dialect twin, raw prose refusing: the style wall isolated in a single specimen, the existence proof that some fraction of the 72 is books-recoverable, and the FIRST ENTRY in the paired-(raw, dialect) format book 1 will systematically produce (the pair's fate attributes the refusal). One problem demonstrated the whole protocol before the protocol was named. (3) **BOOK-1 SAMPLING IS STRATIFIED, BY DESIGN, NOW** — the book is substrate AND census resolver, and the dual role forbids drifting toward cheap-tier-only annotation (substrate throughput) or spread-only (attribution): deliberately spend annotation budget across the refusal spectrum, INCLUDING suspected organ patients (quadratic factoring, completing-the-square, geometric sequences) — a both-refuse verdict on those is the organ's customer list getting its first confirmed names at one annotation each. META-NOTE (the method's signature, performed on its own instrument): the join fired BECAUSE it was registered to be allowed to fail; its failure conscripted the book into a census the roadmap didn't know it needed — negatives that conscript existing work into new instruments (the anchor conscripted the mouth into existence; the join conscripted the book into a census). Book 1 awaits the word; it ships with its own index. - **REGISTERED (2026-07-11): BOOK 1 — the library's first volume, n=18, stratified per the binding sampling law.** Paired (raw, dialect) protocol: both run through the gen-6 lattice gate (5 views, vote>=3, answer key disposes); the pair's fate attributes. ROSTER — TIER N (near-miss, 6): idx 71, 78, 89, 72, 46, 7. TIER S (knotted, style-suspect, 7): idx 21 (sentinel, entry one), 99, 16, 57, 28, 56, 45. TIER O (knotted, organ-suspect, 5): idx 54, 90, 51, 37, 85. TAXONOMY (pinned): STYLE CASUALTY = faithful in-grammar dialect banks (lexical explicitation allowed: literal facts about KNOWN quantities, e.g. 4^2=16, 15th-term->14 steps; supplying an UNKNOWN's value or a rewritten equation is forbidden). ORGAN PATIENT type A = annotation-impossible (no faithful in-grammar dialect; move outside grammar — [90] completing-the-square declared type A up front, its residual runs for the record only). Type B = faithful dialect exists but refuses. ANNOTATION-TIME FINDING (before any GPU): the organ-suspect tier SHRANK under the pen — integer-root factoring IS the tranche-2 Vieta shape (sum+product+selector, in grammar: [54]); function composition UNWINDS into forward relations the CSP inverts natively ([51]); sign-rewrites reduce to positive-form relation sets ([37]). Only [90] resisted. The organ's kingdom is narrower than the census's qualitative read suggested — quadratics with integer roots were annexed by tranche 2 before the organ was chartered. PREDICTIONS (pinned): tier N >=4/6 bank; tier S >=5/7 bank (incl. the sentinel) — style-wall existence at scale; [54] and [37] BANK, [85] REFUSES (repeated-arg mul untrained — type B), [51] uncertain (depth); book-1 substrate growth >=10. Gate: gen-6 ckpt, solve2 n_vars=24 m=300, fdiv at most once per item (double-fdiv is gen-7's known wall). - **VERDICT (2026-07-11): BOOK 1 CLOSES — 15 entries, the style wall confirmed at scale, the organ's kingdom shrunk to a named list.** V1 table: tier N 1/6 (prediction >=4/6 FAILED), tier S 5/7 (prediction EXACT, sentinel 5/5), tier O 4/5 banked with all four specific predictions correct ([54] Vieta BANKED 3/5, [37] 5/5, [51] 4/5, [85] refused). V2 (one taxonomy-faithful retry each, pre-declared): 5/6 recovered, all 5/5 unanimous; [72] stands. FINAL ATTRIBUTION (n=18): **9 STYLE CASUALTIES** (raw knotted + faithful dialect banks — 75% of the sampled knotted slice is books-recoverable; curated sample, upper tier), 5 FRICTION RECOVERED, 1 ORGAN-A ([90] completing-the-square), 3 ORGAN-B standing: [72] novel coupled-linear wiring, [56] 19-var length wall, [85] repeated-arg mul (a grammar GAP, not an organ move — gen-7 one-liner). THE TIER-N INVERSION: near-miss was NOT the cheap tier — its failures shared one cause, the FDIV REGISTER (0/4 in hand-written composition; alg4test's weakness confirmed in the wild), and v2's lexical-literal route recovered 4/4 of them. TWO LATTICE SPECIMENS FOR THE GOODHART FILE: [71] raw went 5/5 UNANIMOUS-WRONG (8 vs gold 9) — first observed; the recognition mouth is the organ that intercepts exactly this (reads the prose foreign upstream) — and [78]'s v1 dialect voted a consistent wrong 12/12/12 (a 3/5 ANSWER-channel error shape from the fdiv register). ONE REGISTER SENSITIVITY: [46]'s deep chain refused under "a plus b equals c" and banked 5/5 under "The sum of a and b is c" — same relations, surface flip. GEN-7 PRICED BY THE BOOK: (1) fdiv into the DAG rotation (5 items waited on it), (2) repeated-arg mul (unlocks number+square), (3) longer chains, (4) coupled-linear wiring, (5) surface-form robustness on deep chains. SUBSTRATE: volume 15 (14 faithful + 1 flagged residual; idx-21 double-banked by design as sentinel re-verification) — substrate n=17 unique. The book's raw prose is gen-7+'s reading-training target (raw -> gold graph from the banked dialect parse); the census re-prices after the ingest. Data: `.cache/book1.jsonl`, `.cache/book1_attribution.json`; scripts `book1_paired_gate.py`, `book1_v2_retry.py`. - **REGISTERED (2026-07-11): GEN-7 — the receipts generation.** Corpus: `algebra_dag7_gen.py` (fdiv/mod pairs in rotation, repeated-arg mul, 8-12-given ladder chains, coupled-linear k1x+y=s1 / x+k2y=s2 blocks; render3 + roundtrip3; smoke 40/40, kinds all present); mixed7 = mixed6 + 3500 dag7; warm from gen-6, CURRICULUM=1, 16k steps. PREDICTIONS (pinned before training): (a) ACCEPTANCE PROBES — the eight refused book-1 v1 dialects re-gated under gen-7: [46]v1 BANKS (the paraphrase probe), fdiv v1s [71,78,7,45] >=3/4 bank, [85] BANKS (sq now in grammar), [72] BANKS (coupled in rotation), [56] BANKS (ladder). (b) REGRESSION BARS (lattice holds its dials): dagtest6 graph-solve >=520/700, bigtest ANSWER >=980/1500, alg2test >=530/800, vtest 600/600 holds; alg4test RISES >=420/800 ANSWER (the fdiv receipt paying); val >=0.87. (c) CENSUS under gen-7: ~UNCHANGED (65-72 knotted) — gen-7 teaches MOVES in the dialect register; the style wall is untouched by design, and the employment-law enforcement census waits on READING-training (the next chapter), not this bump. (d) Mouth bank NOT rebuilt (register unchanged; rides the next register-changing bump). KILL: any regression bar broken -> gen-7 is NOT promoted to gate ckpt; gen-6 stays, diagnose before re-fire. - **VERDICT (2026-07-11): GEN-7 v1 — KILL FIRES, gen-6 keeps the gate; diagnosis crisp.** Bars held: dagtest6 ROSE 541->616 graph-solve, alg2test 551->559, census ~unchanged as predicted (1/28/71), [46] paraphrase probe BANKED, [56] ladder BANKED (5-view vote rescues a 0.27 single-view kind — TTA doing its job). Bars broken: bigtest 963 (<980), alg4test FELL 371->350, vtest 598, val 0.815; fdiv probes 0/4, [85] and [72] still refuse ([72] voted a stable wrong 120 — another Goodhart specimen). ATTRIBUTION (per-kind, dag7test single-view): plain 0.878 (composition circuits HEALTHY — not global undertraining) vs sq 0.264 / ladder 0.273 / fdiv 0.372 / coupled 0.417 — four new factor SHAPES warm-started at once on a skewed diet (fdiv saturated 60% of rows; ladder/coupled ~500 rows each) for 16k steps of a still-climbing val. Consistent with the §6 attention-bootstrap law (new pointer patterns need supervision time). ENGINEERING NOTE (3 kills before the chain ran): precompute held a 15.7GB states array in RAM beside the AM driver's pinned pages -> OOM during every write; root-caused to a disk-backed memmap on BOTH write and train sides (legacy npz path kept for gen<=6 artifacts). RE-FIRE REGISTERED (dag7b): quota-balanced corpus (~1200 per kind incl. plain, fdiv wiring de-saturated 16%->8%, sq 22%->15%), mixed7b = mixed6 + dag7 + dag7b, WARM from gen-7 v1 (val still climbing), STEPS=32000. REVISED BARS: per-kind single-view >=0.55 each new kind; acceptance >=6/8; bigtest >=980 (refund), alg4test >=420, vtest >=598, dagtest6 >=590, alg2test >=530. Same kill: any bar broken -> no promotion. - **VERDICT (2026-07-11): GEN-7B — KILL FIRES AGAIN; gen-6 keeps the gate; the diagnosis graduates from balance to CROWDING.** Val healthy monotone to 0.8736 (balanced test; new best every check from 12k). EVERY new kind improved substantially (ladder 0.273->0.500, fdiv 0.372->0.549, coupled 0.417->0.543, sq 0.264->0.342) yet ALL FOUR missed the 0.55 bar (two within noise of it). Acceptance 4/8 (was 2/8): [46] 5/5, [78], [45], [56] BANK; [71] and [7] now produce the RIGHT answer but vote-shy (votes [9,9] and [45] — parse instability across permuted views, not wrongness); [85] inverse-square still 0/5 (training sqs are FORWARD — a known var squared; the inverse shape, unknown-squared pinned only downstream, is likely absent from the rotation: a GENERATOR gap, not a training-budget gap); [72] coupled still refuses. Bars held: vtest RECOVERED 600/600, dagtest6 ROSE again (616->661 graph-solve; 541 at gen-6), alg2test rose 559->575. Bars broken: **bigtest 1000->963->901 and alg4test 371->350->319 — MONOTONE EROSION across both gen-7 rounds.** This is not the gen-5 tax shape (a dip refunded by the next training); the next training DEEPENED it. It is CROWDING: mixed7b diluted the original register to 55% and pct/seq (alg4's other two thirds) to zero new rows, and the oldest registers paid. The crossover watch logs its first candidate that is NOT a tax. CENSUS: 1/26/73 ~unchanged, as predicted. NEXT-FIRE OPTIONS (await the word — two kills in a row makes this a design decision, not a mechanical re-fire): (a) REPLAY MIX — mixed7c upweights the eroding registers (re-add algebra_nl + alg4 slices) alongside dag7b, retrain; (b) add the INVERSE-SQUARE and inverse-fdiv shapes to the rotation first (the [85] gap), then (a); (c) probe whether the 33-key head is at CAPACITY (param census vs absorbed registers) before spending more steps. Data: val curve + tables in the session log; ckpt `.cache/phase1_gen7b_head.safetensors` (NOT the gate; gate remains `phase1_gen6_head`). - **REGISTERED (2026-07-11, relay adjudication): (c) FIRST — THE CAPACITY PROBE; (a)/(b) sequenced behind its verdict.** The eroding-while-gaining signature (new kinds up, oldest registers monotone-down across two independent rounds, gen-5 tax explicitly fenced off by the second round deepening it) is the registered picture of the CAPACITY CROSSOVER, first symptom on the oldest relations as the diminishing-dividends registration predicted — three weeks early. Data fixes (replay mix, new shapes) are the lever we WANT to work (flattering-remedy principle): if the head is full, a replay mix doesn't cure crowding, it chooses different victims — weighting only works if there is room to weight INTO. PROBE DESIGN (banked machinery only): ALG_HW dial added (512 default); pad-warm gen-7b into a 2x head (1024), train on the IDENTICAL mixed7b corpus (states already precomputed) for 12k steps, read ONE number at fixed data mix. DECISION RULE (pinned): bigtest ANSWER >=960 AND new-kind per-kind >= gen-7b levels − 0.03 -> the wall is the head; growth opens with a measured invoice (head ~3.2M against a frozen half-billion — 2x is a rounding error) and (a)/(b) ride the bigger head. bigtest <930 -> capacity EXONERATED; crowding is a data-mix problem; (b)-then-(a) fires with confidence. 930-960 -> extend +8k once, re-read. PAD-WARM CAVEAT stated: the relational law says padded compositions may not inherit — a null result reads through that lens before exonerating capacity. RIDING REGISTRATIONS: (i) NEW SPECIMEN CLASS — CORRECT-BUT-UNCERTIFIABLE ([71] votes [9,9], [7] votes [45]: right answers, vote-shy across permuted views) — the mirror of raw [71]'s stable-wrongness; the lattice's geometry lives between the two (stability without truth / truth without stability). Parked with the K-dial / view-family question; these two are its acceptance probes if the class grows on wilder prose. (ii) [85]'s INVERSE SHAPES (unknown-squared, inverse-fdiv) ship in gen-8's rotation WHICHEVER branch wins — named generator gap, standing acceptance probe. The probe awaits the word. - **REGISTERED (2026-07-11, Bryce's gut + relay: THE PHYSICS TRIAD) — three instruments for the phase boundary; statistical mechanics of a fixed-capacity head under growing load.** (1) **INTERFERENCE MATRIX (GPU, rides the capacity probe):** crowding = destructive gradient interference in shared weights. Per-register gradient cosine matrix on the gen-7b ckpt (one batch per register, one backward each, pairwise cosines). Joint verdict table with the capacity probe: anti-aligned + erosion-reverses-at-2x = capacity wall (orthogonal subspaces need room); aligned + erosion = data starvation (replay cures); anti-aligned PERSISTING at 2x = genuine task conflict -> the §8.4 LoRA fallback ladder gets its first customer (per-register adapters). (2) **ERASURE-VS-SHARE CORRELATION (zero-GPU, fires now):** Landauer transferred honestly — erasure is never free; when task entropy exceeds head capacity the mix decides who pays. PREDICTION (pinned): per-register erosion gen-6 -> gen-7b tracks INVERSE mix-share of fresh rehearsal (alg4/pct/seq got zero new rows -> pays most; original nl diluted to ~55% -> pays next; dag-fresh registers gain). Holds -> the crossover gets its conservation law: at capacity, expansion pays dividends MINUS erasure. (3) **VOTE-ENTROPY COLUMN (zero-GPU pilot on banked book-1 votes):** the two specimen classes unify as BASIN DEPTH — view permutation is the thermal kick; vote entropy across views is a per-item effective temperature. Fourth lattice column chartered: CORRECT-BUT-SHALLOW (the retraining-target class — rehearsal deepens basins rather than teaching). Pilot on the persisted book-1/acceptance votes; full column when TTA outcomes persist per-item at census scale. GPU items ((1) + capacity probe) hold for the word; (2)+(3) fire on banked data now. - **VERDICTS (2026-07-11, the zero-GPU pair):** (2) **LANDAUER CHECK — DIRECTION CONFIRMED, LAW REFINED.** Composition recovered exactly (dag7 register 45.2% of mixed7b; nl-core 9.5%; tranches ~12% each). rho(fresh-rehearsal share, erosion) = +0.50, n=5 — right direction, short of the pinned "strongly positive," and the DEVIATION is the finding: alg2 IMPROVED (+4.4%) with zero fresh file-share because dag7 rows REHEARSE its kinds covertly (sel/mod/coupled live inside dag7 wiring), while alg4's pct/seq kinds appear NOWHERE in dag7 (true zero rehearsal -> worst erosion, −14%), nl-core's older surface forms partially shared (−9.9%), verbose flat (same relations, held). THE CONSERVATION LAW REFINES: erasure is ordered by UNSHARED-CIRCUIT rehearsal share, not corpus-file share — kind-level, not file-level. Actionable for any replay mix: rehearse KINDS, not files. (3) **VOTE-ENTROPY PILOT — the basin-depth read separates exactly as the physics said:** deep-correct H=0.000, shallow-correct H=0.846, deep-wrong H=0.212, refused H=0.116 (banked book-1 gates, n=36). Entropy cleanly separates SHALLOW from DEEP (0.85 vs ~0.1-0.2) and CANNOT separate deep-correct from deep-wrong — which is the quantitative restatement of why the chain needs the mouth AND the key: temperature is orthogonal to truth. Fourth column (correct-but-shallow) validated at pilot scale; full column when per-item votes persist at census scale. Scripts: `erasure_share_correlation.py`; pilot inline in session log. - **THE LAW NAMED + THE QUADRANTS PINNED (2026-07-11, relay):** (1) **CIRCUIT REHEARSAL, NOT FILE REHEARSAL** — the corpus economy's law: the erasure bill is charged per UNSHARED CIRCUIT; a register whose kinds live inside another register's shapes rides free (alg2 +4.4% at zero fresh rows). Replay design converts from mixing FILES to mixing COVERAGE: gen-8's diet is specified as a KIND-REHEARSAL MATRIX (which circuits each row exercises), deficit = the true-zero-rehearsal kinds (pct/seq, the −14% line items). The law also re-explains the dividends streak at depth: expansion paid BECAUSE new shapes covertly rehearsed old circuits — constructive interference through shared kinds — and broke exactly where sharing hit zero. One law, both signs. (2) **CORRECT-BUT-SHALLOW = FREE CURRICULUM SIGNAL**: rehearsal targets, self-identified by their own temperature — fold into replay to DEEPEN basins (cheaper than any new capability); epigraph + fifth-column candidate added to paper §7. (3) **THE JOINT TABLE'S FOUR QUADRANTS, all pre-written** (capacity probe x interference matrix): (anti-aligned, erosion-reverses-at-2x) -> capacity for orthogonal subspaces; (aligned, reverses) -> PURE capacity, state-counting only — remedy is the bigger head alone, no LoRA, no mix surgery beyond the kind matrix; (aligned, persists) -> starvation, replay cures; (anti-aligned, persists) -> genuine task conflict, §8.4's LoRA ladder gets its first customer. One run, four pre-written verdicts, no cell left to improvise. GPU pair (`cap_probe.sh` + `interference_matrix.py`) staged, on the word. - **VERDICT (2026-07-11, the GPU pair): THE MATRIX IS CLEAN AND NAMES THE MECHANISM; THE PROBE'S CAPACITY AXIS IS CONFOUNDED BY ITS OWN WARM-START.** (1) **INTERFERENCE MATRIX (uncontaminated — measured directly on gen-7b):** the old guard is mutually ALIGNED (nl-core / alg2 / alg4 pairwise +0.22..+0.26 — shared circuits, the dividends streak photographed) and **dag7 is ANTI-ALIGNED with exactly the eroding registers** (nl-core −0.171, alg2 −0.255, alg4 −0.263); verbose/dag6 orthogonal (verbose grad norm 0.10 — fully learned, cosines are noise). THE TWO-FORCE MECHANISM, both instruments agreeing: destructive gradient pressure from the dominant register, OFFSET by covert kind rehearsal where present — alg2 (anti-aligned BUT kind-shared) nets positive; nl-core (anti-aligned, surface differs) nets negative; alg4 (most anti-aligned + zero kind share) pays worst. Interference axis: ANTI-ALIGNED, definitively. (2) **CAPACITY PROBE — mechanically sub-930 (bigtest 809) but the pre-registered pad-warm caveat FIRES:** the 2x head at 12k is worse than its own warm source EVERYWHERE (val 0.8234 vs 0.8736; all per-kind down; alg4 208 — disruption again hitting the least-rehearsed register hardest) and still climbing at cutoff — the probe measured DISRUPTION RECOVERY, not capacity (the relational law's exact prediction for padded compositions). The capacity axis is UNREAD. QUADRANT: provisionally (anti-aligned, persists) — task conflict, the LoRA ladder's customer — but the clean capacity instrument is now registered for the word: **the fair A/B** — 2x head, SAME 32k schedule, SAME mixed7b, only width differs; erosion reversal at matched schedule reads capacity cleanly. Ckpt: `.cache/phase1_cap2x_head.safetensors` (probe artifact, not a gate candidate). The night's arithmetic: the matrix cost one backward pass per register and delivered the mechanism; the probe cost 12k steps and delivered a confound — the cheap instrument won. - **REGISTERED (2026-07-11, relay — the temperament frame + two riders):** (1) **THE ATLAS'S GATE-2 INHERITS ITS DATASET** (one line, no fire): the interference cosines + the kind-rehearsal matrix are an empirical similarity structure over registers/kinds — if it proves tree-metric as tranches accumulate (register families sharing circuit-ancestry), the delta-probe opens with evidence produced as a byproduct. Tripwire unchanged (watches centroid-margin shrinkage, not gradient conflict). (2) **THE TEMPERAMENT DESIGN (the eleventh instinct, musical keys):** registers are keys, kinds are notes, the interference matrix IS the circle of fifths (alg2-inside-dag7 = closely related keys; alg4 = the tritone). The 33-key head is tuned in EQUAL TEMPERAMENT — every key playable, every key compromised. The two remedies are the two historical tunings: grow the head = more strings, same temperament; the §8.4 ladder = WELL-TEMPERED — shared instrument, per-key accidentals. THE SHARPENING THE MUSIC BOUGHT: accidentals ONLY where the matrix shows dissonance — adapters for the anti-aligned pairs alone; the old guard's mutual +0.25 means they WANT one tuning, and adapters there would waste parameters fixing consonance. Sharper than per-register-everywhere, derived from a metaphor, consistent with ten prior instincts. (3) **THE FAIR A/B, staged with its reading frame pre-committed:** 2x pad-warm, FULL 32k schedule, identical mixed7b — only width moves vs gen-7b's own 32k. Honest residual asymmetry noted: gen-7b's warm source was same-width; cap2x's is padded — a clean-cold A/B is the escalation if this one reads ambiguous. RULE: bigtest >=960 -> capacity was the wall; <930 at matched schedule -> exonerated; PARTIAL (930-960, or reversal WITH anti-alignment persisting in the 32k head's OWN matrix — the matrix reruns on the A/B artifact) -> BOTH tunings at once: more strings AND accidentals for the dissonant keys, as the music predicted before the run. Per-kind guard unchanged (>= gen-7b − 0.03). Awaits the word. - **VERDICT (2026-07-12, the fair A/B — overnight): CAPACITY EXONERATED, CLEANLY; the quadrant resolves to STARVATION; the kind-rehearsal law called it.** The one number: bigtest **888** — below the 930 exoneration line at the MATCHED 32k schedule, and statistically the same model as the 1x everywhere: alg4 316~319, alg2 554~575, dagtest 677~670, vtest 599~600, dag7btest 419~420, per-kind all within the ±0.03 guard (sq 0.329, ladder 0.500, fdiv 0.521, coupled 0.551), val 0.8646 vs 0.8736. WIDTH BOUGHT NOTHING. (The 12k probe's disruption transient fully washed out by 32k — that diagnosis confirmed in passing.) THE DEEPER RESULT, from the A/B artifact's own interference matrix: **decorrelation without improvement** — at 2x the anti-alignment softened everywhere (dag7 vs nl-core −0.171->−0.099, vs alg2 −0.255->−0.113, vs alg4 −0.263->**+0.076**, sign-flipped) AND the old guard's mutual alignment dissolved (+0.25 family -> ~0.1/negative): the wider head spread registers into orthogonal subspaces exactly as geometry allows — and behavior did not move. GRADIENT INTERFERENCE IS A SYMPTOM OF SHARED-CAPACITY PACKING, NOT THE CAUSE OF FORGETTING: given orthogonal room, the gradients decorrelate and the erosion stays, because nothing about width changes WHAT IS REHEARSED. QUADRANT: (interference relieved, erosion persists) = STARVATION — REPLAY CURES. Among the physics triad, the entropy/rehearsal instrument called the mechanism; the interference instrument photographed a symptom; the capacity probe (run fairly) exonerated the suspect. CONSEQUENCES: head growth SHELVED (receipt: 888 at 32k) and the §8.4 LoRA ladder's customer WITHDRAWN (receipt: conflict-as-mechanism falsified) — both with evidence, neither by taste. GEN-8 = (b)-then-(a) exactly as sequenced: inverse shapes ([85]'s standing probe) + the KIND-REHEARSAL replay mix (rehearse KINDS, not files; deficit = pct/seq, the true-zero line items). The music's final read: the instrument needed neither more strings nor accidentals — the old songs had simply left the practice schedule. Rehearsal, not tuning. Ckpt `.cache/phase1_cap2x_32k_head.safetensors` (A/B artifact, not a gate candidate; gate remains gen-6). - **PROMOTED TO THE LAW FAMILY (2026-07-12, relay): "DECORRELATION WITHOUT IMPROVEMENT" is the publishable finding — gradient interference is the PHOTOGRAPH of registers packing into shared capacity, not the MECHANISM of forgetting; the mechanism is rehearsal, and orthogonal room does not change the practice schedule. The interference matrix keeps its diploma as a PACKING DIAGNOSTIC; its causal ambitions retire with a receipt. OPERATIONAL LAW: **when erosion appears, check the rehearsal ledger before the architecture** — cheap accounting beat expensive geometry (the entropy instrument called the mechanism for arithmetic on banked tables; the geometry instruments cost GPU-nights to photograph symptoms and clear suspects). Flattering-verdict principle collected on the junction's MOST seductive hypothesis: task conflict was the interesting story, LoRA the elegant remedy; the fair A/B declined both with numbers. Two expensive builds shelved by one overnight run. - **REGISTERED (2026-07-12): GEN-8 — the practice-schedule generation. Diet fully specified by three free instruments; credit clause applied (no double-pay for covertly-rehearsed kinds — sel/mod/coupled ride inside dag7).** THE RATIONS: (1) KIND-RATION — 3000 fresh alg3-register rows (seq/pct/fdiv, the true-zero deficit, −14% line item); (2) SURFACE-RATION — 2500 fresh nl-core rows (nl-core's erosion was surface-unshared per the matrix: the kinds survived, the PHRASINGS starved — two different starvation species, two rations); (3) SHALLOW-BASIN RATION — DEFERRED to gen-9 (needs per-item TTA vote-entropy at corpus scale; registered, not forgotten); (4) INVERSE SHAPES — dag8 rotation adds isq (mul(a,a) with a UNGIVEN, [85]'s circuit) + ifdiv (dividend pinned by quotient+remainder), nogive mechanics in the givens-gate; smoke 30/30. mixed8 = mixed7b + 8500 = 29,500 rows; warm from gen-7b; 32k steps. BARS (pinned): **bigtest >=980 — the starvation thesis's own falsifiable claim: if the replay diet does not recover bigtest, starvation is wrong too and the junction reopens**; alg4test >=420; alg2test >=530; vtest >=598; dagtest >=640; dag7btest >=400; [85] BANKS; acceptance >=6/8; per-kind sq >=0.40, fdiv/ladder/coupled >=0.50. ALL bars hold -> GEN-8 PROMOTES TO THE GATE (first promotion since gen-6) and the census re-runs under it. Any bar broken -> no promotion, diagnose. - **VERDICT FRAME PINNED PRE-MEASUREMENT (2026-07-12, relay — the density-regime discipline applied to the bars themselves):** the bar structure has an asymmetry worth naming before it prints. bigtest recovering while alg4 misses 420 = starvation CONFIRMED + the alg3-ration mis-sized or mis-targeted (kind-ration ARITHMETIC, not mechanism failure — diet-tuning). **The junction reopens ONLY if the RATIONED registers fail to respond to their OWN rations.** A mixed table reads as grocery arithmetic, not thesis-death; only ration-blind erosion kills starvation. Also noted: the smoke's 166-vs-30 rejection rate on inverse shapes is the uniqueness gate EARNING ITS KEEP (an ungiven var constrained only downstream has more freedom to pin) — high mint-time rejection on inverse problems is the gate working, not friction; `nogive` handles [85]'s circuit the solution-first way (the edge condition made UNREPRESENTABLE, as with no-real-roots). STRATEGIC RIDER — what promotion unlocks: gen-8 promoting re-prices the BOOTSTRAP, not just the census — the 2/26/72 ran under gen-6, and the near-miss tier's shared cause (fdiv-in-composition) is precisely what the kind-ration + gen-7b's fdiv gains address. A banked-column jump under gen-8 brings the machine-drafts-the-isotopies economics with it: book 2's annotation budget splits between hand-surgery on the genuinely knotted and gate-disposal of machine proposals. The practice schedule may fund the apprenticeship. - **VERDICT (2026-07-12): GEN-8 — NO PROMOTION (two named bars miss), STARVATION CONFIRMED BY ITS OWN RATION, and the [85] mystery solved at the ENCODING layer.** The table: bigtest **967** (bar 980 — but +66 from 901, recovering two-thirds of the erosion under its surface-ration: THE RATIONED REGISTER RESPONDED; per the pinned frame this is diet-tuning, ration undersized, thesis ALIVE); alg4test **332** (bar 420, +13 only — the kind-ration MIS-TARGETED: 3000 alg3-register rows barely moved the alg4 register; the alg4 generator was archived in the house-cleaning and its register differs from alg3's — gen-9 needs a register analysis or the generator recovered); alg2test 565 ✓, vtest 600 ✓, dagtest 664 ✓, dag7btest 422 ✓ — NO NEW EROSION at 29.5k rows: the diet did not crowd. ACCEPTANCE 6/8 ✓ with the qualitative headline: **[71], [7], [72] all jumped to 5/5 UNANIMOUS** — two correct-but-shallow specimens DEEPENED TO CERTIFIED and the coupled system banked; rehearsal deepens basins, measured. [45] now votes [168,154,168,110] — a NEW specimen (mixed-vote: right answer present, wrong votes competing) for the vote-entropy column. **THE [85] DISCOVERY (verified at source, decode line 495 + gold encoding):** the args pointer decodes as top-2 DISTINCT slots and the gold multi-hot cannot express multiplicity — args=[a,a] is UNREPRESENTABLE end-to-end. Three trainings flat at ~0.33 sq because the target was never learnable AS ENCODED; 600 isq rows changed nothing because they COULDN'T. The third "unrepresentable, not unlearned" finding (family: positional-structure-as-structure, no-real-roots). Gen-9 fix is small: an arg-multiplicity bit (ftype-conditional), gold field, decode branch. CENSUS 1/23/76 — no bootstrap jump (raw prose untouched by dialect-side training, as predicted every time). GEN-9 REGISTERED ITEMS: (1) args-multiplicity mechanism (architecture, small); (2) alg4-register analysis -> correctly TARGETED kind-ration; (3) surface-ration upsized (~2500->4000); (4) the shallow-basin instrument (deferred from gen-8, now with [45]'s mixed-vote specimen as motivation); (5) fdiv borderline (0.483/0.535) rides the retargeted ration. Gate remains gen-6; ckpt `.cache/phase1_gen8_head.safetensors` banked unpromoted. - **CORRECTION + TWO STANDING RULES (2026-07-12):** (1) **THE GEN-8 'MIS-TARGETED RATION' ATTRIBUTION WAS WRONG — SEED COLLISION.** algebra4_nl_train was itself minted by algebra3_nl_gen (seed 81, teeth 0.8); the gen-8 kind-ration re-used seed 81 -> **2500/3000 rows byte-identical duplicates** -> the alg4 register accidentally received a 2x REHEARSAL UPWEIGHT... and moved only +13. THE RATIONED REGISTER FAILED ITS OWN RATION: per the pinned frame, the junction REOPENS FOR ALG4 specifically (capacity already exonerated for it by the A/B at 316~319; rehearsal now exonerated by the accidental 2x). The per-kind diff instrument (gen-6 vs gen-8 heads over alg4test, sliced by seq/pct/fdiv/crt/vieta) fires before any gen-9 ration decision. PROCESS GUARD MINTED: ration seeds must not collide with historical corpus seeds; the mix builder must PRINT DUPLICATE COUNTS (a data-boundary guard per the no-silent-fallback law — this one hid inside a healthy-looking corpus for a full generation). (2) **THE REPRESENTABILITY-AUDIT RULE (relay):** a metric flat across MULTIPLE trainings AND a targeted data intervention is not starved — it is STRUCTURALLY EXCLUDED; three flat trainings trigger a representability audit BEFORE a fourth fires. Applied retroactively it catches [85] two generations early. (3) **DELIBERATE EXCLUSIONS CARRY EXPIRATION TAGS (relay):** the repeated-arg exclusion began as a v0 soundness guard ("mul(x,x) would unsound the pairwise propagator") and AGED INTO the encoding bug when wild shapes arrived — scoping decisions expire the way the integrality jaw did; tag them at birth. (4) Surface-ration expectation clause (relay, pre-said): 4000 rows at the measured +66/2500 exchange rate plausibly clears 980, but the last points of recovered erosion are the hardest-starved — **975 = the ration curve bending, NOT the thesis breaking**; the bar stays 980, the expectation stays humble. - **ALG4 DIFF VERDICT + GEN-9 FINAL SCOPE (2026-07-12, registered before firing):** the per-kind diff (gen-6 vs gen-8, alg4test) shows **UNIFORM decline across every kind** (vieta −0.075, seq −0.069, pct −0.060, fdiv −0.056, crt −0.047 — no single victim). KIND-STARVATION REFUTED FOR ALG4: the register declines AS A WHOLE while unresponsive to 2x rehearsal (the accidental dup-upweight), 2x width (the A/B: 316~319), and decorrelation. THE ONE UNTESTED SUSPECT: **THE SCHEDULE** — curriculum orders coarse->fine by teeth score; alg4's high-teeth rows enter only in the final training phase, which cosine decay runs at annealed LR, and that phase's composition shifted as the mix grew. REGISTERED PROBE (own track, on the word, post-gen-9): a 12k CURRICULUM=0 arm on mixed8, reading alg4test — schedule starvation vs register mystery. GEN-9 SCOPE (kind-ration DROPPED with evidence): (1) ALG_DUP multiplicity mechanism (built; selftest green; env-gated, legacy byte-compat); (2) surface-ration 4000 nl-core rows, FRESH seed 99, collision-checked; (3) mixed9 DEDUPES mixed8 (removes the 2500 seed-collision duplicates; the mix builder prints dup counts — the guard, operational); (4) the SHALLOW-BASIN INSTRUMENT fires (3000-row sample, 5-view vote entropy under gen-8; correct-but- shallow rows oversampled x2 — rehearsal deepens basins, measured on [71]/[7]/[72]). BARS: [85] BANKS (self-grading); sq per-kind >=0.45; bigtest >=980 (humble clause stands); alg4test >=310 HOLD-THE-LINE (recovery rides the schedule probe, not this diet); acceptance >=7/8; alg2 >=530, vtest >=598, dagtest >=640, dag7b >=400. All bars hold -> PROMOTE (census rides). - **THE HASH AUDIT (2026-07-12, Bryce's gut + relay — three-way, zero-GPU): THE ANTI-COLLISION CLASS IS REAL — 42 VERIFIED CROSS-BOUNDARY ISOMORPHS.** (a) GREP CENSUS: the codebase is nearly hash-free — manifest pins are SHA-256/64-bit (safe by orders of magnitude at ~15 artifacts); all dedup is exact-text in Python sets (identity-with-equality, safe by construction); ONE catch: test_kenken_parity used salted built-in hash() as an RNG seed (fixtures non-reproducible across sessions; parity itself unharmed — both arms share the seed) -> FIXED to crc32 (stable). (b) No persisted built-in hash() anywhere. (c) **THE LOAD-BEARING COUNT: canonical WL form (values included, commutative roles sorted, exact-verify by backtracking before counting) over train (mixed8, 29,500 rows -> 26,920 classes; 2,574 multi-member = within-train redundancy, the small-problem pigeonhole) x the 7-test battery: exact-text overlap 0 everywhere (the string dedup held) but **bigtest 27/1500 (1.8%), vtest 13/600 (2.2%), alg4test 1, dag7btest 1 — same knot, different diagram, across the boundary.** Cause: pigeonhole density in the small-problem regime, not a seed bug. FOOTNOTE ON ALL STANDING BARS: bigtest numbers carry <=1.8% isomorph inflation, vtest <=2.2% (deltas across generations unaffected — same fixture both sides). Exclusion list persisted (.cache/iso_contamination.json); REGISTERED: (i) clean-subset re-read of the battery rides the next eval pass; (ii) the canonical digest becomes the MINT-TIME dedup + test-fixture exclusion going forward (knot invariant, not diagram fingerprint — the knots law applied to our own bookkeeping); (iii) paper tables freeze on CLEAN fixtures or report the exclusions. The gut said beware collisions; the deeper hazard was the anti-collision, and it was found before the tables froze. - **TWO CONSEQUENCES PINNED (2026-07-12, relay):** (1) **THE PIGEONHOLE FINDING IS A STANDING CONSTRAINT** — 2,574 within-train redundancy classes means the small-problem regime's structural diversity is finite and partially exhausted; every future ration inherits that ceiling. The canonical digest converts hazard to instrument: the mint now COUNTS ITS OWN KNOT DIVERSITY per batch (redundancy-class coverage alongside kind coverage) — the KIND-REHEARSAL MATRIX UPGRADES TO A KNOT-REHEARSAL MATRIX, the practice schedule tracked at the invariant level. The diagrams-> moves correction, applied to the corpus bookkeeping itself. (2) **CANONICAL DISJOINTNESS BECOMES A GENERATION-BUMP GATE** — checked at every bump so the isomorph class can never silently re-enter; the paper's reproducibility statement earns "train/test disjointness verified up to graph isomorphism," which almost no benchmark can claim because almost none check. THE TWELFTH INSTINCT'S SCORECARD: the gut said collisions, the framing said the anti-collision is scarier, the measurement said both were right in their own jurisdictions — the collision class was EMPTY (hygiene held), the anti-collision class had 42 members (the find). The instinct locates the neighborhood, the framing names the streets, the audit knocks on doors. - **VERDICT (2026-07-12): GEN-9 — NO PROMOTION BY EXACTLY ONE BAR; everything else is the sprint's best table.** THE WINS: **[85] BANKS 5/5 UNANIMOUS** — the multiplicity fix works, the self-grading probe graded itself, three generations of mystery ended by one representability audit; **sq per-kind 0.319 -> 0.751** (bar 0.45, shattered — the encoding fix unlocked the entire kind); **bigtest 1084** (bar 980; gen-6's 1000 EXCEEDED by 84 — the ration curve didn't bend, it OVERSHOT: the starvation thesis fully vindicated, erosion story CLOSED — rehearsal was the mechanism, the diet cures it); dag7btest 422 -> **510**, dag8test 379 -> 510 (the new kinds consolidated: fdiv 0.660, coupled 0.681, ladder 0.506, all bars cleared); alg4test 315 holds the line (schedule probe pending); alg2test 559, vtest 600/600, dagtest 669; val RECORD 0.8826; shallow-basin first census: deep 1432 / SHALLOW 925 (31% of the corpus!) / wrong 29 / refused 614. THE ONE MISS: **acceptance 6/8 (bar 7/8)** — [71] votes [9,9] and [78] votes [16]: RIGHT answers, vote-shy — both were 5/5 under gen-8 and RE-SHALLOWED under gen-9's diet shift. Rehearsal deepens basins; diet shifts can re-shallow specific ones — the correct-but-shallow class claiming its first promotion casualty, in its own vocabulary, one generation after the instrument was built. NO PROMOTION, mechanically; gate remains gen-6. CENSUS 1/18/81 (near shrank, knotted grew — raw-prose reading untouched as always; books remain the path). REGISTERED NEXT (on the word): **GEN-9B, a basin top-up, not a generation** — continue from gen-9 ckpt ~8k steps with a small fdiv-tiny-chain booster (the [71]/[78] shape, ~500 rows) + shallow-census-under- gen-9 oversamples; re-run acceptance + battery. The two vote-shy items are the acceptance probes; bars unchanged. Ckpt `.cache/phase1_gen9_head.safetensors` banked unpromoted. - **REGISTERED (2026-07-12, relay — GEN-9B with the displacement question riding):** (1) **THE BASIN-DISPLACEMENT CONSERVATION QUESTION** — [71]/[78] were 5/5 under gen-8; gen-9's diet (which oversampled shallow basins TO DEEPEN THEM) re-shallowed exactly those two. The uncomfortable reading, carried before the top-up can flatter it away: at fixed capacity, BASIN DEPTH MAY BEHAVE LIKE A BUDGET — rehearsal deepens practiced basins partly by drawing from unpracticed neighbors; the diet ALLOCATES consolidation rather than creating it. The Landauer law one level down: erasure-by-rehearsal was about ACCURACY; this is the same law about CERTAINTY. PRE-REGISTERED READ (rides gen-9b free): the top-up deepens the [71]/[78] shape while KEEPING gen-9's shallow-oversamples — (a) those two at 5/5 but two DIFFERENT probes gone vote-shy -> displacement CONSERVED; the diet question changes from what-to-add to WHAT EQUILIBRIUM TO SEEK; (b) >=7/8 with no new casualties -> depth NOT zero-sum in this regime; the budget worry dies with a receipt and the top-up was just a top-up. (2) **BOOSTER MINTED AT KNOT LEVEL** (the knot-rehearsal matrix's first day on the job): the [71]/[78] fix is specified as N=500 DISTINCT CANONICAL REDUNDANCY CLASSES of the fdiv-tiny-chain shape (counted by the canonical digest at mint; pigeonhole-dense regime — training the diagram again is the named risk), each checked canonically DISJOINT from every test corpus (the bump gate's first live use). (3) TOP-UP TRAINING DESIGN: continue from gen-9 ckpt, 8k steps, LR=1e-4 (cosine), CURRICULUM=0 (a top-up trains near the converged regime; re-running the coarse phase at high LR is the known disruption shape). Bars unchanged from gen-9; all hold -> PROMOTE. - **VERDICT (2026-07-12): GEN-9B — ALL BARS HOLD; **PROMOTED TO THE GATE** (first promotion since gen-6; the gate ckpt is now `.cache/phase1_gen9b_head.safetensors`).** THE TABLE: acceptance **8/8, EVERY probe banked at 5/5 or 4/5** ([71] and [78] back to unanimous; [85], [72], [56], [45], [7], [46] all held — no new casualties anywhere); bigtest **1090** (a second record); alg4test **344** (+29 over gen-9 — see below); alg2test 571, vtest 600/600, dagtest 671, dag7btest 523; per-kind sq 0.751 / coupled 0.697 / fdiv 0.677 / ladder 0.529; val record 0.8890; booster minted 500/500 distinct knots, 0 test-isomorphs admitted (the disjointness gate's first live tour, clean). **THE DISPLACEMENT ANSWER: (b) — DEPTH IS NOT ZERO-SUM IN THIS REGIME.** The top-up deepened [71]/[78] to 5/5 while every other basin held or deepened; the budget worry dies with its receipt, pre-registered. The correct reading of gen-9's re-shallowing: diet SHIFTS jostle specific basins transiently; gentle continued training consolidates without displacement. **BONUS EVIDENCE FOR THE SCHEDULE HYPOTHESIS:** alg4test rose +29 under 8k steps of LOW-LR, NO-CURRICULUM, full-mix training — more than three full generations moved it — consistent with curriculum x cosine starving high-teeth rows of usable-LR steps; the registered CURRICULUM=0 probe gains a prior. **THE TRAINING-REGIME LAW (relay, registered before it generalizes):** the displacement answer's precise scope is the REGIME, not just the diet — gen-9's re-shallowing happened under a full retrain with a shifted diet; gen-9b's consolidation-without-displacement under gentle continuation (8k, low LR, near convergence). **Hard restarts jostle basins; gentle continuation deepens without displacement.** The generation protocol may evolve toward fewer full retrains and more staged continuations; the transaction manifest tracks checkpoint LINEAGE from here. CENSUS TREND (three points, directional, informational): 76 -> 81 -> 89 knotted as basins deepen — the head's consolidation on its own register REDUCES accidental raw-prose carries; the style wall hardens as the dialect sharpens; the bootstrap's raw-prose economics await books, as every census has said. THE ARC CLOSES: the junction opened with two kills and three physics instruments; it closes with rehearsal confirmed as the mechanism (both directions), capacity and conflict exonerated with receipts, [85] representable and banked, the bookkeeping counting knots, and the gate moving on a table with no asterisks. Next chapter: the schedule probe (registered), gen-10's knot-matrix diet, and BOOKS — the style wall is now the tallest thing standing. - **REGISTERED (2026-07-12): THE SCHEDULE PROBE — the coldest-optimizer pair.** Mechanism claim: cosine schedules spend their usable LR on the curriculum's early (easy) phase; high-teeth rows arrive when LR has decayed past learning — THE HARDEST DATA GETS THE COLDEST OPTIMIZER. Design (isolates CURRICULUM alone): two 12k arms, both warm from gen-7b on mixed8 at LR 3e-4 (gen-8's exact condition; m8train states banked; no ALG_DUP — replicates the original regime): ARM A CURRICULUM=1 (control), ARM B CURRICULUM=0. PRIMARY READ: alg4test(B) − alg4test(A); prediction B > A by >=15 answers. GUARD: bigtest(B) must not trail bigtest(A) by >20 (if B lifts alg4 but craters elsewhere, the fix is per-band LR accounting, not curriculum removal). Confirmation -> gen-10 trains flat-mix (or staged LR); refutation (B ~ A) -> the top-up's +29 attributes to low-LR continuation, strengthening the regime law instead. Either way one pair of short runs converts a scatter of "hard register learns slowly" mysteries into a single attributed mechanism or clears the curriculum with a receipt. - **VERDICT (2026-07-12): THE SCHEDULE PROBE — CONFIRMED, 6x PAST THE BAR, AND THE CURRICULUM IS NET-NEGATIVE AT SCALE.** Arm A (CURRICULUM=1): alg4test 296, bigtest 916, dagtest 652, val 0.8388. Arm B (CURRICULUM=0): alg4test **384 (+88; bar was +15)**, bigtest **1032 (+116)**, dagtest 678 (+26), val 0.8618 (+0.023). The guard didn't just hold — B LEADS EVERYWHERE: the curriculum is hurting every register at the 30k-mix scale, not just the high-teeth ones. THE COLDEST-OPTIMIZER MECHANISM CONFIRMED with a sharper corollary: arm B at 12k nearly matches gen-9's 32k bigtest (1032 vs 1084) — **the curriculum was burning roughly two-thirds of every training budget** (the easy-pool phase consumes the hot LR on data the head already knows; the full mix arrives to a cold optimizer). It also retroactively explains the mid-training val dips (the 0.5 -> 0.8 jumps at pool transitions) and every "still climbing at cutoff." JURISDICTION LESSON, again: the curriculum won its ablation on 2026-07-10 in the single-register era; at mixed-register scale the verdict INVERTED — **ablation verdicts expire with their regime**, the way deliberate exclusions do; scope tags on both from here. GEN-10 CONSEQUENCE: flat mix from step one, and the schedule dividend (≈3x effective budget) comes free. Ckpts: sched_probe_armA/ armB (probe artifacts). Gate remains gen-9b (promoted on its bars; gen-10 collects the dividend). - **THREE REGISTRATIONS (2026-07-12, relay — before gen-10):** (1) **SCOPE DECAY, a new species for the §6 family:** ablation verdicts expire with their regime — the curriculum won HONESTLY (2026-07-10, single-register era, sound measurement) and inverted at mixed-register scale. Not audit-that-confirms, not flattering-verdict: a verdict aging out as the system changed under it. OPERATIONAL FORM: every ablation verdict carries a regime tag (register count, mix scale, schedule era); any verdict older than a structural shift re-audits before it is load-bearing again. (2) **THE REGIME CENSUS** (verdicts predating the mixed-register era, by re-audit priority): [HIGH] the LOSS-TERM WEIGHTS (2.0 on args/res/query, 4.0 args_w — set in the single-register era, load-bearing in every run since); [MED] LR=3e-4/BATCH=8 (old, but implicitly re-validated by every healthy val curve — flat-mix changes the regime again, so a small LR sweep rides a future cheap slot); [MED] the 5-view/vote-3 TTA dials (gen-5 era; but exercised daily by acceptance probes — living verification); [LOW] T_ALG=256, N_DIG=3, H_W=512 (H_W freshly re-validated by the capacity A/B; the others are data-bounded, not regime-bounded). None urgent; all now tagged. (3) **THE TEXTURE RULE** (this channel's self-audit): recurring unexplained texture in training curves is an ANOMALY, not scenery — two sightings of the same unexplained curve shape trigger a mechanism probe (the [85] three-flat-trainings rule, generalized to curves). The mid-training val dips and "still climbing at cutoff" were logged repeatedly as texture; the probe they pointed to was worth 2/3 of every training day and fired a week late. (4) BOOKS ECONOMICS NOTE: at 3x effective budget, December's reading-training runs just tripled in affordability — gen-10 carries the FIRST REAL-PROSE INGEST (v0, n=14 book pairs, informational arm) alongside its diet rather than after it. The wall is tallest; the ladder got longer. - **REGISTERED (2026-07-12): GEN-10 — flat-mix lineage continuation + the knot-matrix diet + the prose-v0 arm.** DESIGN: (1) the knot-rehearsal matrix's first dietary act — coverage report (distinct canonical classes per dag kind over the current mix), and a 2000-row dag10 booster quota'd INVERSELY to knot count (thin kinds fed first); (2) mixed10 = mixed9b + dag10; (3) training in the CONTINUATION REGIME: warm from gen-9b (lineage), CURRICULUM=0 (the probe's dividend), 16k steps, LR 1e-4; (4) THE PROSE-V0 ARM (informational, n=14): book-1 dialects parsed under the new head, verified to banked answers, gold graphs attached to the RAW PROSE (factors span-less; span losses auto-mask — the build_gold patch); a 600-step LR 5e-5 micro-continuation, then census + raw-prose acceptance BEFORE/AFTER. No bar — the arm builds and measures the reading-training machinery honestly at n=14 (Brick discipline); DISPLACEMENT GUARD: bigtest under the prose ckpt may not trail gen-10's by >15 or the arm's ckpt is discarded (the main gen-10 ckpt is unaffected either way). PROMOTION BARS (all hold -> gen-10 takes the gate): bigtest >=1090 (hold the record), alg4test >=380 (the flat dividend must show at continuation), acceptance 8/8, per-kind sq >=0.70 / fdiv >=0.62 / coupled >=0.65 / ladder >=0.50, alg2test >=560, vtest >=598, dagtest >=660, dag7btest >=500, dag8test >=500. - **TRIPLE VERDICT (2026-07-13):** (1) **GEN-10 — NO PROMOTION, one bar: alg4test 357 < 380.** Everything else: records nearly across the board — bigtest **1130** (third record), alg2test 585 (record), dagtest 678, dag7btest 537, dag8test 532, sq 0.781, fdiv 0.702, coupled 0.707, acceptance 8/8 unanimous, census softened 89 -> 79 (the knot-diverse booster + flat continuation REVERSED some register hardening). The alg4 reading: gentle continuation (+13) cannot pay a debt the curriculum-era LINEAGE carries — arm B's 384 came from 12k HOT flat steps; the heat-vs-jostle tension is gen-11's design question (hot flat retrain from a clean ancestor vs medium- heat continuation). Gate REMAINS gen-9b. (2) **PROSE-V0 — the displacement guard fired at −243** (bigtest 887 vs 1130): 600 steps x batch 8 on n=14 = ~340 epochs of pure prose = catastrophic interference; ckpt DISCARDED from candidacy (kept as a displacement specimen). The census under it read banked 15 — **CONTAMINATED BY CONSTRUCTION**: the 14 prose rows ARE census-pool members; training-set recall, not reading (the disjointness law's third bite, now prose-side: future prose censuses EXCLUDE trained items). HONEST V0 YIELD: the machinery works end-to-end (14/14 pairs built, span-less gold binds, gradients flow, trained rows parse) — raw prose is LEARNABLE-IN-PRINCIPLE through this path; v1 mixes prose INTO the diet (never a naive fine-tune) and reads a disjoint census. (3) **THE SYNC AUDIT — CANDIDATE 1 CONFIRMED** (Bryce's gut, thirteen-for-thirteen): the manifest sat at GEN-5 all sprint; promotions gen-6..9b were PARSER-ONLY; the composed stack pairs a gen-9b parser with a gen-5 specialist, gen-5 monitor centroids, and a gen-5 mouth — every mouth distance since gen-6 was a gen-5-native reading (the knotted-column hardening may be partly mouth-calibration lag). FIXED TONIGHT: GENERATION.json rewritten to the TRUE stack (gen-9b parser, stale members EXPLICITLY WAIVERED, non-candidates named); REGISTERED: the entourage rebuild (specialist remine on gen-9b errors, monitor centroids in gen-9b slot space, mouth recalibration on the consolidated family) + co-generation assertion in --check + stage-boundary row-count/size asserts. Candidates 2 (artifacts verified exact), 3 (denominators consistent — full fixtures everywhere, exclusion list never applied), 4 (three heads tonight — resolved by the manifest naming the gate) all closed. Discipline -> mechanism, the house conversion, one more time. - **THE LAW + THE QUEUE (2026-07-13, relay — session close):** (1) **NEW LAW-FAMILY ENTRY: PROSE PROMOTIONS DON'T MOVE MACHINES.** The manifest sat at gen-5 for four generations not by negligence but because promotions were LEDGER EVENTS (sentences) while the manifest was a separate artifact no workflow touched — narrative truth and machine-readable truth drifted silently. OPERATIONAL FORM: any state the system depends on must be updated by the SAME TRANSACTION that creates the dependency, or the check must fail loudly — the promotion battery ends by WRITING THE MANIFEST or refusing to print "PROMOTED"; the word and the JSON become one atomic act. Second member of the sync family minted by the same pattern: discipline drifts, gut fires, mechanism ships. (2) **THE ENTOURAGE REBUILD QUEUE (sequenced): MOUTH FIRST** — recalibrate against the gen-9b native family BEFORE any new census claims bank (every mouth reading since gen-6 was gen-5-native; the knotted trend 76->81->89->79 carries an unknown zero-point error), then the FREE RETROACTIVE READ: banked prose vectors against the fresh calibration re-scores the whole hardening history from artifacts on disk — the odometer re-zeroed AND the history corrected. Then: specialist remine (gen-9b errors), monitor centroids (gen-9b slot space), co-generation --check, stage-boundary asserts. (3) GEN-11 DESIGN NOTE (the heat-vs-jostle answer, hypothesized): STAGED — a brief hot phase on the debt register alone, then cold consolidation on the full mix; the top-up pattern with a targeted preamble. (4) PROSE-V1 DESIGN: prose mixed INTO the diet, census read on DISJOINT rows — both lessons inherited at the price of one discarded checkpoint; the arm was always a mechanism check, and the mechanism checks out. - **VERDICT (2026-07-13): MOUTH RECALIBRATED — the retroactive read is the AUDIT-THAT-CONFIRMS: the gen-5 lens error was IMMATERIAL.** New bank drawn from m9btrain (the current family); native threshold TIGHTENED 0.0443 -> 0.0347 (the grown family is more compact — the mouth got sharper, not looser). Retroactive pair-reads: harvest odometer 0.2488 -> 0.2431 (~2%), census pool 0.2558 -> 0.2500, book-1 raw 0.1983 -> 0.1917, read-foreign 100% under both lenses. ALL BANKED MOUTH NUMBERS STAND; the calibration ambiguity resolves as never-material; the sync find's value was the PROTOCOL hole, not a corrupted history. Manifest updated (mouth = gen-9b artifact, waiver retired, hash pinned). RECORD CORRECTION carried from the critique: the census never consults the mouth — the knotted trend was always real parse behavior; the stale lens touched deployment gating, the odometer, and diversity guards only. PROCESS NOTE: the recal unit took five launches — the working-directory omission was repeated FOUR times consecutively before switching to a form where the mistake is unrepresentable (WorkingDirectory= property + absolute paths) — the unrepresentability lesson applies to one's own tooling habits too. CRITIQUE AMENDMENTS BANKED: prose doses re-phrased as oversample multiples of the 14 uniques (x5/x15/x40; "2% of diet" would be x49 = v0's poison in a percentage); the sweep's deliverable = safe-dose slope PER UNIQUE ROW (December's books-planning number). Sequencing stands: entourage (mouth DONE; specialist remine + centroids next, on the word) -> bars -> gen-11. - **THE DOSE LAW GENERALIZED (2026-07-13, relay):** percentages smuggle repetition when the unique pool is small — every diet specification carries BOTH numbers from here: SHARE OF MIX (governs interference) and REPETITIONS PER UNIQUE ROW (governs memorization); they decouple violently at small n. The prose sweep's deliverable, priced: safe-dose slope per annotated row converts December's books question into arithmetic — target census movement / (movement per row x safe multiple) = annotation budget. The books campaign priced by its own pilot. - **REGISTERED + FIRED (2026-07-13): THE SPECIALIST REMINE (entourage step 2) + MONITOR CENTROIDS (step 3).** Recipe = the gen-5 bump's stages 1-4 adapted: fresh repair corpora (nl/alg2/alg3/verbose + the DAG register the gen-5 repair mix never had; seeds 211-215, collision-free), precompute, nack --prep/--train against the GEN-9B parser's errors (6k steps), centroids rebuilt from the m9btrain family in gen-9b fst space. Manifest updated at the end: specialist + centroids waivers retired, hashes pinned — entourage complete, THEN bars, THEN gen-11. - **VERDICT (2026-07-13): THE ENTOURAGE IS COMPLETE.** Remine chain (3rd launch; two None-grad fixes en route — h_dup joins the family with the two-terminal lesson: EMISSION AND GOLD FEED both, or the branch is dead): 3,800 repair rows incl. the dag register the gen-5 specialist never saw; purity filter -208; **1,338 organic gen-9b failures** -> phase1_gen9b_nack (best-by-EMA 4.968); centroids rebuilt, all 7 kinds, gen-9b fst space; manifest retired the last waivers ITSELF as the chain's closing act. The composed stack speaks ONE GENERATION for the first time since 2026-07-10. The sync find is fully closed: hole found, history verified clean, mechanism shipped, entourage rebuilt. NEXT: bars (re-pin with the composed stack readable), then GEN-11 — staged-heat probe, dosed prose arm (oversample-multiple units), and the first battery that writes its own manifest or refuses the word. - **REGISTERED (2026-07-13, Bryce's triple import + relay + critique: KAGGLE/ALPHAZERO/MUZERO):** (1) **MCTS-IN-THE-MINT (strategic; the pigeonhole closed-loop).** Not solver-side (refuted v3-v4 territory; GAC/MRV/LCV IS principled lookahead) — MINT-side: state = partial DAG, actions = add-relation/given/close, value = knot-novelty x gate-survival, the knot-rehearsal matrix as REWARD — the matrix upgrades from reporting instrument to closed-loop curriculum controller; the mint searches toward thin redundancy classes. V1 = GREEDY one-step (canonical-digest peek before the expensive gate; the mint already early-rejects on kind — extend to knot-class population). CRITIQUE AMENDMENTS: (a) novelty alone can mint pathological diversity — the value blends novelty + gate-survival + kind-rehearsal targets; (b) prediction RE-PINNED: knot-classes-per-1000 minted >=2x (the real win); gate-survival gain MODEST (most residual rejections are global-uniqueness failures one-step lookahead cannot foresee). Self-play's deep lesson kept: mint at the frontier of competence — toward the census's temperature gradient. (2) **MUZERO -> TRIAGE, not world model.** The registry IS the world model (owned, exact); latent dynamics in the solve path would trade the zero-leakage bottleneck for drift — anti-thesis. The one licensed address: a small head predicting WHERE THE REPRESENTATION FAILS (P(knotted) for raw prose) — annotation-budget routing, books priced before purchase, mouth-adjacent (selection-safe, zero solve-path contact). DESIGN CONSTRAINT from the banked negative: the mouth-distance join's AUC 0.535 says INPUT-SPACE features do not predict knottedness — the triage head needs PARSE-SIDE features (vote entropy, factor counts, calibration). Waits for census outcomes at books scale; registered, not fired. (3) **CROSS-MODEL x VIEW LATTICE (the afternoon; fire-ready).** Single-model TTA certifies invariance across DIAGRAMS; cross-model consensus certifies invariance across LANDSCAPES — the strictly stronger invariant, aimed at the certification channel's named blind spot ([71] 5/5-unanimous-wrong; the anchor's correlated blindness). CRITIQUE AMENDMENT — member choice by DECORRELATION, not strength: gen-9b/gen-10 share lineage (partial decorrelation only); the banked panel offers true diversity for free — sched_probe_armB (flat regime, gen-7b lineage) and cap2x_32k (2x WIDTH — architectural decorrelation). Panel = gate + one cross-lineage + one cross-width member; 2-3x inference on the CERTIFICATION TIER only; zero training. PREDICTIONS PINNED: cross-model unanimity precision > single-model at meaningfully lower coverage; the coverage GAP = a new instrument (lineage disagreement — the Goodhart rotation's held-out examiner from inside the house). Priority: (3) afternoon, (1) rides gen-11's mint, (2) books-era. Fourteen instincts, all machinery. - **REGISTERED + FIRED (2026-07-13): THE LATTICE PROBE — cross-model x view certification.** PANEL (decorrelation axes, from the checkpoint bench): gen-9b (gate), sched_probe_armB (LINEAGE axis: gen-7b ancestry, flat regime), cap2x_32k (WIDTH axis: 1024d). CERT-V2 RULE (pinned): gate 5/5 unanimous AND both siblings' 5-view majorities agree with the gate's answer. READS (pinned before any vote): (a) bigtest precision/coverage, cert-v2 vs gate-only 5/5 — prediction: precision rises at meaningfully lower coverage; the coverage GAP is the new lineage-disagreement instrument; (b) THE DISAGREEMENT AUTOPSY (decides how the panel GROWS): on gate false-certificates, which sibling dissents — armB breaking more -> lineage is the load-bearing diversity axis; cap2x more -> width earns permanent employment; (c) THE DEEP-WRONG READ ([71]'s class, the only error family with no detector since the anchor): gate stable-wrongs on bigtest + census pool — cross- examination breaking >=1/3 of them = the FIFTH JURISDICTION lands (prevention, depth, detection, recognition, CROSS-EXAMINATION). BENCH NOTE (relay): diagnostic runs leave lineage-decorrelated siblings — the manifest tags them PANEL-ELIGIBLE instead of archiving; every future fair A/B grows the ensemble for free. - **THREE PRE-JOIN PINS (2026-07-13, relay — registered while the jury deliberates, BEFORE the join prints):** (1) **THE NORMALIZER:** cap2x is behaviorally near-identical to the 1x by the fair A/B's own verdict — sparse dissent from cap2x may mean "width axis loses" OR "member never diverse enough to test width." The autopsy reads through the RAW DISAGREEMENT RATE on ALL items per member: equal rates -> axis comparison fair; unequal -> the axis question stays open and panels recruit by MEASURED BEHAVIORAL DISTANCE, not axis theory. (2) **VETO OR APPEAL:** on gate false-certificates broken by a dissent, check the dissenter against the KEY — dissenter disproportionately correct -> the panel is a REPAIR channel (the first mechanism with purchase on committed-wrongs since the survivor arc; jurisdiction five becomes APPEAL); dissents mostly both-wrong-differently -> pure abstention machinery (veto). (3) **GEN-11'S FIFTH DIAL, PRE-COMMITTED NOW so it cannot look fitted: IF cert-v2 lands (>=1/3 stable-wrongs broken at modest coverage cost), the freeze gains the dial and every future battery must hold cert-v2 precision >= 0.998 at whatever coverage it buys, measured per generation alongside the other dials.** If the probe misses, the panel banks as an honest negative WITH its autopsy; the bench note survives either verdict (panel-eligibility is free and the next junction mints wider-gapped members). - **VERDICT (2026-07-13): THE LATTICE PROBE — the gate is CLEANER than its blind-spot narrative, and the panel's real jurisdiction is the WILD register.** bigtest: gate-only 5/5 = 866 coverage (57.7%) at precision **1.0000** — ZERO false certificates in 1500; the [71]-class has nearly vanished from the dialect fixture under gen-9b (ONE stable-wrong in 1500 — and cross-examination broke it, 1/1). CERT-V2: 839 coverage (−27, 3.1% cost) at 1.0000. THE FIFTH DIAL: mechanically LANDS per the pre-commitment (100% >= 1/3 at modest cost) — ADOPTED WITH THE n=1 CAVEAT stated: on in-register text the class it hunts is nearly extinct; its load-bearing jurisdiction is WILD register, where the census read is emphatic — of 10 gate stable-vote raw-prose parses, the panel DISSENTS ON 9. Cross-examination is a second wall behind the mouth: even prose that slips the doorman meets a jury that refuses 90%. VETO-OR- APPEAL: 0/0, undecidable this round (nothing to appeal — the right kind of failure). THE NORMALIZER'S SURPRISE: cap2x disagrees MORE per-item (24.7%) than armB (17.5%) despite the A/B's aggregate equivalence — AGGREGATE EQUALITY MASKED ITEM-LEVEL DIVERSITY (the fair A/B measured means, not overlaps); the axis question stays open and panels recruit by measured behavioral distance. Coverage gap = 27 (the lineage-disagreement instrument's zero-point). Probe cost: 2h05m CPU, zero training. The corpses voted; the gate walked free; the jury found its real beat on the wild side of the wall. - **REGISTERED (2026-07-13, relay + critique): MEANS-VS-OVERLAPS (scope-decay's cousin)** — verdicts about MEANS don't govern claims about OVERLAPS: cap2x was "statistically the same model" by every aggregate and disagreed on 24.7% of items. Behavioral distance is a different measurement than benchmark distance; diagnostic checkpoints get graded on per-item disagreement at archive time (bench protocol). Zero-numerator note carried: 1.0000 on 866 reads as bounded-near-a-tenth-percent; the STRUCTURE is the claim. **THE WILD WATCHER (pre-registered column):** cert-v2's census-side dissent rate (9/10 at gen-9b) logs per generation — NOT a bar (the wild isn't gated) but the instrument watching prose-v1: reading- training working -> dissent falls as both members learn the register; dissent falling toward unanimity WITHOUT the key confirming -> the Goodhart signature, exactly where the rotation law predicts. One column, two watchers. - **REGISTERED + FIRED (2026-07-13): GEN-11 — the five-dial generation.** CHAIN: (A) THE STAGED-HEAT MICRO-PROBE first (3k hot LR 3e-4 on alg4train from gen-9b -> read; 4k cold LR 1e-4 on mixed10 -> read). BRANCH RULE (pinned): hot alg4test >=380 AND cold holds alg4 >=370 with bigtest >=1080 -> gen-11 adopts the staged recipe (hot alg4 preamble, then cold flat mixed11); else plain flat continuation. (B) DIET: mixed11 = mixed10 + dag11 booster 2000 minted with the GREEDY KNOT PEEK (canonical-class dedup against the mix + itself; prediction: booster knot-classes/row >=0.95 and >= 2x dag10's class rate). (C) TRAIN per the branch; battery; per-kind; acceptance. (D) FIFTH DIAL, first enforcement: 3-member lattice (gen-11 + armB + cap2x) on bigtest — cert-v2 precision >=0.998 REQUIRED; census dissent column logged. (E) PROSE DOSE ARMS (x5/x15/x40 of the 14 uniques, each mixed into a 2000-row slice, 2k-step micro-continuations from gen-11): reads = DISJOINT census (the 86 untrained pool items) + bigtest displacement guard (>= gen-11 − 15) per arm; deliverable = the safe-dose slope per unique row (December's number). (F) THE BATTERY WRITES THE MANIFEST: a verdict script checks every bar mechanically and either writes GENERATION.json (with an EXPLICIT one-generation specialist waiver — remine rides the next entourage pass) and prints PROMOTED, or prints the kill — the word and the JSON one act, the law's first enforcement. BARS: bigtest >=1130, alg4test >=380 (THE bar), acceptance 8/8, alg2 >=560, vtest >=598, dagtest >=660, dag7b >=500, dag8 >=500, sq >=0.70, fdiv >=0.62, coupled >=0.65, ladder >=0.50, cert-v2 precision >=0.998. - **PERF AUDIT (2026-07-13, Bryce's question — 'are the easy gains in place?'):** TRAINING yes (TinyJit step + assign-in-place fixed buffers, 0.06s/step — the substrate pattern, long since in place); PRECOMPUTE yes (batched, memmap). THE EVAL STACK NO — found and fixed the big hole: **recompute_states reloaded the 2.4GB Llama weights ON EVERY CALL** (one reload per problem across every census/acceptance/lattice/book gate since Phase-1 began — the llama_loader spam in every log was the bill). Host now CACHED per process; parity BYTE-EXACT vs the pre-edit reference. The trunk TinyJit was attempted and honestly reverted: zero-arg capture RECAPTURES per call with this layer code (13s/batch vs eager, measured under GPU contention) — DEFERRED with the residency smoke's assign-in-place buffer pattern as the known-good recipe (0.34s replay, validated 2026-07-05); the head-forward JIT rides the same deferred item (trunk dominates). Clean benchmarks after gen-11's unit frees the GPU. Every future eval process (including gen-11's own stage E lattice and dose arms, which spawn fresh processes) inherits the cache immediately. - **VERDICT (2026-07-13): GEN-11 — KILL, two bars; the manifest law's FIRST ENFORCEMENT worked exactly as written (kill printed, JSON untouched, no word without the write).** THE TABLE: records nearly everywhere — bigtest **1137**, alg2test 592, dagtest 681, dag7btest 548, dag8test 534, fdiv 0.713, sq 0.781; **CERT-V2 1.0000 at 862 coverage — the fifth dial HOLDS its first enforcement**; deep-wrong still exactly 1; wild-watcher column: 16/19 dissents (84%, from 90%). THE TWO MISSES: alg4test 370 (bar 380; highest since gen-6's 371) and acceptance 7/8 — [45] again, the CHRONIC specimen, votes [154,154,168] (the mixed-vote class's poster child across three generations). STAGED-HEAT REFUTED FROM THIS LINEAGE: the probe's hot phase on pure alg4train reached only 362 (<380) — the branch rule correctly fell back to plain flat, and the finding is sharp: **the alg4 debt is not heat-reachable from the gen-9b/10 lineage** (armB's 384 came from gen-7b ancestry; the debt lives IN the lineage, below the schedule). GREEDY MINT: 2000/2000 distinct knots, ZERO dups — prediction met at 1.00; value diversity makes classes nearly free at booster scale; the peek stays as cheap insurance. **THE DOSE PILOT (December's number, honest): ~ZERO movement per unique row at n=14** — no arm banked disjoint-census gains (<=1); x15's near-column bump (15) is the only weak positive; ALL THREE ARMS FIRED THE DISPLACEMENT GUARD (−47/−24/−19, non-monotonic — the arms are under-averaged at 2k steps). The pilot's verdict: the constraint is UNIQUE ROWS, not repetitions — December's annotation budget must grow n; no oversample multiple substitutes for a book. GOVERNANCE OBSERVATION SURFACED (not relitigated): two consecutive unpromoted heads now beat the gate on most dials; the alg4 380 bar has never been reached by ANY head of the current lineage; [45] is a chronic single-item acceptance gater. The bars held their line — whether the LINE is right is the relay's and Bryce's call, with the lineage-debt finding as the new fact on the table. - **REGISTERED + FIRED (2026-07-13, Bryce's gut #15 + relay: THE LATENT-SPACE AUDIT.** Root named: every geometric instrument was calibrated in SOME generation's latent space; six generations of consolidation rotate/scale/stratify the coordinates under them — the sync-audit's geometry twin (files were the artifact version; this is the coordinates themselves). CRITIQUE CORRECTION carried: the MOUTH IS IMMUNE BY CONSTRUCTION — it reads frozen-trunk space (weights untouched since Phase-1); its threshold tightening can only be corpus membership, never norm growth. Stratification is a HEAD-SIDE (fst) hazard exclusively — the at-risk watchers are the centroid/library family. THE THREE PROBES (no training, banked checkpoints): (A) DRIFT — orthogonal Procrustes between the gen-5 and gen-9b centroid constellations + per-kind alignment: small residual = translation (re-anchoring suffices, standing recal joins the bump); large = reorganization (geometric reads since last validation get footnotes). (B) STRATIFICATION — per-kind fst NORM longitudinal across the bench (gen-6..11 heads, fixed m9btrain sample): prediction = norm correlates with cumulative rehearsal (old-guard high, gen-8+ kinds low); if real, the fix is TWO-CHANNEL reads (angle + radius as separate columns — the density-regime discipline applied to geometry), not abandoning cosine. (C) SEPARATION (the [45] mechanism hypothesis) — per book pair, pooled-trunk cosine between RAW prose and its DIALECT twin; [45] + mixed-vote siblings vs the banked controls: anomalous raw<->dialect geometry for the chronic family = the chronic case is a FROZEN-TRUNK separation limit (not unlearnable — UNSEPARATED AT THE SOURCE; representability's geometric cousin), and the remedy leaves the diet entirely (deeper-prefix question's second customer). META: instruments are trained-adjacent objects — they age with the system they watch; RECALIBRATE-THE-WATCHERS becomes a standing generational duty beside remine-the-specialist. - **VERDICT (2026-07-13): THE LATENT-SPACE AUDIT — all three probes land.** (A) **DRIFT = PURE ROTATION**: raw centroid cosines ~0.59 (the constellations look unrelated in raw coordinates) but Procrustes-aligned **0.988 mean, residual 0.155** — the constellation SHAPE is intact; the space rotated. Re-anchoring suffices (the entourage's rebuild was the right fix); the waist monitor's historical AUC decay now has its mechanism (it was reading rotated coordinates); LAW: never mix generations' head-space coordinates — align or re-anchor. (B) **STRATIFICATION REAL, unexpected shape**: not old-high/new-low but LONGITUDINAL COMPRESSION-THEN-RECOVERY — the whole fst space contracted ~40% at gen-7b (the frame-change generation) and slowly re-inflates (gen-6 ~11-13.6 -> gen-8/9b bottom ~6.4-7.8 -> gen-11 ~6.8-8.6). WITHIN generations the prediction's direction holds: fdiv is the lowest-norm kind on every row since gen-7b AND the weakest per-kind performer — RADIUS TRACKS CONSOLIDATION; the two-channel read (angle + radius) is justified and joins the instrument kit. Cross-generation cosine is doubly unsafe (rotation + scale). (C) **THE CHRONIC CASE HAS ITS MECHANISM**: [45] has the LOWEST raw<->dialect trunk cosine of all 14 book pairs (0.639, z=−2.05) — its taxi surface and rate structure are maximally divergent AT THE FROZEN TRUNK; no head diet can fix a read the trunk refuses to align ([7], the faucet rate problem, is second-lowest — the RATE-PROBLEM FAMILY clusters at the bottom). The mixed-history items [51]/[54] sit HIGH — their past instability was basin-side and rehearsal cured it. THE DISTINCTION IS GEOMETRIC: transient mixed-votes = shallow basins (diet cures); chronic = trunk-level frame distance ([45] leaves the diet conversation and becomes the deeper-prefix question's second customer, with [7] as its sibling watch). Gut #15: all three drawers had something in them. - **THREE RETROACTIVE SETTLEMENTS + THE FOURTH CHARTER REVISION (2026-07-13, relay):** (1) **THE WAIST MONITOR IS EXONERATED** — its AUC decay was filed under selection-hardening (errors evolving to look normal); pure rotation says it aged because NOBODY TOLD IT THE SKY HAD TURNED. The two laws now have a DISCRIMINATING TEST (registered): re-anchor and re-measure — recovered AUC = rotation was the whole story; still-degraded = hardening is real ON TOP. Bench amendment: cross-lineage panel members disagree partly BY COORDINATE FRAME; cert-v2 is immune (votes are answers, not geometry) but any future GEOMETRIC ensemble read needs Procrustes first. (2) **RADIUS IS A CONSOLIDATION CLOCK** — the gen-7b compression scar (~40%, still re-inflating four generations later) records ARCHITECTURAL HISTORY, and within-generation per-kind norm is a free consolidation gauge readable at every checkpoint, predicting ration needs BEFORE the battery: fdiv's chronic low radius says its borderline per-kind era (0.483/0.535) was under-consolidation all along — gen-12's mint weights it accordingly. Two-channel read (angle=identity, radius= consolidation) joins the standing kit. (3) **THE CHRONIC CLASS HAS A BIRTH CERTIFICATE and the taxonomy splits clean**: transient = shallow basin (diet cures, measured on [51]/[54]); chronic = trunk frame distance (measured, z=−2.05). The rate family's clustering is the mechanistic tell: rate problems are where surface narrative and structural content are most ENTANGLED IN NATURAL PROSE — the trunk read a trillion tokens where "fare" and "multiply" co-occur in frames that never separate. **THE FOURTH CHARTER REVISION — the first that GROWS the organ**: before the deeper-prefix surgery, the cheaper candidate is a DISAMBIGUATION REWRITE ("the taxi charges 3 per mile" -> "f = 3 x m" — frame-separation the annotation gate already performs by hand). The organ's true kingdom may be FRAME-DISENTANGLEMENT, not structural-facts-only — the thrice-shrunk surgeon finally gets a patient list written in z-scores. BENCHED WITH CERTIFICATES: [45] and [7] (chronic, trunk-frame class); gen-12's acceptance bars exclude them per this restructure (their cure is the organ's or the prefix's, not the diet's); the organ's eventual training data includes their family. - **REGISTERED + FIRED (2026-07-13/14): BRICK-M — THE SCHEMA MINER (the hierarchical library's measurement-first entry; the layers-of-abstraction instinct split down the C2 line: surface frames [taxi/snow] stay OUT — disguises the mouth/organ handle; structural schemas [RATE/WORK/MIXTURE] are shapes the library may store as TYPED MACRO-FACTORS — compile-time sugar whose expansion the solver sees as primitives: expressiveness without surrendering auditability; embedding-space crossing lives at RECOGNITION [schema retrieval; analogy = shared-schema detection], never at deduction).** THE MINER: banked gold graphs (train corpora tagged by source + the harvest column = book-pair graphs), rooted upstream-closed subgraphs of 2-6 factors, WL-canonical with VALUES ABSTRACTED (RATE is a shape, not a number; ftype/op/sel retained), ranked by frequency x source spread; train and harvest mined SEPARATELY then joined (frequent-in-train-absent-in-harvest = generator habit; the reverse = coverage gap — the knot matrix lifted to schema level). PINNED: (P1) <15 classes cover >60% of occurrences (shallow hierarchy; hundreds = macro-factors die before design); (P2) top classes human-nameable (a top-10 of unnameable wiring accidents = generator artifact — the honest falsifier); (P3) [45] and [7] share a class — the miner doubles as the organ's patient registry. THE GATE: the miner RANKS, never ADMITS — registry admission keeps full birthright; C2's ghost is stopped between mining and minting. Lineage: the musical-keys era's composition catalog — the nouns died, the verb survives, bottom-up, ten months later. - **VERDICT (2026-07-14): BRICK-M — P1 passes at the wire, P2 splits the concept, P3 passes, and the deepest finding is WHERE SCHEMAS LIVE.** Numbers: 20k train rows -> 265,174 subgraph occurrences in 10,232 classes; **top-15 cover 60.5%** (bar 60% — shallow hierarchy confirmed, barely). P2 HONEST READ: the top-20 are nameable but GENERIC — add-chain scaffolding, mod+given, fdiv+given, pct+given: arithmetic PLUMBING, not conceptual schemas. Cause is structural: generated corpora are built FROM primitives, so their statistics return primitives (circular); the conceptual column (harvest) is n=14 — too thin to mine. THE JOIN WORKED AS A DIET INSTRUMENT: 4 harvest-only classes, ALL the sum-of-prefix-terms wiring from [46] (the ladder quota never covered running-prefix-sums — a named train-side gap). P3: [45] and [7] share BOTH their classes — but not exclusively, which is the night's insight: **value-abstracted rate graphs are indistinguishable from generic mul/fdiv plumbing — RATE-ness does not exist at the graph level; it lives in the LANGUAGE-GRAPH BINDING.** The C2 split is thereby vindicated stronger than designed: the graph layer literally cannot see taxi-vs-faucet, so the library's second floor splits into TWO objects — (i) MECHANICAL MACROS on the graph side (CHAIN(k)/ PREFIX-SUM — real candidates, attack the measured ladder/length walls, proposable from THIS mine), and (ii) CONCEPTUAL SCHEMAS as PARSE-SIDE RECOGNITION objects (schema retrieval from prose — minable only when the harvest column grows). EVERY ROAD LEADS TO N: the schema library joins the census, the dose slope, and the organ in waiting on books. Deliverable banked: .cache/schema_mine_top50.json (ranked, never admitted). - **THE BINDING THEOREM + THE CRITICAL PATH (2026-07-14, relay — formal statements):** (1) **THE BINDING THEOREM (the two-channel spine's third and final vindication):** C2 proved operation-type is not classifiable from surface features (LANGUAGE WITHOUT STRUCTURE fails); Brick-M proved schema-type is not recoverable from wiring (STRUCTURE WITHOUT LANGUAGE fails); therefore CONCEPTS ARE IRREDUCIBLY BINDINGS — which is why the parser (the binding organ) hosts schema retrieval, why the silhouette library was always its right home, and why [45] is chronic: its pathology lives IN THE BINDING LAYER, the one place neither a graph fix nor a language fix alone can reach. The architecture's deepest design decision has its completeness proof. (2) **THE CRITICAL PATH, in exactly these words: FOUR UNRELATED MEASUREMENTS — the census, the dose slope, the organ's patient registry, the schema library — ALL BOTTLENECK ON n. When four independent instruments triangulate one coordinate, that coordinate is the critical path by definition. THE BOOKS ARE NOT DECEMBER'S CHAPTER; THEY ARE THE ONLY CHAPTER, quadruple-confirmed.** (3) **THE LIBRARY'S TWO-ADDRESS CHARTER (named before existence, against the attic-C2 ghost):** floor two has two addresses — MECHANICAL MACROS live graph-side (typed subgraph templates, deterministic expansion, solver sees primitives); CONCEPTUAL SCHEMAS live parse-side (recognition objects over prose, retrieval not deduction). Confusing the addresses re-imports C2 through the attic. (4) **REGISTERED PROPOSAL (rank-gate honored): CHAIN(k) + PREFIX-SUM as the first macro-factor candidates** — receipts: the harvest-only classes ([46]'s wiring, zero generator coverage), the ladder/length walls as standing symptoms; expansion deterministic, pointer supervision inherits ladder machinery. Admission = a design decision at gen-12 registry review. FIRED NOW (diet, not registry): the PREFIX-SUM SHAPE joins the DAG rotation — the measured hole gets minted. - **THE EMPLOYMENT LAW + TWO PROMOTIONS (2026-07-11, relay — registered before gen-7):** (1) **THE EMPLOYMENT LAW**: every organ this project charters gets SMALLER on contact with measurement — book 1 is the fourth charter shrinking (organ suspects resolving as readable structure wearing unreadable prose; confirmed kingdom = a named list of four, one a generator line). REGISTERED KILL CRITERION: if the organ's kingdom shrinks below ~5% of harvest refusals after the style wall falls, the "organ" becomes a RELATION TRANCHE plus a chain-length fix and the funnel never grows its generative layer. The deciding measurement is free — the census, re-run after reading-training ships. Nouns die; verbs survive. (2) **[71] PROMOTED — THE MOUTH IS MANDATORY**: 5/5 unanimous-wrong on raw prose is the certification channel's first observed false certificate — unanimity certifies STABILITY, and [71] demonstrated stable wrongness in the wild at n=1. Chain-of-custody, now measured end to end: MOUTH clears register -> LATTICE certifies stability -> KEY confirms truth. The lattice must never see raw prose the mouth hasn't cleared; every link has a named specimen showing why it cannot be removed. (3) **[46] PROMOTED — THE PARAPHRASE ACCEPTANCE PROBE**: the surface flip (same relations, "plus...equals" refuses / "The sum of...is" banks 5/5) makes gen-7's surface-robustness item a paraphrase-augmentation LINE in the generator, not a research question; [46]'s v1 dialect is the pre-positioned acceptance probe (must bank post-gen-7). The tier-N miss stands as a debt WITH AN INVOICE: fdiv-in-composition 0/4 with a stable wrong vote, atop gen-7's worklist. Gen-7's charter is written by real failures for the first time; awaits the word. - **REGISTERED (2026-07-14, relay + critique): THE BOOKS CAMPAIGN CHARTER — the only chapter, chartered the house way.** SIZING: dose pilot says ~zero movement/row at n=14; census demand side says 76-89 knotted with ~75% style-recoverable; bilingual precedent (2,000 paired rows taught the verbose register) gives the PRIOR of hundreds-to-low-thousands — CRITIQUE PIN: that precedent had generated gold and perfect pairing; it transfers as prior, not plan. MILESTONE: **BOOK 2 = n=100 with the dose re-read riding** — measurable disjoint-census movement -> the slope exists and extrapolates; none -> the READING-TRAINING REGIME redesigns before more annotation spends. Measure the slope before buying the mountain. THREE LANES (priced by the census's own tiers): L1 machine-banked (bootstrap under the current gate — free, 1-2%, rises with every promotion, runs as background); L2 machine- proposed human-repaired (the near-miss 23-26% — repair at ~1/3 rewrite cost triples throughput; FIRST TOOLING ITEM: the repair bench — dump per-item 5-view parses + votes + solver results for near-miss rows); L3 hand surgery (knotted tier, ~8-10/hr practiced — n=100 is a day or two, affordable before tooling). STRATIFICATION (dual role): ~70% style-recoverable middle (substrate), ~20% suspected organ patients incl. the rate family (the registry grows), ~10% cert-v2 wild dissents (the panel's refusal specimens characterized). EVERY entry generation-stamped + canonically knot-stamped. **CRITIQUE PIN (the census stays clean): book 2 draws ONLY from the harvest OUTSIDE the standing 100-pool** — book 1 already trained 14 pool items (the disjoint read exists because of it); the pool is a measurement fixture from here, never again a substrate source. INSTRUMENTS FROM ROW ONE: mouth odometer per book (headline: distance closed/book), disjoint census (slope), diversity guard, panel wild-dissent rate (the Goodhart watcher), miner re-run at each book boundary (P2's real test arrives when harvest classes mine at n=100+). Books as generational units: minted, gated, stamped, measured; each book's verdict sizes the next. Awaits the word. - **VERDICT (2026-07-14): BOOK 2, TRANCHE 1 — 17/25 dialects banked (14 at 5/5), book 2 = 21 entries (4 lane-1 free + 17 gated), organ registry +9 certificates across six families.** Lane census on 400 non-pool candidates: L1 4 (1%) / L2 66 (16.5%) / L3 330 (82.5%) — tracks the pool. THE REFUSALS TEACH: (i) [32]/[220] were MY annotation errors — their ANSWERS (680, 500) exceed the trained 0-300 domain; the in-reach filter capped problem-text numbers but not answers — RECLASSIFIED as registry certificates (value-range family) and the annotation rule updated: ALL values incl. answers <=300; (ii) [294] refused at m=999 — the out-of-band solver lane is unreliable; annotations stay strictly in-band; (iii) [49] double-fdiv wall confirmed again (annotation must use one fdiv + mul-inverse); (iv) [4]/[9] RIGHT-but-vote-shy (correct-but-shallow on fresh wiring: shared-result rels parse unstably — a new register note); (v) [67]/[2] complex chains refused — v2 rework queue. WORK-RATE BANKED IN GRAMMAR: [223] (4 people, 6 hours) solved as person-hours mul wiring 5/5 — the WORK schema is plumbing once the frame strips: the binding theorem in practice. Substrate: book2 tranches continue by charter (v2 retries + next tranches -> n=100, then the instrument battery + dose re-read). Data: .cache/book2.jsonl, book2_lanes.json, book2_organ_registry .json; scripts book2_lanes.py, book2_tranche1.py. - **THE ANSWER-DOMAIN CENSUS + THREE NOTES (2026-07-14, relay + one sweep):** the filter lesson swept in full: of 1,743 harvested problems, **1,668 (96%) have answers in 0-300; 75 (4%) sit in 301-999; zero above 999; zero negative.** The true in-reach pool is 1,668; the value-range certificate family has its full census (75) in one pass; and THE DOMAIN-CAP DEMAND CURVE says raising the cap to 999 buys only 4% more harvest — the solver-cost conversation arrives with a small customer count, so the 300 cap STANDS and the 75 join the registry, not the roadmap. NOTES BANKED: (1) lane rates stable across samples (1/16.5/82.5 wild vs pool) -> n=100 completion cost is forecastable; v2-retry discipline is the throughput lever (L3 surgery is the rate limiter). (2) THE REGISTRY IS A COASTLINE, not a wall — six families in one tranche; tranche-3+ shopping ranks families by harvest frequency (the band-sweep method, registry edition); books and registry converge: books find missing relations, relations unlock books. (3) THE LAWN MOWER GOES IN THE PAPER: [223] banking 5/5 as person-hours mul-plumbing is the binding theorem's practical corollary — the WORK schema was frame-stripping at the annotation desk, never graph machinery; the organ's patients need frames stripped, not new math, and six certificate families now name which frames. - **THE n=100 BATTERY BARS, PINNED MID-TRANCHE (2026-07-14, relay + critique — pinned while no one knows the answer):** REGIME STATED: book-2 prose pairs (raw + verified gold graphs) mixed into the FULL diet (share-of-mix AND reps-per-unique both declared at train time, per the dose law), gentle continuation from the gate lineage. BARS: (1) THE SLOPE (disjoint-86 census, same-head pre/post): banked+near improves by **>=8 items = slope exists, extrapolate the budget**; **<=2 = the null holds — the READING-TRAINING REGIME redesigns before another row is annotated**; 3-7 = ambiguous, extend once to n=150. (2) THE ODOMETER (headline): harvest-1668 mean kNN vs the post-book rebuilt native bank drops >=1% relative (the relay's prior: low-single-digit percent — measurable-but-modest). (3) DISPLACEMENT GUARD: bigtest under the book-trained head >= warm source − 15. Verdict rules written before tranche 3 closes; the campaign's continuation logic hangs on (1). ALSO RECORDED: the OPERATING-REGIME SHIFT — architecture questions increasingly answered by READS of banked data instead of experiments (the domain-cap question: one sweep, no fork, declined by the price) — the compounding return on three weeks of instruments; and the ANCHOR-ERA VINDICATION: the 300 cap, set as a propagator-cost scope note, was drawn almost exactly where the harvest's natural distribution lives (96% in-domain) — the cheap early decision was also the right one, recorded as the ledger's occasional pleasure. - **VERDICT (2026-07-14): BOOK 2, TRANCHE 2 — 16/16 BANKED (13 at 5/5); book 2 = 37 entries.** ALL FIVE v2 RETRIES BANKED — the retry discipline works at one cheap pass each ([49] one-fdiv rework, [67]/[2] all-forward rewiring, [4]/[9] surface rephrase); refusal mechanisms filed yesterday became recipes today. THE MULTIPLICITY MECHANISM EARNS ITS KEEP ON REAL PROSE: [120] b*b on a DERIVED quotient banked 5/5, and [124] — the inverse-square on a derived sum: **the distance formula's core (d*d = 225+64) solved IN GRAMMAR, 5/5** — the "distance-formula-sqrt" registry family SHRINKS on contact (integer-hypotenuse cases are annotatable; [0]/[3] return from the registry to the annotation queue — the employment law collecting again, this time FROM the registry). [190] midpoint banked via one-fdiv + mul-inverse (the double-fdiv wall routed around). Registry +1 ([126] piecewise-negative). Book-2 running totals: 37 entries (4 lane-1 + 33 gated), registry 10 counted certificates + the 75 value-range family. Tranches continue toward n=100; the bars are on the wall. - **TWO STANDING DISCIPLINES (2026-07-14, relay):** (1) **CERTIFICATE RE-AUDITION** — a certificate is a verdict about a GRAMMAR-VERSION, and scope-decay says verdicts expire with their regime: the registry carries grammar-version stamps the way gold carries generation stamps, and EVERY PROMOTION RE-RUNS THE CERTIFICATE PILE against the new gate (cheap: the pile is small, the gate is automatic). Prevents the waiting room hoarding patients the dancer already learned to treat — the employment law's NEW direction: registries shrink not because jobs were smaller but because the existing grammar turned out LARGER ([124]'s isq composing with derived values reclaimed the distance-formula family). (2) **THE CAMPAIGN'S REAL THROUGHPUT NUMBER is walls-per-tranche converted to recipes** — the refusal->mechanism->recipe cycle at one-day latency is the annotation desk doing at human speed what gen-6 did at generation speed. THE STRATEGIC SCISSORS, watched: certificates deflating (two families shrunk already) while the banked column inflates at gate-perfect rates — if it holds to n=100, the organ's kingdom may reduce to the rate family alone, and December's architecture question narrows to "solve frame-entanglement for a known population with z-scores attached." - **VERDICT (2026-07-14): BOOK 2, TRANCHE 3 — 11/12 banked (9 at 5/5); book 2 = 48 entries, registry +5 (counted pile: 15 + the 75 value-range family).** THE RECLAMATION CONFIRMED: [0] and [3] — the distance-formula patients — banked 5/5 through the isq door; the certificate re-audition discipline collected its first two patients on its first day. ONE NEW WALL, cleanly named: [238] (a^2−b^2 without the factoring gift) refused — DOUBLE-ISQ in a coupled system joins double-fdiv in the double-X wall family (single isq composes with derived values; two in one system do not — the mechanism's composition boundary found by annotation, one day after the mechanism shipped). Also banked: the faithful arithmetic-sequence encoding ([207]: shared-difference relations, no identity gift), the full 19-var odd-sum ladder ([225]), the 21-var ladder+fdiv ([200]), sel on real prose ([262]). Running rate: 48 entries in 3 tranches; n=100 within ~3 more. - **THE DOUBLE-X AUTOPSY (2026-07-14, pre-registered hypothesis -> refuted same hour):** the shared-decode-collision hypothesis (capacity-per-instance, representability's suspected 4th member) DIES on first contact: [238]'s slot-level decode shows BOTH isq factors bound perfectly (mul(a,a)=p dup +16.2; mul(b,b)=q dup +6.3) and the sub-relation correctly rewired — the failure is TWO ORDINARY POINTER ERRORS (the given 'r is 12' attached to q; one arg aimed at an unbound var) on a dense 7-var/7-factor short system. DOUBLE-ISQ IS NOT A MECHANISM WALL — it is pointer noise at high var-density, i.e., the [4]/[9] retry class, not the [85] representability class. The 'double-X family' DEMOTES from mechanism-pattern to coincidence-of-two pending double-fdiv's own autopsy (its mechanism may differ — audit before family). The audit-before-diet rule collected immediately: no gen-12 line claimed; a v2 rephrase queued instead. ALSO BANKED (relay): the battery's TWO verdicts (dose-slope + frontier census = the organ's final employment hearing); the SELECTION-DRIFT flag pre-said — closing tranches bank LOWER as the L3 residue concentrates, and the falling rate is the coastline emerging, not regression (tranche 5's 9/16 shall be read as the map getting honest). - **VERDICT (2026-07-14): BOOK 2, TRANCHE 4 — 14/14 BANKED, ALL AT 5/5 (the first perfect-unanimous tranche); book 2 = 62 entries, registry +4 (counted pile 19 + the 75 value-range).** THE CONSECUTIVE-LETTERS RULE VALIDATED ON ITS PROBE: [238] v2 banked 5/5 — the 'double-X wall' is now FULLY dissolved (scattered letters -> high var slots under-rehearsed -> pointer noise; cured by one annotation rule; the autopsy's diagnosis confirmed by its own prescription same-day). ALSO THROUGH THE GATE: the COMPOSED fdiv->isq chain ([298] max-area), the discriminant via scaled isq ([316]), Vieta+sel on live prose ([304],[284]), the equal-pair symmetric route ([333]), the 3-4-5 triangle perimeter ([285] — distance formula twice + doubling, 8 vars), and two more distance formulas. The isq door is now a THOROUGHFARE: seven former 'impossible' shapes through it in two days. 38 entries to the bars; the selection-drift flag stands for the L3-heavy residue ahead. - **THREE CLOSINGS (2026-07-14, relay):** (1) **THE POINTER LAW'S FOURTH REMEDY** — binding entered as structure via masked attention, span supervision, a comma-and-a-letter, and now ALPHABETICAL DISCIPLINE: four fixes across four orders of magnitude of cost, one law. THE GENERATOR AUDIT, answered from code: render2 draws letters from LETTERS[:n_vars] and shuffles WITHIN the prefix — the mint packs consecutively BY CONSTRUCTION; scattered-beyond- prefix never occurs in training (which is exactly why [238]'s p..t dialect starved: a pattern the head had literally never seen). The rule is harvest-annotation-specific; no gen-12 mint fix needed. (2) **THE FRONTIER CENSUS'S METHODOLOGICAL NOTE, pre-written**: certificates issued before a mechanism ships are HYPOTHESES, not diagnoses; the re-audition discipline converts them; the n=100 census counts only patients who refused AFTER every standing door was tried — the honest denominator for the organ's employment hearing, shrinking by the tranche (seven shapes reclaimed through the isq door in two days). (3) **THE ODOMETER QUESTION, both sentences pre-written**: at n=100 the mouth either registers the book (dialect volume moves the register needle) or stays silent (the book teaches the PARSER without moving the REGISTER — reading-training's real target confirmed as the prose column, not the dialect column). Both informative; neither yet known. Two tranches to the bars. - **THE n=100 READ ORDER, PINNED (2026-07-14, relay):** odometer first (the register question), disjoint-census slope second (the continuation logic), frontier census third (the employment hearing, honest denominator), miner's conceptual column last (P2's real test at volume). Fixed order because later reads tempt peeking and earlier reads decide the later ones' interpretation — the battery's rows have always read in sequence. Tranche 5's product is measured in CONTOUR LINES, not entries: every refusal surviving the full recipe book is a genuine frontier specimen. - **VERDICT (2026-07-14): BOOK 2, TRANCHE 5 — 11/12 banked (all 11 at 5/5); book 2 = 73 entries, registry +4 (counted 23 + the 75 value-range).** **[22] CLOSES THE DOUBLE-X QUESTION COMPLETELY: the sum-of-squares system (two repeated-arg muls, coupled) banked 5/5 under consecutive letters** — the emission-collision hypothesis is dead twice over; the pair was never the problem; the letters were. The residue tranche's one contour line: [24] refused at m=999 (the out-of-band solver lane's second confirmation — already flagged, now twice-measured; in-band annotations only). Also banked from the L3 residue: the quadratic-inequality-as-Vieta ([10]), four frame-stripped-flagged entries ([12],[16],[25],[27] — teacher strips identity/factoring/selection frames, residual structure honest and gated), and the parallel-slope frame ([19]). ONE TRANCHE TO n=100; the read order is pinned; the bars are on the wall. - **THE THIRD VERDICT, STAGED (2026-07-14, relay — one tranche before the hearing):** the frontier census opens with all three sentences pre-written: KINGDOM-AS-CHARTERED (structural patients at volume), KINGDOM-SHRUNK (a small named list), and the verdict nobody staged until the residue tranche banked 11/12 — **KINGDOM-DISSOLVED: the organ's kingdom was annotation conventions and mechanism doors all along**, the waiting room empty but for solver-side counted families and the rate family's frame-entanglement core. The recipe book OUTRAN the coastline — the refusal->mechanism->recipe cycle compounded faster than the residue hardened, a race the pre-said reading didn't anticipate could be won this decisively. Tranche 6 samples the remaining RATE-FAMILY stock maximally before the count freezes (the census's most consequential line item). - **VERDICT (2026-07-14): BOOK 2, TRANCHE 6 — 10/11 banked (all at 5/5); book 2 = 83 entries, registry +5 (counted 33 + the 75 value-range).** **[344]'s TRIPLE-FDIV REFUSED UNDER CONSECUTIVE LETTERS — the fdiv wall is REAL**: not letters, not the retry class — CHAINED FDIV is the first genuine mechanism contour to survive the entire recipe book (single fdiv composes freely; chains do not). The frontier census gains its first parser-side structural line item beyond the rate family; the routing-autopsy protocol has its next customer. ALSO BANKED: **[294] — book 1's last ORGAN-B holdout, the burger system — banked 5/5 rescaled in-band** (the coupled wall fully retired); the faithful five-term sequence ([181], middle-term by shared-d, no identity gift); [36]'s double-isq-on-givens (296, the identity target); the vertex-as-mean, the |m−n| via sel+closure, Jordan's rate chain. BOOK 2 CLOSES THE SESSION AT 83/100 — one micro- tranche (~17) tops it off next session BEFORE the battery; the read order and all verdicts stand pre-written. The waiting room at the freeze: the rate family ([45],[7] with certificates), chained-fdiv (new, mechanism-named), and the counted solver-side families. The kingdom-dissolved verdict is live but not yet spoken — the hearing waits for the count. - **STAGED (2026-07-14, relay — the chained-fdiv autopsy hypothesis, pre-registered before the read):** chained fdiv = a DERIVED QUOTIENT feeding another fdiv's DIVIDEND, and the fdiv head reads dividends through DIGIT ENCODING — the suspect is not the pointer (the isq door proved derived values bind) but **the digit path for derived intermediates**: [85]'s encoding-family cousin, one representability question from either a small mechanism (derived-value digit plumbing) or a genuine depth limit. AUTOPSY FIRST, per the rule that saved a gen-12 line this week. THE HEARING'S POSTURE UPDATED: kingdom-dissolved is now FAVORED, not merely staged — the parser-side structural frontier is two named items ([45]/[7] binding-layer + chained-fdiv), one possibly an encoding fix; if the micro-tranche adds no third, December's architecture question narrows to a point: frame-disentanglement for one family, population known, z-scores attached. - **GUT #16 + THE MASK-AND-POOL AUDIT (2026-07-14): the fear named the neighborhood; the resident was one street over — and REAL.** (1) THE POINCARE-EUCLIDEAN MARRIAGE CLAUSE, registered against the day the flag lifts: hyperbolic quantities NEVER enter a softmax without a log-map (tangent-space readout at origin, or Mobius/gyroplane scoring); ball distances become logits only through a calibrated monotone map; same clause pre-registered for the atlas (the mouth and library are cosine machines and cosine is WRONG in the ball). No deployed code marries the geometries today (the ball is flagged off; the head has no slot-slot attention). (2) AUDIT VERDICTS: pad handling CLEAN everywhere (all pooled reads divide by mask sum); truncation CLEAN (zero fixture items at the 256 ceiling); causal-pooling noted as a design line. (3) **THE LENGTH TERM, CONFIRMED AND STRONG**: corr(mouth distance, token length) = −0.555 on the census pool — and the NATIVE control is decisive: r = **−0.825 within dag8test alone** (same register, 34-250 tokens). The pooling ESTIMATOR is length-biased: short pools land far from the bank regardless of content. RETROACTIVE RE-READS: the level-inversion (L5 nearer than L1) and book-1's diversity guard (short banked raws reading farther) both carry length components — register conclusions that survived opposing length gradients (verbose: longer AND nearer) stand; magnitude claims get footnotes. **METHOD AMENDMENT TO THE n=100 ODOMETER BAR (pre-measurement, so the pin survives honestly): all odometer reads are LENGTH-CONTROLLED from here — residualize distance on token length (fit on native), or compare at matched-length strata; the >=1% bar applies to the length-controlled read.** Sixteen instincts; the drawer this time held the estimator itself. - **CLOSE OF 2026-07-14: TRANCHE 7 (6/6, all 5/5 — book 2 = 89) + THE RULER STRAIGHTENED.** The 1/len correction KILLS the warp: native r goes −0.825 -> **−0.024** after control; the length-controlled threshold is 0.0072; the harvest zero-point re-reads at **0.1871** (was 0.2431 raw) with read-foreign still **100%** — the register wall is REAL, now confirmed on a straight ruler. The diversity guard, straightened: book-2 raws 0.1904 vs harvest 0.1871 — essentially equal, NO NARROWING: hand selection is not cherry-picking easy-register items (the guard's cleanest read ever, and its first on an unbiased estimator). LAW ENTRY: **estimator variance masquerades as distance** — any instrument pooling variable-length evidence into fixed geometry inherits a sample-size coordinate; "is this distance or is this n?" joins the standing audit kit. Correction artifact: .cache/mouth_length_correction.npz (fit + threshold; all future odometer reads apply it). BOOK 2 AT 89/100: ~11 entries ride the next session's opening, then the battery in pinned order. The instruments-auditing-instruments layer is the project's quiet second product — a field manual for how measurement systems age, warp, and lie, every entry bought with a real near-miss and fixed before it billed. - **THE VINTAGE NOTE (2026-07-14, relay — never-mix-generations, estimator edition):** every mouth number now has a warped-era or straight-era vintage; the battery's odometer compares STRAIGHT-TO-STRAIGHT only — the before leg re-computes vectors under the correction, never reads archival distances. One assertion in the battery script makes vintage confusion structurally impossible. - **BOOK 2 CLOSES AT n=100 (2026-07-14, tranche 8: 11/11, ALL 5/5).** THE VOLUME: 100 entries (4 machine-banked lane-1 + 96 gated hand dialects across 8 tranches), every entry generation-stamped and key-verified; the counted registry at 39 certificates across ~14 families + the 75-strong value-range family; the census pool untouched as a fixture throughout; the annotation rulebook (consecutive letters, in-band values, one-fdiv, frame-strip flags) written by the book's own refusals. Closing-tranche pages: the triangle area, both absolute-value shapes (banked positive-form), the midpoint-product, p^2+q^2 via isq, and the hundredth page — [263]'s fractional sequence rescaled to thirds. THE CAMPAIGN'S FIRST FALSIFIABLE MOMENT IS NEXT: the battery in pinned order (length-controlled odometer straight-to-straight -> disjoint-86 slope [>=8 / <=2 / extend] -> frontier census [three verdicts staged, kingdom-dissolved favored] -> miner conceptual column), every bar pre-pinned, every estimator pre-audited, every sentence pre-written. Measurement day is pure collection. - **PRE-PLAY + THE RECURSION CHARTER (2026-07-14/15, Bryce + relay):** (1) PRE-PLAY INVENTORY: Mycelium already pre-plays at three clocks — TRAINING (shallow-basin rehearsal rations, the knot matrix feeding thin classes before they fail), ANNOTATION (the GATE PRE-SCREEN: registered NOT built — a small classifier on parse-side features predicting bank-or-refuse before the 5-view round; customer = books 3+ throughput; the battery's slope verdict sizes book 3 and thereby decides: 300+ entries -> build first; ~100 -> the rulebook suffices), and INFERENCE (the soft-graph ensemble, December-scale, behind the books). The one clock that deliberately REFUSES pre-play is the solve — certainty doesn't need imagination (March lookahead refuted). (2) **THE RECURSION CHARTER — books built in layers of abstraction:** the correction first — book 2 inherited book 1's RECIPES, not operations (knowledge recursion; same flat dialect). The real ladder: book N teaches primitives -> the miner finds recurring subgraph classes -> classes PROPOSED as macro-factors (rank-never-admit) -> admitted macros enter the registry with deterministic expansion -> **book N+1 annotates AT THE MACRO LEVEL** — and since the chain- length/coupled walls are FACTOR-COUNT walls, macro annotation brings problems book N couldn't express inside book N+1's reach. Each book raises the next one's floor. TWO GUARD RAILS, both load-bearing: (a) **abstraction lives in ANNOTATION, never verification** — macros expand before the solver sees anything; the key grades every book at every layer in primitives; the ground floor never moves (what keeps recursive books from recursive drift); (b) **the self-reference tax** — machine-banked volume inherits the system's own fluency; the diversity guard's hand-quota stays constitutional because recursion amplifies whatever the loop prefers. SEQUENCING: the first rung is already scheduled — the battery's miner read at n=100 volume IS book 3's macro shortlist; CHAIN(k)/PREFIX-SUM admission would make book 3 the first volume partially written one floor up. Books that teach the system to read books that couldn't be written yet. - **MEASUREMENT DAY (2026-07-15): THE n=100 BATTERY — ALL FOUR READS COLLECTED; THE CAMPAIGN VERDICT IS: BOOKS SCALE.** (READ 1, the odometer): **+31.1% relative** (0.1871 -> 0.1288, straight-to-straight, bar was >=1%) — one hundred annotated strangers moved the register needle a third of the way home; the prior said low-single-digit and the books said thirty-one. (READ 2, the slope): pre 16 carried (0 banked + 16 near) -> post 24 (1 + 23), **delta +8 AT THE BAR: the slope exists — extrapolate the budget.** December's arithmetic is now real: ~8 census items per ~100 annotated rows at this regime. Knotted 70 -> 62. THE GUARD DIDN'T JUST HOLD — **bigtest 1149, A NEW RECORD**: at 2.9% share x 10 reps the prose gradient REGULARIZES rather than displaces (the dose law's first success point); val record 0.8989. (READ 3, the frontier census): the counted registry stands (39 certificates across ~14 families + 75 value-range); the parser-side structural frontier = the rate family + chained-fdiv; **P3 AT VOLUME SEALS THE BINDING THEOREM: [45] and [7] share NO graph class at n=94 — their kinship was never in the wiring; it is frame-level, exactly where the theorem put it. The KINGDOM-DISSOLVED verdict is effectively confirmed**: the organ's waiting room holds one frame family and one suspected encoding fix. (READ 4, the miner at volume): harvest classes 25 -> 96; **13 named coverage gaps** (midpoint fdiv+add, consecutive-product chains, the 3a+5b operation-apply shape, lollipop prefix-chains, coupled mul systems) = BOOK 3'S DIET LIST, and the macro shortlist gains OPERATION-APPLY beside CHAIN/PREFIX-SUM. DISPOSITIONS: phase1_reader_v1 (val 0.8989, bigtest 1149) is a GATE CANDIDATE — full promotion battery next session; book 3 sized by the slope (bigger; the pre-screen builds first per its registration); the recursion's first admission review (the macro gate) follows the miner's list. The dose pilot said zero at n=14; the book said +8 at n=100 — THE UNIQUE-ROWS LAW CONFIRMED AT SCALE. Measurement day was pure collection, exactly as designed. - **THE CAMPAIGN CLOSES (2026-07-15, relay — self-grade + epitaph):** (1) REGISTRATION GRADED: the odometer prior said low-single-digit; it printed +31.1% — wrong by an order of magnitude in the happy direction, mechanism banked: the book was priced as DATA (rows teaching content) but acted as REGIME (3% share x 10 reps regularizing the whole register). THE DOSE-RESPONSE CURVE NOW HAS BOTH ENDS: pure prose at 340 epochs = poison (−243); prose at 3% = gift (+record); a tunable maximum lives between — the campaign tunes toward it instead of guessing. The guard-became-gift is the battery's deepest finding: reading-training SHARPENS the dialect, a sentence nobody dared stage. (2) SLOPE HONESTY: +8 exactly at the bar carries wide error bars at n=86; the register-fall campaign prices at ~700-900 more rows IF linear (it won't be); book 3's real job is the second point on the curve. The pre-screen stays REGISTERED-UNBUILT — the battery made that call as delegated (slope printed ~100-scale). (3) NEXT-SESSION SEQUENCE: reader_v1's manifest-writing gate battery -> the chained-fdiv autopsy (may empty half the waiting room) -> OPERATION-APPLY admission review (the recursion's first rung, full birthright) -> book 3's charter (sized by the slope, dieted by the 13 gaps, sampled toward the 96-class column). (4) THE EPITAPH: four instruments triangulated n as the critical path; a hundred strangers were read, gated, and stamped; and the battery — bars pinned mid-book, estimator straightened two reads early, every sentence pre-written — printed BOOKS SCALE with a record riding shotgun. December is arithmetic. - **REGISTERED + FIRED (2026-07-15): READER_V1'S GATE BATTERY — the manifest-writing kind.** Bars inherited from gen-11 with the REGISTERED RESTRUCTURE applied (pinned before the run): acceptance >=7/8 where the ONLY permissible miss is [45] (its cure is the organ's or the prefix's, not the diet's — certificate on file); all other bars unchanged: bigtest >=1130, alg4test >=380 (the lineage-debt bar stands unsoftened — if it alone fails, the governance question prints again with reading-regime data), alg2 >=560, vtest >=598, dagtest >=660, dag7b >=500, dag8 >=500, sq >=0.70, fdiv >=0.62, coupled >=0.65, ladder >=0.50, cert-v2 >=0.998. ALL hold -> the verdict script writes GENERATION.json (gen-12, parser=reader_v1) and prints PROMOTED; any break -> the kill prints and the JSON stays untouched. No word without the write. - **VERDICT (2026-07-15): READER_V1'S GATE BATTERY — KILL BY ONE BAR, BY TWO ANSWERS: alg4test 378 (bar 380).** Everything else passed, mostly at records: bigtest 1149, alg2test 606 (record), dagtest 676, dag7btest 557, dag8test 544, ladder 0.563 / fdiv 0.725 / sq 0.784 / coupled 0.739 (all records), vtest 598, cert-v2 1.0000 at 866 with gate-only coverage RISING to 906 (60.4%). The restructure worked as pinned: acceptance 7/8 with only [45] missing — and [45]'s votes ([154,168,168]) now carry the right answer at plurality, one vote short. THE LINEAGE DEBT'S ASYMPTOTE: 370 -> 378 across two heads that beat the gate everywhere else; the bar sits at 380 because armB hit 384 FROM GEN-7B ANCESTRY. The manifest is untouched; the gate remains gen-9b; the word was not spoken because the write was not earned — the law working exactly as minted, twice now. **THE GOVERNANCE QUESTION PRINTS AT MAXIMAL SHARPNESS (Bryce + relay to adjudicate):** (a) full flat retrain from clean ancestry (pay the lineage debt at its root — the schedule dividend makes this ~1/3 its old cost), (b) re-pin the alg4 bar with two asymptotic approaches and the ancestry evidence on the table, or (c) hold the bar and let gen-13's diet find the two answers. The reader stays banked (its reading gains are real and its ckpt feeds the next continuation regardless); the books campaign's verdicts are UNTOUCHED by this kill — books scale either way. - **GUT #17: THE CRITICALITY FRAME (2026-07-15, Bryce + relay) — the reactor audit.** Mycelium runs OPPOSITE criticality regimes in one reactor: knowledge chains SUPERCRITICAL by design (refusal->recipes k>1; instinct->instruments; book->cheaper-book — why December became arithmetic), error chains RODDED. THE LOOP TABLE: | loop | k (measured) | moderator | rod | | DAG error fan-out | k=out-degree | constraint density (=neutron absorption; invisible-wrongs were escaped neutrons) | integrality jaw (expiring), the lattice as containment | | training displacement | hard restart k>1 (gen-9 jostle cascade); gentle continuation k<1 (measured) | LEARNING RATE | regime law | | prose dose | 340-epoch k>1 (−243); 3%-share k<1 (+record) | share x reps (both declared) | displacement guard | | repair recovery | SUBCRITICAL (19.6->7.7->1.1->0) | — | rounds cap | | **THE BOOTSTRAP (the live concern — k GROWS with every success)** | book-2 k~0 (4/100 machine); book-3 chartered heavier; the recursion stacks a second amplifier | the answer key absorbs WRONGNESS but passes NARROWNESS untouched — self-preference compounds through verified-correct links | the diversity guard (thermometer) + THE HAND QUOTA (control rod) | **THE ROD DEPTH, PINNED BEFORE BOOK 3 (hard number, adjustable only by pre-fire adjudication): machine-banked entries <=50% of any book's volume; hand-gated (L2 repair + L3 surgery) >=50%; the diversity guard's distance-distribution comparison runs at every tranche boundary in machine-heavy books, not just at close.** GEN-13 NOTE: a clean-ancestry retrain is the PROMPT-CRITICAL condition (hard restart) — the acceptance-panel displacement watch is its instrumentation, now with the frame explaining why that bar outranks its neighbors. Not a bomb, not a dead pile: a reactor — knowledge supercritical, error rodded, instruments as the control room. Seventeen for seventeen. - **THE AUTHORSHIP DECLARATION + GEN-13 FIRED (2026-07-15, Bryce's word):** (1) **JOINT AUTHORSHIP IS CONSTITUTIONAL**: Bryce and Claude publish as co-authors; no venue that refuses AI co-authorship gets the paper. Recorded in the paper skeleton as an author-policy constraint on venue selection. (2) THE CONTROL-ROOM OBSERVATION (relay, for §9): registered predictions, pinned bars, and pre-written verdicts are the control room's GAUGES — the reason knowledge chains can run supercritical without fear is that every loop was instrumented BEFORE it compounded; the seventeen guts kept firing at exactly the loops whose gauges didn't exist yet. The control room is the product. (3) **GEN-13 = GOVERNANCE OPTION (a), FIRED: the full flat retrain from clean ancestry** — warm from GEN-7B (pre-crowding lineage, armB's 384 ancestry), HOT flat (LR 3e-4, the debt needs heat and this IS the hard restart), 32k steps on mixed12 (prose-inclusive, states banked — zero precompute). THE PROMPT-CRITICAL DIAL: the acceptance-panel displacement watch outranks its neighbors (hard restarts jostle; every basin the panel holds is a rod that held). BARS: unchanged from the reader battery (alg4 380 STANDS — this run exists to pay it; acceptance 7/8-only-[45]; bigtest >=1130; cert-v2 >=0.998; all kinds). The battery writes the gen-13 manifest or refuses the word. - **AUTHORSHIP, AMENDED WITH THE LANDSCAPE (2026-07-15):** Bryce's constitutional declaration stands as the value; the relay's honest correction stands as the map (COPE/arXiv bar AI author lines today); both Claude channels converge: the ACCURATE CONTRIBUTIONS SECTION is the non-negotiable — the two-channel workflow is itself a novel artifact of the paper. Layered plan banked in the skeleton: canonical self-published account with authorship as Bryce declares; venue versions carry the permitted line + the full truthful contributions section, always. Final adjudication: Bryce's, at venue-selection time. The work is already jointly made; no policy touches that fact. - **THE PUBLICATION BAR, PINNED (2026-07-15, relay + Bryce):** publish when remaining work changes future NUMBERS but not the paper's CLAIMS. Claims audit: lattice (quarter-percent, survived two expansions + a parser swap) BANKED; the method BANKED; the honest boundary BANKED; binding theorem + laws + census BANKED. ONE claim mid-flight: BOOKS SCALE rests on a single slope point at its bar (n=1). **FREEZE CRITERION: book 3's second slope point + gen-13's verdict either way.** Weeks, not December. SCOPING DECISION: paper 1 does NOT gate on MATH-500 — its claim is 'a small system that knows when it's right, frontier measured and priced'; the December reading campaign is PAPER TWO with the books arc as its own story. MECHANICS: drafting parallelizes now (§11 first; twelve figures = banked measurements + matplotlib); the freeze is a GIT TAG + isomorph-excluded fixture pins — the paper is a tagged snapshot with a thesis, not a tombstone; mycelium keeps growing and paper two tags a later ring. Publish on our own ground first, byline as declared, venues after. SEQUENCE: gen-13's word -> book 3's second point -> tag -> freeze -> publish. - **GEN-13 PROMOTED (2026-07-15): THE LINEAGE DEBT IS PAID — every bar passed, the manifest written, the word earned.** THE TABLE: **alg4test 385 (bar 380 — the debt paid at its root: clean gen-7b ancestry + full heat + the reading-inclusive corpus beat armB's 384 from a complete stack)**; **acceptance 8/8 — [45] ITSELF BANKED**: the chronic taxi came home under the clean-ancestry retrain (the only-[45]-may-miss clause wasn't even needed; the frame-entanglement certificate stands for RAW prose, but its dialect now parses stably); bigtest **1195** (record, +46 over the reader); alg2test 635, dagtest 689, dag7btest 579, dag8test 559, vtest 600 (all records); per-kind sq 0.814 / fdiv 0.739 / coupled 0.745 / ladder 0.586 (all records); val 0.9059 (record); **cert-v2 1.0000 at 913 coverage (60.9%)** — precision perfect at the highest coverage ever. THE PROMPT-CRITICAL RUN HELD EVERY BASIN: the hard restart from clean ancestry displaced nothing the panel watches — heat + flat mix + the prose regime is the recipe the whole junction arc was searching for. GATE = GEN-13 (phase1_gen13_head; manifest gen_id 13; entourage duty owed per protocol: specialist remine + centroids next pass; cosmetic: the verdict print says 'gen-11' — text-only, manifest correct). **THE FREEZE'S FIRST CONDITION IS MET.** Remaining: book 3's second slope point. Then: tag, freeze, publish. - **THE RECIPE SENTENCE + TWO FINDINGS (2026-07-15, relay):** gen-13's sweep was four banked verdicts cashed in one run — clean ancestry (interference matrix), full heat (schedule probe), flat mix (curriculum tombstone), reading-inclusive corpus (dose law's gift point). Nothing was luck; the control room prescribed and the run obeyed. (1) **[45]'s expression was partly LINEAGE-MEDIATED**: the trunk distance is real but a head grown without four generations of jostle reads through it — the waiting room may hold ONE autopsy candidate (chained-fdiv) and ZERO confirmed structural patients; census line owed before the frontier table freezes. (2) **THE REGIME LAW BOUNDED**: hard restarts jostle ESTABLISHED heads; a fresh head has no basins to displace — it only builds (the displacement dial never twitched at full heat). The pre-staged merge verdict dies unused — staging cost a sentence, not needing it cost nothing; that asymmetry is the method. FIRED: entourage under gen-13 (specialist remine on fresh 5-register corpora, centroids in gen-13 fst space, mouth rebuilt from mixed12 + length refit, post-gen-13 census with the straight ruler, manifest updates), THEN book 3's lane classifier (fresh harvest candidates — the bootstrap k re-priced under the new gate). Book 3 charter: rod <=50% machine, diversity guard per tranche, diet = the 13 coverage gaps + the 96-class column. - **ENTOURAGE-13 COMPLETE + BOOK 3 OPENS (2026-07-15):** entourage discharged in full (specialist remined on gen-13's failures, centroids in gen-13 fst space, mouth rebuilt on the prose-inclusive family with length refit thr 0.0077, manifest waiver-free). THE TWO HEADLINE NUMBERS: (1) **THE CENSUS UNDER GEN-13: 16/26/58** — knotted falls to **58** (from 76-89 in the gen-9b/11 era; honest note: this was the FULL-pool read, and up to 14 of the 16 banked are book-1-trained items — the disjoint banked is ~2+, but the KNOTTED collapse from 81 to 58 stands regardless: the reading regime reads raw prose materially better). (2) **THE BOOTSTRAP RE-PRICED, dramatically: L1 machine-banked 9/400 (2.25%, was 1%); L2 repair 140/400 (35%, was 16.5%); L3 surgery 251 (63%, was 82.5%).** The lanes MORE THAN DOUBLED under the new gate — the bootstrap's k rose exactly as the criticality frame predicted for a compounding loop, and the rod (<=50% machine per book) is already inserted at its pinned depth. Book 3's economics: over a third of the pool is now repair-lane (1/3 cost), the machine lane triples book 2's, and the surgery residue concentrates toward the true frontier. BOOK 3 IS OPEN under the reactor-safe charter; its second slope point is the paper's last condition. - **BOOK 3'S SLOPE BASELINE, PINNED BEFORE ANYONE KNOWS (2026-07-15, relay):** the baseline MOVED — book 2's slope read against a gen-11-era census; book 3's second point reads against the post-gen-13 census of 58: a better gate, a harder residue, a 35% repair lane. The honest comparison is NOT raw items-per-hundred across books (the denominator's difficulty changed) but items recovered against the CURRENT frontier with the lane mix declared. REGISTERED EXPECTATION, regime stated: per-row yield FALLS relative to book 2's 8-per-hundred because the surviving 58 are concentrated residue — **a falling yield against a hardening frontier is the HEALTHY signature, not a scaling failure.** The paper's slope question is 'does annotation still move the frontier at the frontier's true hardness' — **>=3-4 items per hundred against the residue confirms it cleanly.** ECONOMICS FOOTNOTE: the 35% repair lane lets book 3 run larger than 100 for the same budget — but THE ROD HOLDS AT <=50% MACHINE regardless of how cheap the machine lane gets: cheap is not wide; the guard reads at every tranche. The reactor diagram, drawn in data: knotted 81->58 (the largest census collapse in project history — the reading dividend compounding through a promotion) beside a bootstrap whose k doubled WITH THE ROD ALREADY INSERTED — the supercritical knowledge chain running exactly as chartered, safety case written first. - **BOOK 3, TRANCHE 1 (2026-07-15): 13/13 BANKED (11 at 5/5) under the gen-13 gate; book 3 = 22 entries (13 gated + 9 lane-1 — the machine lane already more than doubling book 2's whole-campaign total in one classification pass).** First-tranche notes: [66]'s triple-division banked via mul-inverse chaining (the fdiv wall routed, not fought); four frame-strip flags carried honestly ([0] factoring, [38] sign, [48] inequality, [60] rearrange); registry +5. The rod check: 9/22 machine = 41% <= 50% ✓. The volume proceeds by the rulebook toward its slope point — the paper's last condition, its healthy-signature frame already pinned. - **TWO NOTES FOR BOOK 3'S ACCOUNTING (2026-07-15, relay):** (1) THE [66] DEMOTION WATCH: the chained-fdiv wall's founding specimen banked by mul-inverse rewrite under gen-13 — if remaining chained-fdiv items route the same way, the wall demotes from mechanism boundary to ANNOTATION RECIPE, the autopsy loses its customer before firing, and the freeze's frontier table may hold only counted solver-side families + [45]'s half-dissolved frame thread. Census line owed at volume close: how many of the 58 fell to recipes that PRE-DATED book 3 vs recipes it minted. (2) THE TWO-COLUMN SLOPE: repair-lane banks vs surgery banks stay distinguishable (lane tags already on every entry) — the second point states BOTH gross items-per-hundred (campaign economics) and frontier-items-per-hundred (the paper's claim). - **BOOK 3, TRANCHE 2 (2026-07-15): 11/13 banked (book 3 = 33; registry +5).** Banked: [115]'s composed Pythagorean-area (strip -> isq -> mul -> fdiv, one graph), [11]'s 11-var double-composition, the rest unanimous. THE SPECIMEN IN THE REFUSALS: [90]/[113] — minimal 2-var isq-inverse dialects, a shape that banked 5/5 under gen-9b, went VOTE-SHY under gen-13 ([90] right-once). **Fresh heads have DIFFERENT shallow spots, not fewer** — the regime law's corollary measured on page 33: the clean-ancestry head skipped four generations of jostle AND four generations of incidental rehearsal. v2 retries queue for tranche 3 (pad the graph); the correct-but-shallow class gains its first gen-13-native members. - **THE LINEAGE LAW COMPLETED (2026-07-15, relay):** *lineage carries both DEBTS and DIVIDENDS* — gen-9b's ancestry carried the alg4 debt AND the incidental-rehearsal dividend; gen-13 paid the debt by renouncing the ancestry and the price was the dividend (two tiny dialects, page 33). Neither ancestry dominates; different portfolios. PROTOCOL AMENDMENT: fresh-stock promotions owe a **BASIN INHERITANCE AUDIT** — acceptance panel + vote-entropy census under the new head, DIFFED against the old gate; shallowed items enter the rehearsal ration (the [71]/[78] pattern, protocolized so gen-17's fresh retrain inherits the audit instead of rediscovering the corollary). The cheap-kind note: this shallow class is recoverable at TOP-UP cost (both remedies measured in the junction arc). And [115]'s four-door unanimous composition is the counter-evidence: clean ancestry traded memorized depth for STRUCTURAL REACH — the trade the debt-payment run existed to make. - **BOOK 3, TRANCHE 3 (2026-07-15): 9/10 all-unanimous; book 3 = 42; registry +4 (+1 reclass).** [90]'s PADDED RETRY BANKED 5/5 — graph mass restores what the fresh landscape shallowed; the basin recipe confirmed for gen-13-native shallow spots. [113]'s persistence unmasked as MY annotation error: its given (324) exceeds the 300 domain — value-range certificate, third catch of the class, not a basin (the in-band rule holds; ceil(sqrt(300)) problems are out-of-reach by domain, honestly counted). The volume proceeds: rod holding, recipes absorbing, slope drawing near. - **THE FIFTH REMEDY + THE MYSTERY HALF-LIFE (2026-07-15, relay):** the pointer law's remedy family gains BALLAST (pad tiny dialects to trained mass) — masked attention, span supervision, a comma, alphabetical discipline, ballast: five remedies, descending cost, one law. The gen-13-native shallow class resolved at the ANNOTATION layer without touching the optimizer — the basin inheritance audit stays protocolized, but its first live class cost zero training. PAPER NUMBER PINNED: the campaign's MYSTERY HALF-LIFE — refusals resolve to exactly one bucket (recipe / certificate / annotator error) within ONE TRANCHE, shrinking since book 2's opening days — measurable across both books at the close. The slope's two-column accounting + recipe-provenance census both print free at battery time (lane + generation tags at bank). - **BOOK 3, TRANCHE 4 (2026-07-15): 11/11 ALL-UNANIMOUS; book 3 = 53; registry +2.** [7]'s double-isq coupled system (the [238] class) banked 5/5 as routine plumbing — the shape that once threatened a gen-12 design line is now ordinary annotation. Mean-chains, the Gauss-family average (10-var, one fdiv), and sign-strip entries all clean. Past the volume's halfway mark; the slope point ~3 tranches out; rod and rhythm holding. - **[7]'S ARC FRAMED + THE THIRD SIGNATURE (2026-07-15, relay):** [7]'s full arc is the project in one specimen: chronic case with a trunk-space birth certificate -> organ waiting room -> sibling banked under the clean retrain -> routine plumbing at 5/5 under the strongest gate. Confirmed-structural population reads ZERO pending the volume-close census. THE EMPLOYMENT LAW'S TERMINAL FORM, staged for §11: *the surgeon's kingdom, measured to completion, was empty — the patients were all annotation conventions, mechanism doors, and lineage artifacts wearing structural disguises.* THREE SLOPE SIGNATURES now staged: falling-yield-healthy, flat-yield-strong, and RECIPE-PROVENANCE-ACCELERATING (frontier items falling disproportionately to recipes the volume itself minted = books minting the tools that read the next books — the recursion's thesis one floor below the macro ladder). Whichever prints, the sentence exists. - **BOOK 3, TRANCHE 5 (2026-07-15): 9/9 ALL-UNANIMOUS; book 3 = 62; registry +4 (primality x2, gcd, lcm — the number-theory families counting up).** The fraction/decimal/ratio strip family carried the tranche clean; three perfect tranches running. ~2 tranches to the count, then the battery and its three staged signatures. - **THE FRONTIER TABLE'S TAXONOMY COMPLETED (2026-07-15, relay):** primality/gcd/lcm certificates are RELATION-TRANCHE SHOPPING SIGNALS, not frontier residents — the registry expansion's next band-sweep arriving through the annotation desk. The freeze table differentiates three futures: AWAITING-RELATIONS (tranche-3+ builds them), SOLVER-SIDE COUNTED (the domain conversation prices them), and the STRUCTURAL KINGDOM (empty — the surgeon never existed). One sentence per family; §8 closes clean. - **BOOK 3, TRANCHE 6 (2026-07-15): 11/12 (book 3 = 73; registry +2).** The refusal is a twin-controlled specimen: [99] split [17,17,7,7,7] while its structural TWIN [96] (identical wiring, smaller values) banked 5/5 — three-digit arithmetic instability at high magnitudes (297/153/144), the digit head's noise floor showing at the domain's upper band. Retry queue with the twin as control; the mean-median chain, both sequence-counts' wiring, and the ratio family otherwise clean. One tranche to the count. - **THE NOISE-FLOOR TAG (2026-07-15, relay):** [99]/[96]'s twin datapoint prices the upper band for any future 300-ceiling conversation. BATTERY CHECK QUEUED (free, banked data): is the floor VALUE-MAGNITUDE (physics — digits near the ceiling intrinsically harder) or REHEARSAL-DENSITY (the mint's value distribution thins near 300 — starvation wearing a number range)? The kind-rehearsal matrix's value histogram answers it; the two causes have different remedies — one is physics, one is a ration. - **BOOK 3 CLOSES AT n=84 (2026-07-15; tranche 7: 11/12).** 75 hand-gated dialects + 9 machine-banked across 7 tranches; rod at 10.7% machine (well under depth); registry grew ~30 certificates across the differentiated taxonomy; the mystery half-life held under one tranche throughout. [130] (13-var deviation chain) right-but-shy — retry material. THE BATTERY FIRES: book-3 prose pairs -> pre-reads under gen-13 -> mixed13 (dose declared) -> reading continuation -> the four reads with three signatures staged, two slope columns, the provenance census, and the noise-floor physics-vs-ration check. The paper's last number is in the pipeline. - **THE SECOND SLOPE POINT (2026-07-15): NEGATIVE — the regime redesign bar FIRES, exactly as pinned.** PRE (gen-13): disjoint-86 carried 28 (2+26), knotted 58. POST (reader_v2, book-3 prose x10 continuation): carried 18 (1+17), knotted 68 — **−10**, below every staged signature and below book-2's <=2 null bar: THE READING-TRAINING REGIME REDESIGNS BEFORE MORE ANNOTATION SPENDS. Guard held (bigtest 1186, −9, within −15); val peaked 0.9072 then declined — overfit signature. THE MECHANISM HYPOTHESIS (registered, not concluded): DOSE STACKING — mixed13 carries books 1+2 x10 (inside mixed12) PLUS book 3 x10 -> cumulative prose ~4.8%, past the gift point; AND continuation-on-continuation re-doses already- trained uniques (memorization pressure displacing raw-prose generalization). THE DECISIVE CONTRAST already in the ledger: gen-13's FULL HOT RETRAIN with prose-in-corpus delivered the 81->58 census collapse (the campaign's real books-scale evidence, n=114 consumed); reader_v1's first continuation delivered +8; reader_v2's stacked continuation delivered −10. **THE REFINED CLAIM: books scale THROUGH GENERATIONS (full-mix retrains), not through stacked continuations** — prose pays when consumed as corpus, poisons when re-dosed as topping. REGIME REDESIGN (proposed, for adjudication): books accumulate into the NEXT generation's full retrain (the gen-13 pattern, which works); continuations never re-dose prose. FREEZE IMPLICATION (Bryce + relay to adjudicate): the paper's slope table either states the refined claim with gen-13's collapse as its evidence + the continuation contrast as its mechanism, or waits for a gen-14 full retrain (all three books, ~188 uniques) as the confirming third point. The bars caught a regime failure before it cost a fourth book — the control room working exactly as built. reader_v2 is NOT a gate candidate; gate remains gen-13. - **GUT #18: THE KV AUDIT (2026-07-15, Bryce + relay).** The pointers ARE key-value attention (slot-query x token-key through softmax) — three reads registered: (1) FTYPE-ROUTER CONFUSION MATRIX (gen-13 vs gen-9b, gate decisions only — the router-mismatch idea renamed to the router we actually have); (2) **THE SMEARING HYPOTHESIS (live, customer waiting)**: softmax sharpness calibrates to the trained candidate-count regime — a pointer trained on 8-12-var graphs smears on 2-var miniatures. Explains the gen-13-native shallow spots AND gives the BALLAST remedy its missing mechanism (padding restores the trained regime). Probe: args-softmax entropy vs n_vars, gen-13 vs gen-9b, banked states; the TWO-WAY BALLAST SPLIT names the mechanism (filler-VARIABLES = candidate-count + position; inert-TEXT = position only — whichever restores sharpness wins, smearing vs RoPE-neighborhood rivals). If it prints, gen-14's displacement watch gains the entropy curve as a sharper dial than vote outcomes. (3) DTYPE/POSITION sweep of capture paths (fp16 uniformity assert — no-silent-fallbacks). - **GUT #18 VERDICT (2026-07-15): SMEARING CONFIRMED, MECHANISM = POSITION.** Entropy curves: gen-9b at tiny graphs **0.003** (four generations of small-graph rehearsal = razor calibration) vs gen-13 **0.212** (70x) — converging at trained sizes where gen-13 is SHARPER (0.558 vs 0.648). Fresh heads' shallow spots ARE temperature-calibration bands: the diet's mass distribution sets where the pointers are sharp. THE SPLIT'S SURPRISE: text-ballast (0.042) beats var-ballast (0.154) — **POSITION wins**: tiny dialects fail because query/factor tokens sit in RoPE neighborhoods training never used for those roles; inert prose pushes them home. THE FIFTH REMEDY REFINES to its cheapest form: pad with TEXT, not graph mass (annotation rulebook updated). Gen-14's displacement watch gains the small-n entropy curve as a direct temperature dial. The pointer law's remedy family,再 one law deeper: binding enters as structure, and STRUCTURE INCLUDES POSITION. Eighteen for eighteen — the drawer held the thermometer of the one attention system we built ourselves. - **THE REMEDY LAW + GEN-14'S PRE-READ (2026-07-15, relay):** the pointer-law remedy family reaches terminal form — masked attention -> span supervision -> a comma -> alphabetical discipline -> WHITESPACE: five fixes, six orders of magnitude of cost, one law, the cheapest newest. **The employment law for remedies: they too get smaller on contact with mechanism.** GEN-14 PRE-REGISTERED (pinned before the run): its diet naturally exercises a wider positional band (three books of variable-length prose pairs) — expectation: gen-14's small-n entropy curve sits BETWEEN gen-13's (0.212) and gen-9b's (0.003) — the books accidentally paying the calibration debt the clean ancestry incurred. Prints -> the books-scale claim gains a third mechanism (PROSE AS POSITIONAL REHEARSAL); doesn't -> ballast stands at zero cost. GEN-14 = the freeze's last experiment: full hot flat retrain from clean ancestry, all three books (~188 uniques) in corpus, temperature dial armed on the displacement watch, three slope sentences staged. Awaits the word. - **GUT #19: THE CONDUCTIVITY AUDIT (2026-07-15/16, Bryce + relay).** The residual stream is the defect-free lattice BY DESIGN (additive gradient flow + deep supervision = current injected at every floor); the impurities are the MULTIPLICATIVE elements — gates, LayerNorms, saturated softmaxes. THREE READS: (1) gate saturation + LN-gain longitudinal (free, tonight); (2) **THE MEISSNER PROBE — the pointer law's WHY, staged**: a saturated softmax EXPELS gradient from non-selected keys (the error becomes superconducting in the wrong channel, insulated against correction); conditioning routes through the saturation and is expelled; span supervision drills through and injects at the key — the mechanism-level account of six sightings of 'pointers move only by structure or supervision.' Probe: gradient magnitude at the correct key via CE-through-saturated-softmax vs direct span supervision, on a banked wrong-pointer specimen; predicted orders-of-magnitude gap. (3) THE GRADIENT LOGGER rides gen-14's train loop (per-module norm mean+variance every 500 steps — the project's first dissipation map; customers: the [99]/[96] upper-band question, fdiv's consolidation lag, the early transient). Summary for the ledger: the lattice was built superconducting on purpose; the wall that defined a month may have been a MEISSNER PHASE — errors so cold they expelled every field aimed at them. - **CONDUCTIVITY READS 1a/1b (2026-07-16): EMPTY BY ARCHITECTURE — the strongest possible print.** The deployed head has NO gates and NO LayerNorms (the breathing block never shipped in the 35-key production head; all LNs are frozen-trunk = fixed impedance, immune by the mouth's own construction). The head is a pure additive/bilinear circuit — the lattice is cleaner than the metaphor feared, and the ENTIRE resistor budget concentrates on the SOFTMAX family (pointers + CE heads). Gut #19 sharpens to one suspect: the MEISSNER PROBE is now the audit's whole remaining body, riding gen-14 with the gradient logger. If saturation- expulsion prints, the pointer law's six sightings get one mechanism and the month's defining wall was a Meissner phase. - **GUT #20: THE MIRROR AUDIT (2026-07-16, Bryce + relay).** THE SYMMETRY FILE, formally named (five sightings, one law): [85]'s identity palindrome (args=[a,a]), Vieta's symmetric root pairs, ill-defined selectors self-gating, [22]'s sum/difference twins, and the mixed-vote twin-key signature — **binding requires distinguishability; every symmetric structure must either break the symmetry or grade as a multiset — never bind through it.** THREE READS: (1) EFFECTIVE-VIEW-COUNT on banked certificates (fires now): sentence permutation manufactures dart independence from permutable ASYMMETRY — symmetric problems collapse five views toward fewer effective darts, and unanimity gets easier exactly where evidence is thinnest; the quarter-percent bound gains a per-item effective-K clause before the tables freeze. (Solver side immune: the uniqueness gate can't bank interchangeable variables — the wave can't stand where the gate won't let it form.) (2) TWIN-KEY POINTER ENTROPY (rides the KV machinery): mirrored mention pairs make near-identical keys; fresh heads may hold thinner twin-key margins — minted symmetric specimens vs matched controls, gen-13 vs gen-9b. (3) 1001 itself is out of band (cap 300) — the literal palindrome is unrepresentable by domain; it enters only as the digit heads' twin-key stress case if the cap conversation ever reopens. - **MIRROR AUDIT, READ 1 (2026-07-16): the standing wave is real, measured, and SMALL.** Effective-K census on bigtest under gen-13's votes: 1,477/1,500 items at full effK=5; **23 items (1.5%) at effK=3-4 — all unanimous-CORRECT, zero unanimous-wrong.** The certification table gains its per-item effective-K column as an honest fine-print clause; the quarter-percent arithmetic stands (reduced-dart certificates exist but none misfired). Found by the gut days before the freeze instead of by a reviewer after — the audit-that-confirms, with a clause as its fee. Read 2 (twin-key entropy) rides the gen-14 window with the KV machinery. - **GEN-14 FIRED (2026-07-16, Bryce's word): the freeze's last experiment.** The proven recipe at full strength: hot flat 32k from gen-7b clean ancestry on mixed13 (all THREE books x10 in corpus, ~188 prose uniques, states banked). Bars = gen-13's, unchanged. THE CONVOY READS AT CLOSE: battery + manifest-writing verdict; the ENTROPY PRE-READ (small-n curve between gen-13's 0.212 and gen-9b's 0.003 = prose as positional rehearsal, the books-scale claim's third mechanism); the DISJOINT CENSUS as the THIRD SLOPE POINT at the recipe that works (pre: gen-13 carried 28/knotted 58); Meissner + twin-key next session. Gradient logger DEFERRED (the JIT consumes grads; surgery not worth blocking the run). Three slope sentences staged; the tag waits on the verdict. - **GEN-14 PROMOTED (2026-07-16): the freeze's last experiment returns with everything.** ALL BARS: alg4test **388** (new record — the debt stays paid from clean ancestry), bigtest 1195 (ties), acceptance 8/8, cert-v2 **1.0000 at 912**, every kind over its bar. Manifest written; GATE = GEN-14. **THE ENTROPY PRE-READ PRINTS: nv0-3 = 0.010** — from gen-13's 0.212 to nearly gen-9b's 0.003 razor: **PROSE AS POSITIONAL REHEARSAL CONFIRMED** — the books paid the calibration debt the clean ancestry incurred, the books-scale claim's third mechanism, pinned before the run and printed by it. Large-n sharpest ever (0.455). **THE THIRD SLOPE POINT: SATURATION** — disjoint carried 25/knotted 61 vs gen-13's 28/58: book 3's 74 additional uniques moved the frontier ~0 (−3, noise). THE COMPLETE CURVE: +23 knots for books 1+2 (114 uniques, gen-13), −10 for stacked continuation (regime artifact), ~0 for book 3 at the same distribution (gen-14). **THE CLAIM'S FINAL FORM: annotation moves the frontier until the reachable register saturates — 81->58 for ~114 uniques, marginal yield ~0 thereafter at fixed problem-distribution; the remaining 58 are counted, family-sorted, and priced (the registry taxonomy).** A saturating curve with its mechanism triple-confirmed (corpus-consumption, positional rehearsal, the continuation contrast) is a STRONGER paper than an open linear slope: the campaign measured its own completion. **THE PUBLICATION BAR'S CONDITIONS ARE MET: tag, freeze, publish.** - **THE FREEZE'S FINAL GRADING (2026-07-16, relay):** the entropy registration graded — 'between the lineages' predicted, 0.010 printed: wrong in the happy direction; 188 naturally-varying prose uniques were nearly COMPLETE calibration payment. The books didn't just teach reading — THEY RE-TEMPERED THE POINTERS. **THREE MECHANISMS UNDER ONE CLAIM: prose as register (the mouth's needle), prose as regularizer (the gift point), prose as positional rehearsal (the temperature dial).** No single-mechanism story survives the table. **THE SATURATION SENTENCE, both halves for §8:** +23-then-~0 measures the completion of THIS DISTRIBUTION'S teachable content — what saturation does NOT claim is that books are done: harder strata, new registers (AMC prose), post-tranche-3 relation coverage are DIFFERENT distributions — book 4's charter, paper two's territory, the examiner rotation already scheduled. The curve saturated; the library didn't close. DRAFTING BEGINS: §11 first (honesty is the spine), twelve figures from banked measurements, the contributions section plain: two channels, one ledger, twenty-for-twenty, every claim gated by machinery that couldn't be flattered. - **S11 EDITORIAL PASS (2026-07-15, relay critique -> ledger-checked fixes):** six edits applied to paper/draft/s11_honest_limitations.md. (1) Denominator clash fixed: the 75-member value-range family is HARVEST-WIDE (75 of 1,743, the answer-domain census) and now stated separately from the 58-item fixture residue. (2) The 2% foreign-prose figure rephrased to name its instrument (anchor answer-accuracy that motivated the mouth), 'banked' removed. (3) Vintage pinned: 1195/1500 verified as gen-14's own battery row (ties gen-13's record); 'at freeze' made explicit; the 58-vs-61 census reads disclosed as within vote noise. (4) Chained-fdiv sentence updated to post-book-3 truth: founding specimen [66] resolved by mul-inverse rewrite, [344]'s triple-fdiv is the single surviving refusal, mechanism-vs-notation question stated open (the demotion watch, honestly carried). (5) Dialect swept from S11: darts->views, banked->verified phrasing. (6) Two limitations ADDED: every generation comparison is an n=1 training run (no seed variance, mitigations named); the annotator is the system's author (answer key verifies correctness not representativeness; 'gold' = author-written + answer-verified). S11 now 10 paragraphs; the closing line unchanged. - **S7 DRAFTED (2026-07-15): the headline artifact, four movements per the relay's brief.** paper/draft/s07_certification_lattice.md (~1,250 words + figure block). 7.1 the lattice as decision structure (four rungs zero-parameter/gold-free; chain of custody as four invariances register/rendering/lineage/truth; the epigraph carried by the entropy quadruple 0.000/0.846/0.212/0.116 at n=36, scope stated). 7.2 dials at freeze (1195/1500 one-shot; cert-v2 912 at 1.0000; the frontier's trajectory 0.9982@38.1% -> 1.0000@60.8%; zero-numerator language CROSS-REFERENCED to S11 not repeated; the 570R/1W broken certificate stated as the channel's own counterexample). 7.3 specimens load-bearing ([71] 5/5 unanimous-wrong -> mouth's mandate; panel wild dissent 9/10 and 16/19 -> second wall; [78] 3/5 stable-wrong -> answer-vs-certify dial split). 7.4 instrument aging by mechanism (rotation finding 0.59->0.988 Procrustes, 'nobody told it the sky had turned', monitor exonerated; the rotation law in full with the monotone- decline prediction RE-REGISTERED in the paper's own text; held-out- examiner portfolio discipline). Figure block pinned in-file: F-7a frontier, F-7b entropy basins, F-7c chain-of-custody diagram (drawn; candidate Figure 1). ALSO: the S11 half-nit applied — hand-quota mitigation re-aimed at the machine lane's self-preference (same commit). - **THE FIGURE CONTRACT + F-7c DRAWN (2026-07-15): the paper has a face.** (1) paper/figures/figstyle.py — the style contract every figure imports: one palette (Okabe-Ito, roles named ok/kill/wild/ alt/gate), one rcParams block, and the SELF-CITING STAMP: every saved figure carries freeze tag + gen id + parser hash in a visible footer AND embedded in PDF/PNG metadata (Subject/Keywords = the manifest's full hash block + fixtures read) — a figure detached from the paper still names its evidence; retrofitting-fourteen- figures day is now unrepresentable. (2) paper/figures/ f7c_chain_of_custody.py -> out/f7c PDF+PNG — candidate Figure 1, drawn per the relay's masterstroke note: the SPECIMENS LIVE IN THE CHAIN. Five trajectories against the four gates (register/ rendering/lineage/truth): the in-register item runs the chain to CERTIFIED; [71] dies at the mouth with its ghost dashed through the vote ('would vote 5/5 — unanimous, wrong'); wild stable votes die at the panel ('dissents 9/10, the second wall'); [78] splits off at the vote to ANSWERED-NOT-CERTIFIED (0.833); the one broken certificate (570R/1W) runs everything and dies at the key. The key drawn dashed: measurement only — grades the machinery, never deploys. Stamp reads paper-1-freeze-4-g085296d. - **S7 EDITORIAL PASS + THE PLOTTED PAIR (2026-07-15): three catches applied, two figures banked.** (1) THE LOAD-BEARING CATCH: 'every rung zero-parameter' collided with rung 2's own specialist — S7.1 rewritten to the true claim: the DECISION MACHINERY is zero-parameter (vote counting, unanimity, rank-sum, distance threshold); trained components produce candidate answers, never verdicts; 'decision-path purity, not an absence of learned parts, is what the certification claims rest on.' Honest and stronger. (2) Panel independence de-overstated: 'three models of distinct training histories (one lineage, one width), per-item behavioral disagreement measured rather than independence assumed.' (3) The dangling examiner named from the ledger: at freeze the out-of-path seat belongs to THE EXTERNAL ANCHOR (designed as held-out examiner, never in any training/acceptance path); re-rendering held the seat until the vote was promoted; staged next chairs = library cross-check + paraphrase views. Style: one name per organ — 'recognition gate' formal, one 'doorman' as color, second one removed; F-7c's gate box relabeled RECOGNITION GATE to match. FIGURES BANKED under the contract: F-7a precision-coverage frontier (ladder 3/5 0.9832@51.7 / 4/5 0.9925@44.3 / 5/5 0.9982@38.1 + channel trajectory gen-9b 866 gate, 839 panel -> gen-11 862 -> freeze 912@60.8 at 1.0000; zero-numerator note on the plot) and F-7b entropy basins (0.846 shallow vs 0.212/0.116/0.000; both arrows carry the thesis — separates shallow-from-deep, cannot separate wrong-from-correct). Remaining inventory (skeleton §12): twelve figures staged, each needs its banked artifact located. - **CONTRIBUTIONS SECTION DRAFTED (2026-07-15): the paper's handshake, written as a claim registry.** paper/draft/contributions.md. Five claims in the relay's descending-takeaway order, each with evidence pointer AND its own limit inline (the S11 cross-references made load-bearing, not decorative): (1) the lattice with its minimality-by-named-specimen argument; (2) the method as artifact — atomic promotion/manifest, nine tombstones, the ledger AS supplementary material ('offered for audit, not trust'), S7.4's standing bet as live exhibit; (3) the reading campaign — the incorruptible gate, triple-confirmed mechanism, measured completion; (4) the binding theorem proved both directions with the two-jaws design as constructive consequence, reach-beyond-register honestly conjectured; (5) the instrument-aging field manual — rotation-not- decay, length-as-distance, selection-against-gates, 'the laws' forms travel, the constants do not.' AUTHOR BLOCK in three paragraphs: Bryce (direction, adjudication, twenty registered instincts all finding something real, annotation surgery, the policy itself); Claude (design channel + execution channel, checking each other); and THE MACHINERY (neither author) — 'the results belong to a discipline, not to a hand' as the paper's strongest authorship statement. Counts verified against the ledger before writing: 20 instincts (gut #20 = mirror), 9 tombstones, ~188 uniques, ~82% surgery lane. - **THE THREE CLAIM-FLAGSHIP FIGURES (2026-07-15): the registry's re-prioritization executed.** All under the contract, all self-citing. (1) F-9a SATURATION CURVE (claim 3): 81 -> 58 -> 61 vs cumulative uniques (0/114/188), the Delta+3 disclosed as vote noise ON the plot, marginal-yield~0 named, and the S11 scope drawn as a hatched unmeasured region ('the curve saturated; the library did not close') — the screenshot-without-caption attack is pre-empted in the pixels; the excluded stacked-continuation read disclosed in a footnote. (2) F-5a ROTATION-NOT-DECAY (claim 5), three panels per the relay's stronger form, computed from the REAL gen-5/gen-9b centroid npz files with the audit's exact SVD — the script ASSERTS reproduction (raw 0.593 / aligned 0.988 / residual 0.155 printed at render) and panel C shows per-kind residue honestly: sel (0.965) and rel_add (0.976) are the least-aligned kinds, a detail the mean hid. (3) F-3a BINDING THEOREM (claim 4), specimens-as-rows per the relay's design note, zero schematic spaces: [45] taxi + [7] faucet prose verbatim from book1.jsonl (same RATE frame, z=-2.05; visibly different knots: 3-var chain vs 5-var double-mul junction, no shared class at n=94) against bigtest[1187] + vtest[116] — the hash audit's ONE cross-fixture isomorph pair (digest 468be959), terse vs verbose register, drawn with identical knot layouts. Caption = the theorem. Figure count: 6 banked (F-3a, F-5a, F-7a/b/c, F-9a). - **S9 DRAFTED + THE CASCADE FIGURE (2026-07-15): the method section, with claim 2's illustrations riding inside per the relay's curatorial call (worked example over census).** paper/draft/ s09_method.md, five movements: (9.1) the protocol — bars before builds, density regimes stated, promotions mechanical, with the stale-manifest audit DISCLOSED as the rule's origin story ('there is no state of the system that exists only in prose'); the ledger as supplementary, 'long, unedited, and contains our mistakes at the same resolution as our results.' (9.2) the survivor arc as worked example — nine refutations narrated to the arc's own closing line; depth-over-breadth per the relay ('a reader shown forty kills at once admires it without believing anything in particular'). (9.3) Table 9-1: THIRTEEN LAWS with sighting counts from the ledger (metric-decision 4; pointer-never-fixed-downstream 5; density-regime 5; front-loaded-decay 4; selection-jurisdiction 3; acceptance-law 3; structural-entry 2+; prevention 2; five mechanism-grade singletons incl. the standing bet). (9.4) the method applied to itself — the discriminating-test pattern quoted. (9.5) two channels + adjudicator; twenty instincts stated as CHECKABLE ('an instinct that had to survive formal registration and a mechanical read is data; the same instinct applied directly would have been anecdote'); ends pointing at S7.4's bet. F-9b BANKED: the cascade drawn from the arc's final-ledger line — five kills (kill-red), the pivot at #6 (gate-blue, 99.6%), three pricings (13.9% / 6% retired / 3.0%@0.165), closing box with the accounting line verbatim. Figure count: 7. - **S3 DRAFTED + THE FRESH CENSUS + THREE CROSS-READ CATCHES (2026-07-15): the vehicle gets its chassis, and the census re-run catches a stale headline number.** THE CENSUS FINDING: the relay's re-run-at-the-tag instinct was right — parser gen-14 = 4,000,813 params (docs quoted ~3.2M; the multiplicity bit + reader-era growth postdated the last count); specialist = 4,004,909; frozen trunk slice (embed + L0-L3, counted from the safetensors header) = 505,954,304; jury = armB 4,000,300 + cap2x 13,767,724. S11's scale paragraph corrected to 4.0M/8.0M/~506M. CLAUDE.md still says ~3.2M (correction owed at next doc pass). THE CENSUS TABLE'S PUNCHLINE: row 2 (both jaws) EQUALS row 1 (parser+specialist) — the solver adds zero trainable parameters, 'nothing on the verification path for training pressure to corrupt'; leverage 63x frozen per trained. S3 four movements: (3.1) theorem->design derivation ('the rare architecture section that argues necessity'); (3.2) components + census; (3.3) as-built-vs-as-designed — Poincare tier and notebook replaced by hard membership and the repair signal, 'a designed OBJECT replaced by a measured ACTION. The nouns died; the verbs survived.'; (3.4) laws as constraints — pointer law five remedies at birth, the DISCOVERED DIALECT as the IR nobody designed, the two-channel spine proved load-bearing. CATCHES APPLIED: survivor- arc naming unified (contributions), five remedies (Table 9-1), channels frame unified ('two machine channels and a human adjudicator'). - **S8 DRAFTED (2026-07-15): the wound-and-cure section, framed per the relay's credibility note.** paper/draft/s08_external_anchor.md, four movements: (8.1) THE CREDIBILITY SENTENCE LEADS — the anchor is the only measurement whose INPUTS carry no author fingerprint (all other fixtures generated or author-annotated); 'the examiner rather than the exam,' verdict reported first and at full strength. (8.2) the wound as registered: 2/97, the 63 impossible certificates, flat abstention 67.5/66.1 ('the system did not know what it did not know'); the mechanism as the section's theorem — all five renderings thrown by the same arm; 'unanimity certifies reading stability, and stability coincides with truth only in-distribution'; the ledger's honest sentence quoted verbatim (distribution-bounded certification). (8.3) the funnel frame + the cure: zero-parameter gate, selection-safe by construction, AUC 1.0000, foreign refused 100.0% at 1% native false-refusal, 160/160 anchor false-certs refused — THEN the gate's own S11 entry reported as part of the result: the length warp (r -0.825 -> -0.024), wall confirmed at 100% on the straightened ruler, zero-point 0.1871, vintage asserted by the battery. (8.4) the gradient at its true faintness (0.236-0.273 vs 0.044; symbol-dense nearest — prose-style- before-vocabulary hypothesis logged open); demand census (62.2% plain-integer); close = 'recognition buys honesty now; coverage buys capability later,' anchor left seated as S7.4's examiner. Figure candidates pinned in-file: F-8a mouth separation, F-8b the length warp before/after. The S5/S6 MERGE DECISION banked from the relay: one section, 'The repair stack and its boundary,' arc outcome as boundary, S9.2 owns the narrative. - **S4 + S5/S6-MERGED DRAFTED (2026-07-15): the body closes; only the front door remains.** S4 (paper/draft/s04_corpus_discipline.md), four movements per the brief: (4.1) solution-first + gates, with the three specimens under one principle — perfect-square discriminant dissolving no-real-roots, self-gating selectors, nogive — 'three edge policies, zero new mechanisms; the generator's grammar simply cannot say the broken thing.' (4.2) teeth/bands as measured axes + the curriculum tombstone in one honest regime-tagged sentence. (4.3) grading policy audited as an instrument: 802 -> 5 lucky-unforced (0.6% luck bound) -> 797 forced; the 16.6% right-asked-wrong-graph class as a DESIGN PARAMETER (stable 16.6/17.2 across draws) — why answers grade through the solver and parse-accuracy is never conflated with answer-accuracy. (4.4) the closer: disjointness UP TO ISOMORPHISM (WL digests, the 42 found and excluded, bump gate) — 'the difference between we-deduplicated and we-know-no-knot-is-on-both- sides-of-the-wall.' S5/S6 MERGED (s05_repair_stack.md) per the banked decision, thesis stated up front ('measured to its boundary, and the boundary is a population, not a mystery'): portfolio (dense-ranker agreement 0.840 / rare-flag centroid, combo wins the tail), withhold-and-solve (26% free, zero silent-wrong), selective retransmission (field-flags beat gold localization, leakage zero, 148/627), the half-life (19.6->7.7->1.1->0), then the boundary in one paragraph with S9.2 owning the narrative — closer: 'the boundary did not end the repair story; it relocated it upstream.' FRONT-DOOR PRE-REGISTRATION BANKED (relay): abstract gets ONE number per claim (912@1.0000-zero-numerator / fourteen-generations / saturation / 2%-then-refused / 8.0M-on-506M), no number not in a drafted section's own text. - **S10 DRAFTED (2026-07-15): the campaign's room — a results section wearing a narrative, per the framing constraint.** paper/draft/ s10_reading_campaign.md, five measured beats: (1) lane economics 1/16.5/82.5 stable pool-vs-wild (4/66/330 on the 400-draw), with the gen-13 re-classification (16.5->35 repair) read as 'the lane split is not a constant of the domain but a moving readout of how much the librarian has learned'; census-never-substrate stated. (2) THE GATE DEMONSTRATED ON ITS AUTHORS — the harvest gate's first day (0/5, 'the zero was the system working') as S11's annotator-paragraph made concrete, plus the three value-domain catches ([32]/[220]/[113], the last unmasked after days as OUR error); 'a gate that cannot be flattered by its own builders is the campaign's license to call its data gold.' (3) the rulebook written by refusals, one rule per named wall; the MYSTERY HALF-LIFE quoted (every refusal -> recipe/certificate/annotator- error within one tranche, 'counted, not remembered'). (4) the triple-confirmation with its three instruments: odometer +31.1% length-controlled + census slope +8/100 (register); the 2.9%x10 record-as-side-effect vs poison-at-saturation (regularizer, dose law carries it); entropy 0.212->0.010 pinned-before-printed (positional rehearsal) — 'three effects, three instruments, no shared failure mode.' (5) completion: F-9a + the waiting room emptying ([7] chronic->plumbing as the two sentences of color) + the closer: 'every wall had become a recipe, a certificate, or plumbing... The library taught the librarian.' Door rhythm confirmed: artifact, discipline, construction, limit, confession. - **THE FRONT DOOR DRAFTED (2026-07-15): title, abstract, intro — the house is whole, asterisk pending S2.** paper/draft/s00_front_door.md. TITLE PROPOSED (Bryce's veto at venue selection): 'Certify, Answer, Flag, Abstain: A Chain of Custody for Machine-Read Mathematics' — the four words lead per the lattice-is-the-brand vote; the subtitle names the mechanism. ABSTRACT ~200 words under the pinned rules: opens on the decision claim ('output should not be an answer; it should be a decision'); one number per claim from the pinned set (912@1.0000-zero-numerator / 8.0M-on-506M / 2%-then-refused-100% / saturation / fourteen generations); the boundary stated PROUDLY in-abstract; closes on ledger-as-supplementary + the standing bet. Every abstract number verified present in a drafted section. INTRO in the five-movement rhythm (artifact, discipline, construction, limit, confession): thesis lands in paragraph one ('a depth gauge called a compass'); Fig. 1 cited in paragraph two; the ledger sentence quoted whole; the campaign in three sentences with the 0/5 incident; the limit at full strength with S11 recommended as the skeptical reader's first stop; reader's map closes. Assembly notes pinned in-file (renumbering, S2's literature list). REMAINING before assembly: S2 related work; the S8 figure pair (F-8a/b); optional inventory figures; tag-check pass. - **FIVE CROSS-READ CATCHES + S2 DRAFTED (2026-07-15): the last room.** CATCHES: (1) S7.2 campaign pointer S9->S10. (2) The designed- logical-form tombstone re-homed: one sentence added to S10's rulebook beat (top-down attempt died in a registered kill; dialect written bottom-up) and S3.4's citation repointed S6->S10. (3) NUMBERING DECIDED: renumber at assembly (s07->6 etc.), recorded in s00's assembly notes with the reader's-map/cross-ref/figure-name update riding the tag-check pass. (4) 'Figure N' placeholder on the checklist. (5) THE SUBSTANTIVE ONE, verified against the manifest not memory: ALG_FTYPES=6 IS current at freeze — the fix was the missing relationship sentence, added to S3.2: 'the six factor types are the parse-side surface, not the relation inventory: registry relations enter as solver-side predicates bridged onto these types — which is how double-digit relation kinds ride on a six-way head output.' S2 (s02_related_work.md) against the delivered brief, seven paragraphs: selective prediction (zero-parameter decision machinery + boundary-as-headline vs threat-to-validity); calibration ('it calibrates the wrong axis for certification'); SELF-CONSISTENCY HEAD-ON (three load-bearing differences: deterministic input re-renderings not samples; unanimity as certification tier; 'S8 is the out-of-distribution invoice'); CONFORMAL its own paragraph (the mouth as an explicit exchangeability check; complementary, honestly credited as offering MORE than our bound under its precondition); propose/dispose (trained verifiers 'move the corruptible component rather than removing it'); ARQ ('our contribution is the boundary measurement, not the loop'); AGING-NOVELTY CLAIMED CAREFULLY ('the ingredients are old; the articulation as a deployment law with a succession plan and a standing bet is, to our knowledge, new'). [cite] slots resolve at assembly. THE HOUSE IS WHOLE; remaining = F-8a/b + the assembly/tag-check pass. - **F-8a/b + THE ASSEMBLY PASS (2026-07-15): the paper stands as one document.** FIGURES: F-8b (length warp) reproduces the recal's native fit FROM BANKED ARTIFACTS ONLY (dag8test pooled states + gen-9b bank, numpy kNN, zero GPU) and ASSERTS the ledger's numbers at render — printed raw r=-0.825, residual r=-0.024, thr=0.0072 to the digit; two panels, warped and straightened, 'is this distance, or is this n?' on the plot. F-8a (register map): one ruler, three real populations — native fixture histogram (banked states, n=700), the census pool's 100 PER-ITEM banked dots, MATH-500's banked band — with the wall and the 160/160 refusal annotated; raw vintage stated on the axis, corrected reads pointed at F-8b. S8 prose now cites both. ASSEMBLY (paper/assemble.py -> paper1_assembled.md, 1,130 lines): 11 sections in final order; RENUMBERED per the decision (7->6, 8->7, 9->8, 10->9, 11->10) on headers and every paragraph ref; figures mapped to sequential Figure 1-9 (chain=1 as branded); Table 1 census / Table 2 laws; working-note comments stripped; assembler ASSERTS no unmapped ref survives. One seam caught in the read: S3.1's '(S4 gives both proofs)' promised what S4 doesn't deliver — softened to the isomorph-audit pointer + the ledger. TAG-CHECK: 57 headline tokens swept; 14 flags, ALL placement-only (figure-vs-prose), ZERO not-in-ledger — every number in the document traces. REMAINING (the relay's highest-stakes item): [cite] verification by web lookup, minimal bibliography, each reference read before cited; then the fresh top-to-bottom stranger read, both channels; then theshapeofthought.ai. - **THE CITATION PASS (2026-07-15): 22/22 VERIFIED, ZERO CUTS — the last wall built to code.** Four parallel verification agents ran the pinned twenty against actual sources (title/authors/venue/year/ identifier confirmed by lookup, abstract read before any characterization); two additions (Guo et al. 2017 temperature scaling; Kadavath et al. 2022 self-evaluated confidence) were caught OUTSIDE the pinned list by the integration sweep and verified inline — the protocol admitted no exceptions, including its own list. THE PASS'S CATCHES, each a trap avoided: (1) MATH-500 PROVENANCE — the 500-subset originates in Lightman et al. 2023 ('Let's Verify Step by Step'); the NAME is post-hoc (HF dataset card); S7's anchor now cites Hendrycks et al. 2021 for the dataset + Lightman for the subset. (2) WL-INDISTINGUISHABILITY — 1-WL is complete for almost-all, not all, graphs; S4's disjointness claim STRENGTHENED: digest-equality is coarser than isomorphism, so the exclusion is conservative — removes at least every true isomorph. (3) GOODHART'S ACTUAL WORDS — the famous 'measure becomes a target' is Strathern's paraphrase; S2 now quotes the 1975 original ('any observed statistical regularity will tend to collapse...'). (4) Autoformalization softened (formal checking is downstream of translation). (5) Shanmugam's naive-averaging-imperfect thesis cited AS SUPPORT for unanimity-not-averaging. (6) Chow double-cited (1957 for age, 1970 for the rule); rehearsal attributed to the review literature, not McCloskey-Cohen's abstract; Zhang survey corrected to TPAMI 2020. ARTIFACTS: paper/bibliography.md (29 cited entries each with a cited-for note — the citation tag-check — plus 3 verified-in-reserve: Baars, Dehaene-Changeux, Papyan); s02 fully keyed (assembler asserts no [cite] survives); body citations inserted (Kschischang S3, WL pair S4, MATH provenance S7, preregistration pair S8, rehearsal pair S9); References section appended at assembly (1,186 lines). REMAINING: the stranger read, both channels, then the door opens. - **THE STRANGER READ PASSES + FINAL SWEEP (2026-07-15): THE PAPER IS DONE.** Both channels read the assembled document fresh; verdict = the house passes (renumbering resolved everywhere; S2 characterizations match the bibliography's cited-for notes; the workflow accounting consistent across S8.5/contributions; the references' [Editorial.]/[Unrefereed.] honesty tags in-voice). THE BLOCKER CUT: the title block's working note (governance text under the published title) removed at source. THE PARENTHETICAL SWEEP, per-instance and deliberate: S2's process note PROMOTED to prose in the paper's voice ('the same use-matches-source standard the rest of the paper applies to its own numbers'); S10-limitations' 'drafted first' note KEPT (the intro references it; self-aware house style); contributions' 'claim registry' note KEPT (same). FIGURE ORDERING decision recorded: forward references from the intro (Figure 9 cited before Figures 3-8 appear) are ordinary practice and stay. FINAL STATE: paper1_assembled.md, 1,184 lines, eleven sections + References (29 verified entries), ten figures under the self-citing contract, two tables, every number traceable, every citation verified, every catch closed. NEXT (Bryce's hands): theshapeofthought.ai under the declared byline; ledger as supplementary; repo public at the tag; reproduce-tables armor. - **THE PDF (2026-07-15): paper1.pdf rendered — the paper has a physical form.** assemble.py now EMBEDS all nine figures at their anchor paragraphs with written captions (whitespace-normalized anchor matching; assert guards the count); render_pdf.py (markdown -> HTML -> weasyprint, A4, DejaVu Serif, page numbers) carries the byline per the declared policy: 'Bryce Roche · Claude (Anthropic)' with the freeze tag under the date. Title question OPEN (Bryce brainstorming poetic options: guided-by-primes / the-shape-of-thought / the-shadow-of-intelligence); the renderer takes the title from the assembled doc's first line, so re-titling is a one-line change + re-render. - **THE TITLE RULING + THE GLITCH INSPECTION (2026-07-15): four words, one title, everywhere.** THE COLLISION CHECK CONFIRMED the relay's memory: 'The Shape of Thought: How Mental Adaptations Evolve' (H. Clark Barrett, Oxford University Press, 2015) — a cognition book, exactly the adjacent field; adopting the name would cost discoverability and invite the derivative jab. RULING per the relay's cast vote: the paper is 'Certify, Answer, Flag, Abstain: A Chain of Custody for Machine-Read Mathematics' in every voice; brand unity achieved by HOSTING (title = contribution, address = brand); 'Guided by Primes' banked for Paper II; 'The Shadow of Intelligence' to the essay, owned in prose. THE GLITCH INSPECTION: neither reported artifact reproduces in the rendered PDF — page 9's rung list carries its numerals and the References their bullets (verified visually, pages 9/24/25) — the orphaned-marker symptoms match text-extraction artifacts on the reading side, not the typesetting. THE INSPECTION'S REAL CATCH: page 9 read 'cross-model panel' where Figure 1 and S6.2 say cross-lineage — the one-name-per-organ rule applied, both occurrences unified, re-rendered. paper1.pdf, 25 pages, is the publication artifact.