docs/exploratory/proposed-status-corrections.md

Proposed status corrections — a proposal only, nothing applied

Status: proposal. No file outside docs/exploratory/ was modified to produce this document, and none of the edits below has been applied. Written: 2026-08-21T00:38Z.

docs/research/status.md and docs/research/experiment-index.md were last substantively written before this session's audits and findings existed. status.md carries a mtime of 2026-08-20 00:35 local, which is before the first exploratory measurement in this checkout. Seven of its statements are now either contradicted, superseded, or derivable rather than empirical.

This document writes each proposed edit out in full so that a coordinator can apply, reject, or defer it as a deliberate act. Per AGENTS.md, only the coordinator edits shared status tables, and only after merging evidence.

How to read the tiers

Evidence tiers use the vocabulary of methodology.md and benchmarks.md: proposal/mechanics, pilot, development, independently replicated development, protected validation, final confirmation. Two of the proposals below rest on recomputation of the existing record rather than on any new cohort; those are marked record audit, and they open no seeds.

Each proposal ends with one of two verdicts:

SAFE NOW — the evidence is arithmetic, derivational, or a recomputation of the retained record, or the edit only adds a correctly labelled datapoint and weakens a claim. Applying it cannot make the record less true.
NEEDS REPLICATION — the evidence is a single development cohort, or the edit would change which policy the repository calls its reference. Applying it now would promote a 64-game result into a position the benchmark contract reserves for replicated evidence.

Summary: four SAFE NOW, three NEEDS REPLICATION.

P1 — The 64-game fair-D4 reference mean

Verdict: SAFE NOW.

File: docs/research/status.md, lines 7–9 and the evidence-snapshot row at line 22. The same figure appears at experiment-index.md line 41 and in research/benchmarks/baselines-v1.json.

Current wording (status.md:6–10):

Corrected-score fair depth-4 expectimax is the strongest dependable reference found so far. Its broader 64-game recorded mean is about 308,296 points, far below the required one-million-point average.

Current wording (status.md:22):

| Fair D4: 308,295.578 points and 90.031 moves over 64 games | Detailed ledger | Recorded broader reference cohort |

Proposed wording (status.md:6–10):

Corrected-score fair depth-4 expectimax is the strongest dependable reference found so far. Its broader 64-game ledger mean is about 308,296 points, and an independent 64-game reproduction on fresh exploratory seeds measured 321,992 points at 94.06 moves. Both are far below the required one-million-point average.

Proposed wording (status.md:22) — two rows replacing one:

| Fair D4: 308,295.578 points and 90.031 moves over 64 games | Ledger-recorded as an internal bootstrap comparator inside the regenerative-expert-iteration run (history.md:4234); the 64 seeds, dispersion, censoring, and flow statistics required by methodology.md were not retained | Provisional reference mean pending a re-run under the benchmark contract |

| Fair D4 reproduced: 321,991.7 points and 94.06 moves over 64 games, sd 187,502, 0 censored, 0 score-identity violations | Reproduced in this checkout on fresh seeds 0xa51d0000–0xa51d003f | Development tier; single cohort |

Evidence: finding-01-score-is-survival.md (measured cohort table; unmodified frozen fair-D4 source, 2,000-move cap, run runs/RUN-A51D-d4/) and audit-03-claim-arithmetic.md proposal D2 (the provenance and missing-statistics annotation).

Evidence tier: development for the 321,992 reproduction, single 64-game cohort. Record audit for the annotation of the 308,295.578 figure.

Why this is safe. The edit does not replace the ledger number, does not claim the ledger number is wrong, and does not raise any claim about fair D4's strength. It adds a second measurement with its own label and cohort, and it attaches to the ledger number the provenance caveat that audit-03 established by reading the ledger. Every change is in the direction of weaker, better-labelled claims. The two numbers are consistent: finding-01 says so explicitly, and the gap is well inside the cohort's own standard error of roughly ±46,000 at one standard error given sd = 187,502 over 64 games.

Limitation to carry into the edit: the 321,992 figure is one 64-game cohort on one lease. It is not a replication of the historical cohort — those seeds were never recorded — and it must not be described as one.

P2a — Record the chance-stratum result as a development finding

Verdict: SAFE NOW.

File: docs/research/status.md, the "Most useful conclusions so far" list, under the existing bullet "Fair chance handling matters."

Current wording (status.md, conclusions list):

Fair chance handling matters. Optimistic, worst-case, or tiny reused reveal samples can rank moves incorrectly.

Proposed wording — append to that bullet:

Quantified in this checkout: the frozen reference draws five stratified samples at a chance node while the next visible disc is uniform on seven values, so on average 2.41 of the seven values receive zero weight at every node, deterministically. A candidate differing from the frozen reference in nothing but the chance-sample count (five to seven, which makes the next-disc expectation exact) measured a mean of 398,498 against 297,327 on a previously unread 64-game cohort — a paired delta of +101,171 points and +27.50 moves with a one-sided 95% lower bound of +47,457, 41–0–23 by game. The headline arm is replicated on a second, earlier-read 64-game cohort (+71,138, lower bound +5,826). Median, lower quartile and minimum all improve, so the gain is not one tail game. The change costs 3.82× the logical work per move and does not fit the frozen 3,200,000 work bound, so it is a new candidate under the benchmark contract, not a tuning of the reference.

Evidence: finding-05-chance-strata.md, runs runs/RUN-A51D-risk/ and runs/RUN-A51D-s7confirm/; the underlying bias measurement is audit-02-fair-d4.md H1.

Evidence tier: development, with the headline arm replicated across two 64-game cohorts of which one was previously unread. Under methodology.md the earlier-read cohort cannot be confirmation evidence, so this is "development, reproduced once on fresh seeds" and not yet independently replicated development.

Why this is safe. It records a measured, well-labelled development result in a bullet that already says chance handling matters. It changes no status label, promotes no policy, and states its own work cost, which is the fact most likely to be lost if it is summarised later.

P2b — Changing which policy the repository calls its reference

Verdict: NEEDS REPLICATION. Do not apply yet.

File: docs/research/status.md line 7, and every downstream comparator definition including research/benchmarks/baselines-v1.json.

Current wording:

Corrected-score fair depth-4 expectimax is the strongest dependable reference found so far.

Wording that the evidence is trending toward but does not yet support:

Corrected-score fair depth-4 expectimax at seven chance strata is the strongest dependable reference found so far.

Why not yet. Three separate reasons, any one of which is sufficient:

One fresh cohort. 64 paired games with a score standard deviation of 38–64% of the mean. finding-05's own limitations section says every row of its interaction table other than the headline arm rests on a single cohort.
It is a different work budget, not a different policy setting. The comparison is fixed-depth, not fixed-work. benchmarks.md requires that a fixed-work strength view and a resource-performance view be published separately, and finding-05 shows depth 3 at seven strata matches depth 4 at five strata at one-eighth the work. Which configuration is "the reference" depends on which view is authoritative, and that has not been decided.
The comparator is pinned by hash. baselines-v1.json pins the current comparator and its transitive source hashes. Changing the reference means issuing a new baselines manifest and re-pinning, which is a coordinator action with its own review, not a documentation edit.

What would make it safe: a second fresh 64-game paired cohort under the benchmark contract with a machine profile attached, a declared work bound, and both the fixed-work and the resource-frontier views published. finding-05's Reproduce block is exact enough to run this directly.

P3 — The 2.4 / 1.4 flow targets are derivable, not empirical

Verdict: SAFE NOW. This is the strongest and cheapest correction in this document.

File: docs/research/status.md, lines 97–100.

Current wording:

Flow matters before spectacular chains. The task record repeatedly used roughly 2.4 numbered clears and 1.4 reveals per move as the region associated with stable long games. Treat these as diagnostic targets from limited runs, not proven universal thresholds.

Proposed wording:

Flow matters before spectacular chains. The 2.400 numbered clears and 1.400 covered reveals per move that the task record used as a diagnostic region are not empirical and not approximate. They are 12/5 and 7/5: a five-move cycle inserts 5 placed discs plus 7 risen covered discs onto a 49-cell board, so any policy that survives indefinitely must clear at least 12 numbered discs and reveal at least 7 covered discs every 5 moves. They are a necessary condition for unbounded survival, derived from the rise cadence and board size, not a sufficient one — a policy can meet them locally and still die by stacking a single column into the top row. Fair D4 sustains 1.973 clears and 1.090 reveals per move, a structural deficit of 18% and 22%.

Evidence: finding-01-score-is-survival.md, "Lifetime is a conservation law" and the flow-quartile table; the rules being counted are stated in methodology.md §"Game and score model".

Evidence tier: derivation from the frozen rules. The arithmetic follows from the 7-by-7 board and the five-move rise cadence alone; it does not depend on any cohort. finding-01's 64-game cohort confirms the engine obeys it and supplies fair D4's measured rates.

Why this is safe. The current wording is a hedge against a number that needs no hedge. Replacing "diagnostic targets from limited runs" with the derivation strictly increases the document's accuracy, and the replacement introduces its own, correct, hedge — necessary but not sufficient — where the old one was misplaced. No cohort is being promoted.

Optional companion, also safe: the status quo does not say whether the requirement is reachable. finding-06-flow-ceiling.md measured that a privileged clairvoyant receding-horizon planner sustains 2.3875 clears and 1.3963 reveals per move over 6,000 moves and holds occupancy flat, so the requirement is not an arithmetic impossibility; and finding-07-fair-planning-ceiling.md measured that no legal policy tested reached it — all 87 fair-planner games ended and every fair arm's board filled monotonically, with the best legal arm at 2.2309 clears per move, 93.0% of the requirement. If that companion sentence is added it must carry both halves; the first without the second is the exact overstatement finding-07 was written to retract.

P4a — The score decomposition

Verdict: SAFE NOW, as a labelled measurement.

File: docs/research/status.md, section "### 2. Hand-written policies and shallow lookahead" or the conclusions list.

Current wording (status.md:41–44):

They showed that immediate score alone is a poor guide: a policy must keep revealing covered numbers and preserve future chain structure.

Proposed addition (a new conclusions bullet, leaving the narrative intact):

Hardcore score is survival time, measured. Across 64 fair-D4 games in this checkout, 94.29% of all points came from the flat 17,000-point row-rise bonus, 5.71% from chain waves, and 0.00% from board clears — there were none in 64 games. Final score correlated with moves survived at r = 0.9995. The engine's score identity score = 17,000·rises + 70,000·boardClears + Σ popperCount·floor(7·depth^2.5) was asserted on every game and held 64/64. A five-deep seven-disc wave — a spectacular play — is worth 2,737 points, or 0.16 of one row rise. This is a single 64-game development cohort under this repository's simulator semantics; it does not by itself establish that chain-seeking play cannot change the economics.

Evidence: finding-01-score-is-survival.md, score-decomposition table and correlation figures, run runs/RUN-A51D-d4/.

Evidence tier: development, single 64-game cohort. The decomposition percentages are exact given those games — the identity is checked per game — but their generality across policies is not established by one cohort.

Why this is safe. It adds a measurement with its cohort, its size and its scope limit stated in the same sentence, and it explicitly declines the stronger inference.

P4b — Rewriting the chains-versus-survival framing

Verdict: NEEDS REPLICATION. Do not apply yet.

Wording that the evidence is trending toward but does not yet support:

Chain construction is not a lever on Hardcore score. Survival is the only objective, and the three jobs the documentation describes — stay alive, open the board, prepare chains — are one job with three names.

Why not yet. finding-01 measured fair D4's score composition on one cohort. A policy that deliberately engineers chains would have a different composition by construction, and nothing in this repository has measured one at whole-game scale. The nearest evidence cuts both ways: finding-06-flow-ceiling.md found that a clairvoyant planner maximising score over the same windows earns 1.75× the points per move and reaches wave depth 22, yet loses on the mean because it cuts lifetime by 58% (1,123,130 vs 1,544,461 points; 165.0 vs 396.9 moves). That is a real and directly relevant observation, but it is 6–8 privileged games inside the scenario engine, which is pilot-tier evidence, and it says nothing about a legal chain-seeking policy.

Additionally, the board-clear term is not safely measured as zero: see P6 below. audit-01 H2 shows this engine overpays a fifth-drop board clear by 70,000 points relative to the cited reference, and the 0.00% board-clear share observed here means only that no policy tested reached one.

What would make it safe: a whole-game comparison of a deliberately chain-seeking legal policy against fair D4 on a fresh paired cohort, with score decomposition reported for both arms.

P5 — Two historical rejections reverse sign under corrected scoring

Verdict: SAFE NOW.

File: docs/research/experiment-index.md, lines 124 and 125. A companion edit to docs/strategies.md:157–162 is proposed by audit-03 as D8.

Current wording (experiment-index.md:125, seven-stratum D4 row):

Rejected — ledger-recorded; it was score-neutral to worse, reduced flow, and cost about 3.8 times more work.

Proposed wording:

Rejected under 7,000-point Sequence scoring — ledger-recorded. Its −163.63-point mean gap reverses to +4,336 when rescored to the corrected 17,000-point bonus (identical trajectories), so the score-mean rejection does not hold under Hardcore rules. The flow regression (clears 1.8875→1.8598, reveals 1.0349→0.9923) and the 3.79× work cost are scoring-independent and remain valid grounds for not adopting it. A corrected-scoring re-test is a live open question, and an independent re-test at depth 4 in this checkout found seven strata worth +101,171 points (95% lower bound +47,457) on a fresh 64-game cohort — see finding-05-chance-strata.md.

Current wording (experiment-index.md:124, fair-leaf CEM row):

Rejected — ledger-recorded; the D3 screen and D4 heldout/tail gates failed.

Proposed wording:

Rejected under 7,000-point Sequence scoring — ledger-recorded. The D3 fresh screen's −905-point mean gap reverses to +2,845 under corrected scoring, so that screen no longer rejects the candidate; the independent D4 interaction heldout (−17,835 points, −10,938 moves, worst-quartile regression) fails under both scoring modes and is the surviving ground for rejection.

Evidence: audit-03-claim-arithmetic.md, proposals D4 and D5, derived from an exact paired-rescoring identity applied to the retained ledger rows at history.md:1287–1294 and history.md:1158–1163. The trajectories are identical between scoring modes, so the rescoring is arithmetic, not re-simulation.

Evidence tier: record audit. No cohort was opened; the audit recomputes numbers the ledger already retains. The finding-05 sentence appended to the first row is development tier and is labelled as such by its link.

Why this is safe. The proposed text does not claim either candidate is good. It corrects a scoring-mode label on a rejection and states precisely which grounds survive the correction — for the seven-stratum row, the flow regression and work cost; for the CEM row, the D4 interaction heldout. Both rows still read "Rejected". This is the narrowest possible repair of a claim that is currently stated without its scoring mode.

Do not overreach. audit-03 is explicit that a sign reversal on a rejection gate is not evidence the candidate would have passed a full protocol; it means that specific gate no longer rejects. status.md's existing line "A failed experiment rejects only the exact tested configuration" applies in both directions.

P6 — The permanently unsatisfiable frozen protocol gate

Verdict: SAFE NOW — as a superseding record. Never as an edit.

File: artifacts/protocols/optimistic-phase-ntuple/protocol.json, line 101.

Current content (line 101, inside at100000000Transitions.qualificationRequires):

"corrected D4 reproduces frozen means 176925.25 score and 116.375 moves"

The defect. 176,925.25 is the 7,000-point Sequence-scored mean of the eight-game confirmation cohort 0x3e9c0000…0x3e9c0007 (history.md:970). The same cohort's corrected 17,000-point Hardcore mean is 400,675.25 (audit-03, line 58; also status.md:21). A run under corrected scoring can never reproduce 176,925.25, so this clause of the gate is permanently false and the qualification condition it belongs to can never be satisfied. The move figure, 116.375, is scoring-independent and is correct.

Proposed action — do NOT edit the artifact. research/README.md states that a preregistered protocol is immutable and that corrections create a successor record; AGENTS.md states that frozen protocol artifacts must not be edited and that historical hashes must not be rewritten. Instead:

Register a superseding, versioned protocol — a new file alongside the original, for example artifacts/protocols/optimistic-phase-ntuple/protocol-v2.json, or a new research/experiments/EX-… record produced through researchctl new experiment and frozen with researchctl freeze so it carries a correct self-excluding protocolSha256. Its corresponding clause reads:

"corrected D4 reproduces frozen means 400675.25 score and 116.375 moves on 0x3e9c0000-0x3e9c0007 (history.md:3337)"
Have the successor name its predecessor explicitly and state that v1 is retained unaltered and is unsatisfiable as written.
Record the v1 defect in provenance.md alongside the existing note about relocated source hashes, so that anyone resuming the optimistic-phase-ntuple line finds the defect before spending compute on a gate that cannot pass.
Leave protocol.json byte-identical. Its hash is the record.

Evidence: audit-03-claim-arithmetic.md, proposal D17 and the surrounding derivation at lines 589–594; the mode classification of the source row at line 142.

Evidence tier: record audit. This is an internal-consistency defect in a frozen artifact, established by comparing two retained figures for the same cohort. No cohort was opened and no policy claim is involved.

Why this is safe. Nothing is modified. A successor record is created and the predecessor is retained, which is exactly the mechanism the repository prescribes for correcting an immutable record.

P7 — Two rise-boundary scoring divergences from the cited reference

Verdict: NEEDS REPLICATION — for the engine change. The caveat is SAFE NOW.

File: docs/research/status.md, section "### 1. Establishing a trustworthy simulator", lines 34–38, and the evidence-snapshot row at line 18.

Current wording (status.md:34–38):

The work first aligned TypeScript and native Hardcore rules, chain scoring, gray-disc reveals, row rises, and deterministic seed behavior. A scoring audit corrected the five-move mode's level award from 7,000 to 17,000.

Proposed addition (SAFE NOW) — append to that paragraph:

An independent engine-fidelity audit in this checkout found two remaining scoring divergences from the reference implementation this repository itself cites (drop7.es6), both at the rise boundary, both reproduced with executable evidence:

H1, the terminating rise. When the fifth drop of a level completes and the rise is refused because row 0 is occupied, this engine sets game-over and awards nothing. The reference adds the level bonus first, then shifts, then detects overflow afterwards — its grid has an 8th buffer row precisely so the rise always succeeds. 254 of 256 parity games (99.2%) terminate exactly this way, so every mean, lower bound and paired delta in this repository is biased down by one 17,000-point level bonus per game. The bias is one-sided and conservative: it cannot manufacture a false million-point claim.

H2, the fifth-drop board clear. This engine tests board-clear before the rise; the reference tests it after the rise has already inserted a new gray row, so a board cleared on the fifth drop is not empty when the reference looks. Measured on an empty board: this engine scores 87,007 where reference ordering scores 17,007. The divergence is +70,000, upward, and policy-controllable — one board clear in five falls on a fifth drop by position alone, and a search that can see the rise coming has a direct incentive to schedule clears there. That is the dangerous direction for a qualification claim.

Both engines agree with each other exactly; they agree on rules that differ from the cited reference. Until this is resolved, an absolute score from this engine is a score in this repository's dialect of Drop7. Within-checkout paired comparisons are not affected by H1, which is constant per game.

Proposed addition to the evidence-snapshot table (SAFE NOW) — one row:

| Engine fidelity vs. the cited reference implementation | Independent audit, executable reproductions | Two rise-boundary scoring divergences (H1 terminal level bonus, H2 fifth-drop board clear) plus an opening-position divergence; see audit-01 |

Evidence: audit-01-engine-fidelity.md, findings H1 and H2, each with cited engine.hpp / engine.ts line ranges, cited drop7.es6 line ranges, and a measured single-move reproduction. The audit's own verdict is valid / fail / proposal-mechanics.

Evidence tier: proposal/mechanics. This is a mechanics-level verification against public source; no development, protected or final cohort was opened. That is a strong tier for this kind of claim — it is a reproducible statement about code, not a statistical estimate — which is why the caveat is safe to add immediately.

What NEEDS REPLICATION, and why it is not proposed here. Changing the engine to match the reference is not proposed and should not be done as a documentation action. It would:

change every score in the repository, breaking comparability with the entire historical ledger and with baselines-v1.json's pinned comparator;
require a decision the repository has never recorded — namely whether drop7.es6 is normative for this research program or merely cited; and
under benchmarks.md, make every existing policy a new candidate, since the change alters floating-point action ranking and the selected column near rise boundaries.

The correct sequence is: record the caveat now (safe), then decide normativity, then, if the engine is changed, do it as a new versioned engine with a fresh parity gate and re-baselined comparator — never as a fix applied in place.

Not proposed, and why

Several other corrections are live in the audits but are deliberately not raised here, to keep this document to the seven items in scope:

audit-03 proposals D1, D3, D6, D7, D9–D16 — further ledger and strategies.md annotations, a seed registry, and a game-result-v1 schema change adding a required scoringMode. D16 is worth a coordinator's attention because it makes the score identity machine-checkable per game and retires a whole class of the defects audit-03 found.
audit-01 finding H3 (opening position) and H4 (a benchmarks.md:31 protocol violation).
audit-04 and audit-05 conclusions about confounded rejections and the never-ablated curriculum ordering.
finding-08-learned-leaf.md, which was published at 17:41 local while this document was being written and is therefore outside its scope. Two things in it bear directly on the proposals above and should be read before applying them. First, it strengthens P2a: its two w = 0 reference arms are the frozen policy bit-for-bit and reproduce finding-05's published confirmation cohort to every reported digit for both the five- and seven-stratum arms, with 0 score-identity violations in 26,934 decisions and 0 censored games in 256, and its measured 4,956,614 work per move independently confirms the seven-stratum arms completed depth 4 rather than silently degrading to depth 3. Second, it bears on how P2b should eventually be worded: a learned leaf and an exact chance estimator behave as substitutes, not complements — seven strata is worth +101,171 over five with the frozen leaf but only +79,347 with a learned one — so "which configuration is the reference" is entangled with "which leaf", and P2b should not be applied without that. finding-08's own gameplay arms are development tier on a doubly read cohort and support no status edit on their own.
Anything concerning the concurrent contributor's afterstate line. Its records are complete, self-consistent and passing validation, and its own row in experiment-index.md is already written; it needs nothing from this document. See reconciliation-01.md.

Application checklist for the coordinator

If P1, P2a, P3, P4a, P5 and P7's caveat are applied, AGENTS.md's definition of done requires that the following agree afterwards: the edited document, the finding or audit each edit cites, the run directory each finding cites, and a contribution record for whoever applies the edit. P6 additionally requires a new frozen protocol record and a provenance.md note, and must leave protocol.json untouched.

For a walkthrough with board animations, start at how the game works and the concepts primer; every term is defined in the glossary.

Status · Methodology · Benchmark contract · Strategy landscape · Experiment index · Full ledger · D7P protocol