Observations made while writing narrative summaries for the machine-readable records (web/content/research/*.mdx). No record was edited. These are reported so a future auditor can decide whether a successor record is owed; under AGENTS.md a frozen or promoted record is immutable, so any correction creates a successor rather than a rewrite.

Status: observations only, not a claim that any result is wrong.

1. Quantile-coverage level disagrees between criterion and observation

RS-20260820T094500Z-5c1e9a04 states the gate criterion as "80% quantile interval coverage within [0.70, 0.90]" while its observed field reports "0.631 coverage of the nominal 76.5% outer-quantile interval". The nominal level named in the criterion (80%) and in the observation (76.5%) differ. The same mismatch appears in EX-20260820-afterstate-pilot-h40-29b8588a, whose gate.passCriteria says 80% / [0.70, 0.90] while its own metrics.secondary for later iterations says 76.5%.

Effect on the reading: the check failed either way, so the pass/fail outcome is unaffected; only the stated nominal level is ambiguous. The overlay avoids quoting the coverage number for that iteration.

2. updatedAt earlier than createdAt

EX-20260820-afterstate-pilot-h40-k64-d7a9faf5 has createdAt 2026-08-20T09:50:00Z and updatedAt 2026-08-20T09:04:08Z — updated 46 minutes before it was created. Its amendment timestamp (14:30:00Z) is also later than updatedAt. EX-20260820-afterstate-pilot-h40-29b8588a shows the same amendment-after-updatedAt pattern (updated 09:01:20Z, amendments 14:30:00Z).

Effect on the reading: none on the science; it does mean timestamps cannot be used to order these records reliably.

3. "Only change" claim is narrower than the disclosure

EX-20260820-afterstate-pilot-h40-k64-d7a9faf5's hypothesis says raising the scenario count from K=8 to K=64 is "the only change" from EX-20260820-afterstate-pilot-h40-29b8588a, but its own reuseDisclosure also moves the held-out origins to a fresh seed block and folds iteration 1's held-out roots into training. The disclosure's closing sentence ("this experiment changes only the scenario count and the held-out origins") is the accurate statement; the hypothesis sentence understates it.

Effect on the reading: two variables moved between those iterations, so a difference between them cannot be attributed to K alone. The overlay says so.

4. Root counts differ between summary and gate check

RS-20260820T184500Z-63c0a8e2 evaluates "2,689 fresh held-out roots" in its summary while its corpus gate check reports "4,858,880 rows over 2,793 roots". The difference is plausibly corpus roots versus roots that were actually evaluable (for example, roots with fewer than two legal actions cannot have a top-two near-tie), but the record does not say so.

Effect on the reading: the override-gate percentages are computed against 2,689 and are internally consistent with that denominator; the overlay uses the record's own numbers and does not reconcile the two counts.

---

Part 2 — discrepancies found while writing the console's approach pages

Twelve writers read the ledger, the experiment index, the findings and the machine-readable records in order to describe each approach for a general reader. Doing that end to end surfaced places where two retained sources disagree, or where a summary document has fallen behind the record it summarises. Nothing below was fixed by editing a record; the console pages carry both readings where they conflict, and this list exists so a coordinator can decide what deserves a successor record or a status update.

Several of these were already known: docs/exploratory/proposed-status-corrections.md proposes exact wording for seven of them, and the audit-01…audit-05 notes raise others. Those are marked (already proposed).

A. Summary documents that are behind the record

A1. status.md still describes the afterstate line as unimplemented. Its evidence snapshot row reads "AFBR-40 afterstate idea | Task-record only | Proposal; no source, checkpoint, or result", and its closing paragraph says the approach "has no implementation, protocol, model, or measurements in this repository and must not appear in a list of attempted results." The repository now holds a full implementation under approaches/afterstate-learning/distributional-afterstate/, a theory record at lifecycle: assessed / assessment: not-supported-as-tested, four experiment records and four result records. status.md §7 also stops before the top-two override gate, which is the only preregistered held-out test in which a learned intervention improved on fair D4 at all (in both half-folds; it failed because the second half's margin was below the preregistered bar).

This one also propagated into the console. The concepts primer originally repeated the "no implementation" claim; it was corrected during this pass and now describes the four pilots and their outcomes.

A2. Findings 06–13 appear nowhere in status.md or the experiment index. grep -c lifetime-objective docs/research/experiment-index.md returns 0. A reader following the console's evidence links from that family reaches only docs/exploratory/. Not a contradiction, but the "current evidence boundary" is materially behind the retained record.

A3. The seven-stratum verdict is stated as settled and is not. The index and strategies.md describe more chance samples as "score-neutral to worse". audit-03 C2 rescores the ledger's own paired means to +4,336 under corrected scoring, and finding-05 measures +101,171 points / +27.5 moves on 64 fresh paired games (95% lower bound +47,457), replicated on a second cohort. (already proposed — audit-03 rows D4 and D8.)

A4. The rollout veto is listed as runtime-paused; it has since been rejected on evidence. finding-03 retested it at corrected scoring on 32 paired games (paired mean −46,510.5, 95% lower bound −91,924.6, 9–4–19, both flow rates worse), run validity valid, outcome fail, tier development. The index still says runtime-paused.

A5. The 2.4 clears / 1.4 reveals figures are labelled as measurements but are arithmetic. status.md calls them "diagnostic targets from limited runs, not proven universal thresholds". finding-01 derives them exactly — 12 discs onto 49 cells every five moves gives 12/5 = 2.400 and 7/5 = 1.400 — and says so. Both finding-01 and finding-06 raise this against status.md. The console pages present them as arithmetic and cite finding-01.

B. The repository's headline reference number

B1. The 308,295.578 / 90.031 over-64-games figure has no identified cohort. audit-03 H2: it appears exactly once in the ledger (history.md:4234) as an internal bootstrap comparator inside the regenerative expert-iteration experiment. Which 64 seeds, what dispersion, and what censoring are not recorded, so it does not meet methodology.md's own cohort-reporting requirements — yet it propagates to status.md, README.md, the experiment index, and research/benchmarks/baselines-v1.json. (already proposed.)

Note that a different 64-game fair-D4 cohort is fully retained, with per-game rows, on lease SEEDLEASE-A51D (finding-01): mean 321,991.71875, median 266,282, max 1,017,234, 0 censored. The console's heavy-tails page uses that one and states explicitly that the two must not be pooled.

C. Internal inconsistencies inside individual records

• C1. RS-20260820T094500Z-5c1e9a04 states its criterion as "80% quantile interval coverage within [0.70, 0.90]" but observes "0.631 coverage of the nominal 76.5% outer-quantile interval". The nominal level disagrees between criterion and observation; the same mismatch is in the matching experiment record. The check failed either way, so no outcome changes.
• C2. EX-20260820-afterstate-pilot-h40-k64-d7a9faf5 has updatedAt (09:04:08Z) earlier than createdAt (09:50:00Z), and amendments dated later than updatedAt. Its result references a run stamped 09:04:11Z, so createdAt looks like the wrong field. Timestamps in this family cannot be used to order records.
• C3. The same record's hypothesis says raising scenarios from K=8 to K=64 is "the only change" from the previous iteration, while its own reuseDisclosure also moves the held-out origins and folds the previous iteration's held-out roots into training. Two variables moved, so a difference between those iterations cannot be attributed to K alone.
• C4. RS-20260820T184500Z-63c0a8e2 evaluates "2,689 fresh held-out roots" in its summary while its corpus gate check reports 2,793 roots. Plausibly corpus roots versus evaluable roots, but the record never says so.
• C5. Two panel measurements of exact fair D4 are quoted interchangeably in adjacent prose: pairwise 0.6585 / regret 0.2766 (preflight) versus 66.8191% / 0.276577 (martingale-dual audit). Different frozen protocols and tie handling; they should not be quoted across records.

D. Scoring-mode and reproducibility hazards

• D1. history.md:57's archival disclaimer scopes itself to "experiments below that identify levelBonus: 7000" — a literal string that appears nowhere else in the file, so the disclaimer selects nothing. (audit-03 C3.) Consequence: several ledger sections are 7,000-point Sequence-scored with no in-file label. The console pages state the mode per source and attribute the reconstruction to audit-03 where the ledger is silent.
• D2. Sources split into those that pin the old scoring with a static_assert(kLevelBonus == 7'000) and those that do not. The unpinned 7,000-era sources (for example the n-tuple temporal-coherence and phase-conditioned programs, and the learned-guidance search sources) now build against the 17,000-point engine, so re-running them will silently disagree with their own recorded numbers.
• D3. artifacts/protocols/optimistic-phase-ntuple/protocol.json requires a corrected-scoring D4 to reproduce a 7,000-point figure (176,925.25 / 116.375), which is permanently unsatisfiable. The artifact is frozen and was not touched; a superseding protocol is the remedy. (audit-03 C1.)

E. Verdicts that do not match the code they describe

• E1. The index's direct-policy verdict describes teacher roll-ins drifting from their teacher, but that program contains no teacher, no imitation target and no roll-in mechanism — it is cross-entropy weight search against complete games. Either the verdict belongs to a different configuration or it is misfiled.
• E2. The index's denoised-value row implies score means that were never retained; only move means exist (78.75→81.25 screen, 79.5→88.125 confirmation). (audit-03 L1 already flags this.)
• E3. "CFPI" names two different things in history.md — a depth-3 comparator in the fair-only horizon screen, and the conservative fitted-policy-iteration program — and the ledger never says whether they are the same. The console pages do not attribute one's numbers to the other.
• E4. The index lists all three evolution programs as "Rejected — ledger-recorded", but the linear evolution.cpp appears in the ledger only as the source of reused features, with no result section of its own.
• E5. The PyTorch PPO index row stops at the aborted direct run and omits a completed third stage (32 iterations, 16,384 games, then a 64-game development cohort read and rejected). strategies.md does mention it.
• E6. strategies.md records the primal-dual actor-critic at "about 176,000 points and 54 moves"; the ledger records 175,834 / 55.006.

F. Engine and interface facts worth surfacing

• F1. Hidden gray values are drawn at reveal time, not stored. audit-01 M2: this repository's engine assigns a fresh value at the moment of reveal, in row-major order, while the original game fixes the value when the disc is created. To a player the two are indistinguishable; for research they are not, because there is no persistent latent board unless latent mode is enabled. The console's rules page originally stated the original game's behaviour as though it were this engine's; it was corrected during this pass.
• F2. The censoring cap is not one number. The benchmark contract says 2,000 moves; the engines default to 500, and retained runs use 500, 1,000 or 2,000. Every quoted cohort should state its own cap. (audit-01 L1.)
• F3. The gray-throughput policy is registered publicInformation: false. Its chance-sample salt hashes level and moves-played, both outside the deployable interface. It is a legitimate playground entry, not a legal candidate, and the console shows its extended-state badge.
• F4. "Throughput" means two different things. In this repository it is discs cleared and revealed per move, not games per second. A draft console page had adopted the wrong sense; it was corrected.

G. Documentation defects

• G1. docs/exploratory/finding-15-depth5-exact-estimator.md carries the H1 heading "Finding 14", which is also the number of finding-14-leaf-reweight.md. A citation to "finding 14" is ambiguous.
• G2. docs/exploratory/finding-12-fair-planner-ceiling-extended.md §6 contains an unfinished table row — a four-cell placeholder against an eight-column table — for an H = 9 clairvoyant arm. No number was quoted from it.
• G3. docs/exploratory/README.md's contents table is stale: it lists through finding-03, marks two notes "in progress", and omits findings 05–08 and the gpu-02/gpu-03 notes. (reconciliation-01 gap 4.)
• G4. approaches/lifetime-objective/flow-ceiling/README.md (the operational note, not the rendered page) still presents figures that finding-12 withdrew: 58.8% gap closure, since corrected to 27.1% on a matched fresh cohort; 2.2309 clears per move, cohort-inflated versus 2.0260 on fresh tapes; and 117.75 mean moves for fair D4 versus 93.56.
• G5. strategies.md reports the perfect-information oracle above one million points without noting that all twelve of those games were censored at the 500-move cap, so the figure is a lower bound on the oracle's own play and is not comparable with uncensored cohorts.
• G6. The implied move target for a million points is quoted two ways: about 294 moves in status.md/finding-01 arithmetic, and "roughly 285" in the ledger's evolved-public-policy entry.