Gate-1 amendment proposal 2: mean-based classifier gating

[MaxGhenis]

Gate-1 amendment proposal 2: mean-based classifier gating

This PR is an inert amendment OBJECT, exactly like PR #57. It adds a
gate_1.amendment_proposed subsection to gates.yaml and changes
nothing: every locked threshold is byte-identical (the pairs-view
c2st_auc_max: 0.53 line this proposal reinterprets is untouched), and
the harness scores only the locked block. It becomes live only if a
fresh adversarial referee round runs against it and maintainer
ratification-by-merge follows (a later flip PR, as PR #59 did for
amendment 1).

The evidence

• PR Candidate-10 diagnostics: seed noise vs microtexture #66 — the 20-seed measurement (runs/c10_diagnostics_v1.json):
  the committed candidate 10 (run 12, PR Gate-1 candidate 10: the inner-validated composition #64) scored on 20
  protocol-identical splits has pairs-view c2st mean 0.5234 ± 0.0059,
  i.e. 5.0 standard errors of the mean below the 0.53 line. Only 2 of
  the 20 seeds clip 0.53 — and both are the original gate seeds 0–1. The
  parametric probability of the observed run-12 outcome (≥2 of 5 gate
  seeds over the line) given that true mean is ≈0.13. The microtexture
  forensics read the clipping as noise: the same seed-invariant,
  sub-threshold residual drives the classifier in every seed, with no
  coherent failing-vs-passing signal (reading: "noise",
  signal_is_same_across_seeds: true).
• PR Gate-1 candidate 10: the inner-validated composition #64 — run 12 (runs/gate1_rank_knn_v4.json): candidate 10, the
  near-miss that fails the locked per-seed rule 3/5 on pairs c2st.
• PR Inner-validation harness and the candidate-10 design sweep #63 — inner validation (runs/inner_sweep_v1.json): the nested
  selection that set candidate 10's constants without outer-holdout
  contact.

The per-seed rule rejects a candidate whose true pairs-view
distinguishability sits 5 SE inside the ratified allowance, for
thousandths-of-an-AUC sampling excursions on individual splits.

The exact changes

Change 1 — pairs c2st gating moves from per-seed to the across-seed
mean, at the same ratified line. The gate condition becomes: mean of
pairs c2st over the five locked seeds ≤ 0.53. This is the same
ratified line with the same derivation (ctx20 floor mean +
4.2 × per-seed floor sd, rounded to 2 = 0.53). The line does not move; the
amendment reinterprets it as the absolute distinguishability allowance the
lock chose, now tested where its per-seed form is measurement noise.

Considered and NOT proposed: a statistically-tighter mean-of-5 line,
floor mean + 4.2 × sd/√5 = 0.5194 (exact committed-artifact value; a
rounded-input hand computation reads ~0.5196). This is declined because
moving the line from 0.53 to 0.5194 is a material tightening of the
ratified allowance that no referee round ratified — it would fail
candidate 10 (mean 0.5279 > 0.5194) and reintroduce the noise-vs-signal
ambiguity this amendment removes. The referee should weigh this choice
explicitly.

Change 2 — a per-seed catastrophe cap alongside the mean rule. No
single seed's pairs c2st may exceed floor mean + 8 × per-seed floor sd =
round(0.5108054 + 8 × 0.0045535, 3) = 0.547 (exact committed-artifact
derivation; a rounded-input hand computation reads 0.548 — the exact value
is what binds). tests/test_gates_derivations.py binds this like every
locked geometry threshold. The cap restores per-seed protection at a level
no faithful candidate reaches (candidate 8's max seed 0.5399 is under the
cap yet its mean 0.5307 fails the mean line, so the two together are
strictly more discriminating than either alone).

Change 3 — protocol clarification (no threshold change). The C2ST
classifier's scikit-learn version must match between floor derivation and
candidate scoring, and artifacts must record it. Evidence: PR #66 found a
silent cross-venv mismatch — committed floors built under sklearn 1.9.0
(.venv), gate candidates scored under 1.8.0 (.venv-gate, populace-fit's
<1.9 pin). Per-seed floor c2st shifts ~±0.002–0.006 between versions, but
the floor mean is version-stable at ~0.5109, so no committed number
changes. This clause prevents a future silent drift.

Verdict recomputation — every committed run

Re-scored under the amended pairs-c2st rule (mean ≤ 0.53 and every seed
≤ 0.547), holding all other locked thresholds, the amendment-1
benefit_space block, the battery, and the pooled Q0 gate fixed. Every
number below is recomputed from the committed artifacts by
test_amendment2_verdict_recomputation_matches_artifacts.

Run	PR	pairs mean (≤0.53)	max seed (≤0.547)	battery	pooled Q0 (|·|≤5)	old → amended
baseline	#40	0.5122 ✓	0.5196 ✓	0/5 ✗	—	FAIL → FAIL (battery)
candidate 2	#43	0.5191 ✓	0.5255 ✓	0/5 ✗	—	FAIL → FAIL (battery)
candidate 3	#44	0.5228 ✓	0.5328 ✓	0/5 ✗	—	FAIL → FAIL (battery)
candidate 5a	#50	0.6007 ✗	0.6121 ✗	0/5 ✗	—	FAIL → FAIL (mean, cap, battery)
candidate 5a′ (segment splice)	#51	0.5466 ✗	0.5510 ✗	5/5 ✓	—	FAIL → FAIL (mean, cap)
candidate 5b	#52	0.6696 ✗	0.6748 ✗	0/5 ✗	—	FAIL → FAIL (mean, cap, battery)
candidate 6	#53	0.5469 ✗	0.5556 ✗	0/5 ✗	—	FAIL → FAIL (mean, cap, battery)
candidate 7	#55	0.5231 ✓	0.5291 ✓	2/5 ✗	+9.2953 ✗	FAIL → FAIL (battery, Q0)
candidate 8	#58	0.5307 ✗	0.5399 ✓	0/5 ✗	—	FAIL → FAIL (mean, battery)
candidate 9	#62	0.5373 ✗	0.5499 ✗	2/5 ✗	−17.8853 ✗	FAIL → FAIL (mean, cap, battery, Q0)
candidate 10	#64	0.5279 ✓	0.5330 ✓	4/5 ✓	+0.0376 ✓	FAIL → PASS

Exactly one committed run's overall verdict changes: candidate 10 (PR
#64), FAIL → PASS. No other run flips — segment splicing's pairs mean
0.5466 (~0.547) > 0.53 still fails; candidate 7 still fails pooled Q0 at
+9.30 and the battery. Under the amendment candidate 10's across-seed mean
0.5279 ≤ 0.53 and every seed ≤ the 0.547 cap, so geometry passes 5/5 (its
non-c2st pairs thresholds, runs coverage, and per-seed benefit_space all
pass on all five seeds); battery 4/5; pooled Q0 +0.0376 ≤ 5.

Process statement — stated plainly

This amendment was proposed after run 12's near-miss, because the
post-run diagnostics (PR #66) proved the per-seed pairs-c2st rule
noise-dominated at thousandths-of-an-AUC margins. Changing a gating rule
after seeing a specific candidate miss it is a goalpost move on its face.
The defense is that (a) the ratified 0.53 line does not move, (b) the
change is motivated by a 20-seed / 5.0-SE measurement independent of which
candidate triggered it and applicable to any candidate, and (c) a per-seed
cap is added so protection is not merely relaxed. Whether that defense
holds — or whether the timing makes this an illegitimate goalpost move — is
exactly what the fresh adversarial referee round must adjudicate. This
proposal discloses that candidate 10, the run that motivated it, flips to
PASS, and that no other committed run does. This PR merges nothing live;
it is inert until refereed and ratified.

Tests

tests/test_gates_derivations.py gains four binding tests (all green):
the cap derivation (0.547 from the committed ctx20 floor), the mean rule
(0.53 with the ratified derivation reference), the considered-and-rejected
line (0.5194), and the full verdict recomputation (every disclosed
pairs-mean / cap / battery / Q0 number and verdict re-derived from the
committed run artifacts).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
social-security-model	Ready	Preview, Comment	Jul 7, 2026 1:54pm

[MaxGhenis]

Verdict: AMEND BEFORE RATIFYING.

The core measurement point is sound and would hold without candidate 10: a per-seed c2st gate with per-seed sd ~0.006 and a 4/5 rule is genuinely noise-sensitive, so a mean statistic is defensible for any candidate. But the two free parameters that decide candidate 10 — the line and the seed set — were both resolved to its benefit, the headline oversells the margin ~4x, and the triggering candidate self-rescues. Fixable, not fatal; not ratifiable as-is.

1. Exchangeability is violated, and the direction guts the headline (check 1). The five gate seeds are not a random draw for this candidate. Seeds 0 and 1 are the global max and 2nd-max c2st of all 20 (0.5330, 0.5315; 3rd is 0.5275); permutation P(a random 5-subset has mean ≥ the locked-5 mean 0.5279) = 143/15504 = 0.009; Welch t(locked vs fresh) = 2.54. It is candidate-specific, not split difficulty: the floor is flat across seeds (locked 0.5106 vs fresh 0.5111) while candidate excess-over-floor is 0.0173 locked vs 0.0108 fresh. Mechanism is exactly as flagged — inner validation uses inner seed = 1000 + outer seed and selected V1 λ=0.1 by aggregating over outer seeds 0–4, so 0–4 are design-touched, not fresh. Consequence: the advertised "5.0 SE below the line" is the margin of the 20-seed mean (0.5234); the gate actually scores the locked-5 mean 0.5279, which is 0.8–1.2 SE below 0.53 (0.80 SE on the 20-seed per-seed sd; 1.16 on the locked-5 sd). Pooling the 15 easier fresh seeds — drawn after this candidate — manufactures a cushion the gated statistic does not have.

2. "Same 0.53 line" hides a large operating-characteristic shift (check 2). k=4.2 was ratified as a per-seed line paired with 4/5. Reused as a mean line it inherits stringency calibrated for a different estimator. Pass probability (σ=0.0059, iid approx), per-seed(≥4/5≤0.53) → mean(≤0.53): true mean 0.525 → 74%→97%; 0.529 → 28%→65%; 0.530 → 19%→50%; 0.533 → 3%→13%. The over-rejection at 0.525 is the legitimate fix; the doubling-to-tripling of pass rates at the line is a real loosening. The rejected 0.5194 is too tight (fails a true-mean-~0.522 candidate). Honest answer is an intermediate line — or a mean over more pre-registered seeds, which sharpens the OC around 0.53 so "the number doesn't move" also means "the gate doesn't loosen."

3. The cap is decorative (check 3). Across all 11 committed runs the cap (0.547) changes zero overall verdicts: it never binds a mean-passer (max c2st of every mean-passer ≤ 0.5330) and is redundant for mean-failers. Fine as a prospective one-seed-catastrophe guard; it does no work on the evidence and should be described as such.

4. The version-pin clause contradicts its own thresholds (checks 3+4). Change 3 declares floor-derivation and candidate-scoring sklearn versions must match; the proposal then ships lines derived from the 1.9.0 floor while candidates score under 1.8.0. The mean line survives (floor mean is version-stable, 0.5109 both). The cap does not: it uses the 1.9.0 per-seed sd 0.004553, but the version-matched (1.8.0) sd is ~0.0050–0.0054, giving a cap of 0.551–0.554, not 0.547. So the claim that "the derived 0.53/0.547 lines are [version-]unaffected" is false for the sd-derived cap — and it's demonstrable: c5a' (max 0.5510) fails at 0.547 but passes at 0.554. The clause is violated at birth.

5. Disclosure: numbers verify, emphasis misleads (check 5). I recomputed from committed artifacts — c10 mean 0.5279 PASS, c8 0.5307 fails mean, splice 0.5466/max 0.5510 fails both; sole flip = candidate 10; test_amendment2_verdict_recomputation and all 19 derivation tests green. The process_statement names the goalpost timing plainly — good. Gap: nowhere does the disclosure state that the gated locked-5 mean is a p≈0.01 unfavorable draw sitting ~1 SE from the line while the headline touts 5 SE. Also note "noise, not signal" is loose — the excess over floor is real signal at 6.8 SE (0.0124 ± 0.0018); it's tolerated signal within the 0.53 allowance, and only the per-seed clipping location is noise.

6. Precedent: this one is accommodation-shaped (check 6). Two amendments in two days, each after a failure — but amendment 1 (PR #57) added strictness and candidate 7 still failed; amendment 2's triggering candidate flips to PASS. That asymmetry is the pressure-valve tell.

Required fixes.
A. No candidate may be the ratifying pass for a gate rule proposed after its own run. Require a fresh candidate — or a c10 re-run on a freshly pre-registered seed set — to clear the amended gate. (1, 6)
B. If moving to a mean, gate over a larger pre-registered seed set (≥~15–20), not the 5 that contain both global-max clippers: robust to the unfavorable draw, sharpens the OC at 0.53, still passes c10 (20-seed mean 0.5234). If staying at n=5, justify 0.53 against the per-seed rule's operating characteristic — not "the number is unchanged" — and disclose the gated statistic is ~1 SE, not 5, below the line. (1, 2)
C. Re-derive the cap (and re-confirm the mean line) from a floor built under the scoring sklearn version (1.8.0), or drop the cap and pin the version separately. As shipped the cap is off the wrong version (0.547 vs matched 0.551–0.554). (3, 4)

[MaxGhenis]

Reworked per the AMEND-BEFORE-RATIFYING verdict

Thanks — the verdict was right on every count. I've reworked the proposal on this branch (the locked thresholds block and amendment_history are byte-identical; all changes are inside amendment_proposed). Every number below is re-derived from committed artifacts and bound in tests/test_gates_derivations.py; I trusted no figure from the review or my own notes.

Required fix A — no self-rescue (findings 1, 6)

The triggering candidate no longer rescues itself. Added the verbatim clause:

No candidate's committed run verdict changes under a rule proposed after that run. A gate rule amendment applies only to runs registered after its ratification.

• Candidate 10 REMAINS FAIL. The old verdict_recomputation (which flipped it to PASS) is gone. It is replaced by illustrative_retroactive_application with applied: false — the recomputation table survives as disclosure only, every run now carries committed_verdict: FAIL (unchanged) alongside a hypothetical would_be_verdict_if_applied. The sole would-flip is still candidate 10, but it flips nothing.
• path_to_pass: a pass requires a FRESH registration — a new candidate number, even for a byte-identical spec, run one-shot on seeds 0–19. It discloses that candidate 10's 20 per-seed scores are already public (runs/c10_diagnostics_v1.json, 20-seed mean 0.5234), so re-registering that spec has a known-in-advance pairs-c2st outcome; the run still executes end-to-end and must clear every block.

Required fix B — mean over a larger pre-registered seed set (findings 1, 2)

The pairs-c2st mean now gates over 20 pre-registered seeds (0–19), not the 5 locked seeds that contain both global-max clippers. The line is unchanged at 0.53 (same ratified derivation: floor mean + 4.2·sd, round 2); only the estimator moves. Only pairs-view c2st moves to the 20-seed mean — every other pairs geometry metric, the battery, and the benefit-space block keep their per-seed 4-of-5-of-locked-5 form (stated explicitly in scope_note).

The new operating_characteristics block makes your finding-2 point as data (normal iid approx, math.erf, rounded to 2dp; the test recomputes every cell):

true mean	per-seed ≥4/5	mean-of-5 (rejected)	mean-of-20 (proposed)
0.525	0.74	0.97	1.00
0.529	0.28	0.65	0.78
0.530	0.19	0.50	0.50
0.531	0.11	0.35	0.22
0.533	0.03	0.13	0.01

Reading the 0.533 row (above the line): per-seed 0.03, mean-of-5 0.13 (looser — your objection), mean-of-20 0.01 (tighter). At 0.530 the mean is 50% by symmetry. So the 20-seed mean sharpens the operating characteristic at the line rather than loosening it — "the number doesn't move" now also means "the gate doesn't loosen." The tighter mean-of-5 line (0.5194) is retained in considered_and_rejected (still machine-checked) with your reasoning: the honest answer to over-rejection is more seeds, not a tighter number.

Required fix C — version-matched cap (findings 3, 4)

The cap is re-derived from the version-matched (sklearn 1.8.0) floor — the 20-seed floor_c2st_distribution in runs/c10_diagnostics_v1.json, the version candidates actually score under:

cap = round(matched_floor_mean + 8 * matched_floor_sd, 3)
    = round(0.5109686 + 8 * 0.0053630, 3)
    = 0.554

not the 0.547 the 1.9.0 sd (0.0045535) produced. The old 0.547 and its 1.9.0-sd derivation are moved to considered_and_rejected with the reason, and it's demonstrable exactly as you said: candidate 5a′ (max seed 0.5510) fails at 0.547 but passes at 0.554.

The version-pin clause's self-contradiction is fixed by estimator:

• Mean line is version-stable. Floor mean 0.5108054 (1.9.0, seeds 0–4) vs 0.5109686 (matched 1.8.0, 20 seeds); the 0.53 line rounds identically under both. So the mean keeps its ratified derivation and does not move.
• The sd is not (0.0045535 vs 0.0053630), so an 8-sd cap does move with the version — which is precisely why the cap must derive from the matched floor. Clause no longer violated at birth.

The cap is described as your finding 3 requires: a prospective one-seed-catastrophe guard that, on the 11 committed runs, changes zero verdicts (it never binds a would-mean-passer; it applies per-seed across all 20 seeds of a future run).

Finding 5 — disclosure (emphasis fixed)

A new disclosure block states plainly, as data:

• The gated statistic under the pre-amendment rule was the locked-5 mean 0.5279, sitting only ~0.80–1.16 SE below 0.53 (0.80 on the 20-seed per-seed sd, 1.16 on the locked-5 sd) — not the 5.0 SE the 20-seed-mean headline advertises.
• Locked seeds 0–1 are the global max and 2nd-max of all 20 (0.5330, 0.5315); P(random 5-of-20 mean ≥ 0.5279) = 143/15504 = 0.009.
• The locked seeds are design-touched (inner validation used inner seed = 1000 + outer seed over outer seeds 0–4).
• The excess over the version-matched floor is real signal at 6.8 SE (0.0124 ± 0.0018) that the 0.53 allowance tolerates; only the per-seed clipping location is noise. The loose "noise, not signal" phrasing is replaced accordingly.

process_statement keeps the goalpost-timing honesty and adds that the 15 fresh seeds were drawn after candidate 10's run, and that the mean rule is proposed regardless because its operating characteristic dominates the per-seed rule for any candidate — your own concession that "the core measurement point is sound and would hold without candidate 10."

Tests

tests/test_gates_derivations.py is updated: the cap binds to round(matched mean + 8·sd, 3) = 0.554 from the diagnostics artifact; the mean line binds to the unchanged 0.53 with the ratified derivation and records the estimator change; the OC table recomputes from the stated normal approx via math.erf (no scipy); the seed set is asserted == list(range(20)); applied: false and all-FAIL committed verdicts are asserted; the no-self-rescue clause text is asserted; both rejected lines (0.5194, 0.547) still bind.

Full suite green from the worktree: 388 passed, 23 skipped (python -m pytest tests/ -q; the 6 net-new tests are the amendment-2 bindings). uvx black -l 79 and uvx ruff check are clean. Pushed as ae0c166; the locked thresholds block and amendment_history are byte-identical (diff confined to amendment_proposed; no runs/ artifact touched).

No merge; the proposal stays inert pending a fresh referee round and maintainer ratification.

[MaxGhenis]

Verdict: RATIFY AS-IS.

Verification round on the reworked proposal (head ae0c166), against the six findings from the AMEND round. I recomputed every load-bearing number from the committed artifacts in a detached read-only worktree — trusting neither the PR text nor my own prior notes. All three required fixes (A/B/C) are genuinely resolved, not renamed or suppressed, and the rework introduces no new defect. The proposal stays inert (applied: false) pending maintainer ratification; this verdict is that the rework answers the findings.

1. Fix A — no self-rescue (findings 1, 6): RESOLVED. Grepped the YAML directly. The no_self_rescue clause is present and unambiguous ("No candidate's committed run verdict changes under a rule proposed after that run… applies only to runs registered after its ratification"). Every one of the 11 committed runs carries committed_verdict: FAIL (candidate 10 included); the retroactive table is applied: false and framed as disclosure only; exactly one would_be_verdict_if_applied: PASS exists (candidate 10) and it flips nothing. path_to_pass requires a fresh registration and discloses that candidate 10's 20 per-seed scores are already public in runs/c10_diagnostics_v1.json (20-seed mean 0.5234), i.e. a c13 re-registration's pairs-c2st outcome is known in advance — exactly the framing fix A required.

2. Fix B — mean over 20 pre-registered seeds (findings 1, 2): RESOLVED. The gated rule is now the across-seed mean of pairs-c2st over seeds 0–19 ≤ 0.53, applied prospectively; scope_note and pass_rule keep every other geometry metric, the battery, and benefit-space on the per-seed 4/5-of-locked-5 form. I recomputed the entire operating_characteristics table myself (normal iid approx, math.erf, per-seed sd 0.0059, line 0.53): per-seed ≥4/5 → [0.74, 0.28, 0.19, 0.11, 0.03], mean-of-5 (se=sd/√5) → [0.97, 0.65, 0.50, 0.35, 0.13], mean-of-20 (se=sd/√20) → [1.0, 0.78, 0.50, 0.22, 0.01] — all match to 2 dp. The sharpening claim holds: at 0.533 (above the line) mean-of-20 = 0.01 < per-seed 0.03 (tighter), while mean-of-5 = 0.13 > 0.03 (the loosening the referee objected to); at 0.530 both means are 0.50 by symmetry. Moving to 20 seeds removes the mean-of-5 loosening.

3. Fix C — version-matched cap (findings 3, 4): RESOLVED. Recomputed the cap from the matched-floor diagnostic_1_seed_extension.floor_c2st_distribution in runs/c10_diagnostics_v1.json (sklearn 1.8.0, the scoring version): round(0.5109686 + 8 × 0.0053630, 3) = round(0.553872, 3) = 0.554 — matches. The old 0.547 is in considered_and_rejected and reproduces from the 1.9.0 floor (round(0.5108054 + 8 × 0.0045535, 3) = 0.547) with the version-mismatch reason. Version-pin is now internally consistent: the mean line is version-stable — I confirmed round(mean + 4.2·sd, 2) = 0.53 under both the 1.9.0 floor (raw 0.52993) and the matched 1.8.0 floor (raw 0.53349) — while the per-seed sd is not (0.0045535 vs 0.0053630), which is exactly why the sd-derived cap must use the matched floor. The c5a′ demonstration checks out: max seed 0.5510 fails 0.547 but passes 0.554.

4. Finding 5 — disclosure: RESOLVED. The disclosure block states the gated pre-amendment statistic is the locked-5 mean (I recomputed 0.52787 → 0.5279) sitting ~0.80–1.16 SE below the line — confirmed as the SE of the 5-seed mean: margin/(0.0059/√5) = 0.806, margin/(0.0041/√5) = 1.160 — not the 5.0-SE headline of the 20-seed mean (recomputed 20-seed-mean margin 5.009). The permutation P(random 5-of-20 mean ≥ 0.5279) recomputes to 143/15504 = 0.0092 exactly (I counted the 5-subsets). Design-touched seeds (inner seed = 1000 + outer over outer 0–4) are named. The excess-over-floor framing recomputes to 6.826 SE (0.0124 ± 0.0018), correctly characterized as tolerated signal with only the per-seed clipping location as noise.

5. Tests: PASS and genuinely binding. Full suite from the worktree: 388 passed, 23 skipped, 0 failed (exit 0). The 10 amendment-2 tests recompute each stated number from the artifacts. I confirmed they are load-bearing by perturbing an isolated scratch copy (worktree untouched): cap 0.554→0.555, an OC cell 0.01→0.02, applied: false→true, and dropping seed 19 from the seed set each break their specific test; restoring returns 10/10. test_amendment2_illustrative_block_is_not_applied also asserts the old verdict_recomputation key is structurally absent, so the self-rescue cannot silently return.

6. Scope + accommodation: CLEAN. git diff origin/master...HEAD touches only gates.yaml (+706/−0, a single append hunk inside gate_1 after amendment_history) and tests/test_gates_derivations.py (+373/−0, a single append hunk after line 294). Zero deletions: the locked thresholds block (including the pairs-view c2st_auc_max: 0.53), amendment_history, the 19 pre-existing derivation tests, and every runs/ artifact are byte-untouched. On the second question — no new free parameter is resolved in candidate 10's favor. The cap's k=8 does zero work on the evidence and cannot affect c10 (its max seed 0.5330 clears the cap at any k ≳ 4.1) and is anchored at ~1.9× the ratified 4.2 mean-line multiplier, not tuned to c10. The 20-seed set is the full measured set (anti-cherry-pick; it includes the two global-max clippers) and is justified by the candidate-independent OC-sharpening argument the referee itself requested. The line stays at the ratified 0.53, which is if anything conservative relative to the matched-version raw value (0.5335). The one residual — that the 20-seed mean flatters c10 relative to the gated locked-5 mean — is both disclosed explicitly and rendered moot by fix A (c10 stays FAIL; a fresh registration is required). The accommodation is neutralized structurally, not argued away.

All six findings resolved; no regression. The goalpost-timing judgment the proposal defers to the maintainer is handled exactly as the AMEND round required — disclosed plainly, self-rescue removed. Ratifiable as-is.