vyndr/specs/re-adjudication.md

RE-ADJUDICATION UNDER CUMULATIVE CORRECTION

2026-08-04. Counter byte-identical. Nothing recalibrated, because nothing needed to be.

There is nothing to re-adjudicate: the proven set is empty, and has always been empty. Verified three ways below.

And the cumulative correction did not catch a false positive last session. It caught nothing — there was nothing in the proven set to catch. What it did was tighten α from 0.0026 to 0.0013 within one session, which demonstrated the mechanism working, not a demotion.

The real exposure was elsewhere, and it was real: a promotion could bypass the cumulative correction entirely. That hole is now closed structurally, so the retroactive pass this order describes can never become necessary.

---

1. The registry state, verified (STEP 1 + 2)

scripts/proven-status.js  → PROVEN_SET: EMPTY
                             hits LOSES · total_bases INCONCLUSIVE · strikeouts INCONCLUSIVE

featureRegistry           → PROVEN features: ["recent_frequency_prior"]
                             validatedSkills(BOMBER|GHOST|TORCH|FLAME|SCALPEL|SINKER) = {}
                             conditioning entries recorded: 0

The single PROVEN feature is recent_frequency_prior — the incumbent counter itself, proven by the S78 ablation showing it is ~100% of the champion's resolution. It is the baseline every challenger is measured against, not a conditioning interaction, and demoting it would leave the model with nothing to grade from. No conditioning entry has ever reached PROVEN, so there is no archetype-slot to re-run and no grade that depends on a demoted feature.

STEP 3 (demote) and STEP 4 (recalibrate) are therefore vacuous — correctly. readjudicateAll([], …) returns NOTHING TO RE-ADJUDICATE — the proven set is empty, which is the honest output rather than a no-op glossed over.

2. The hole that WAS open, and is now closed

The order's underlying worry — make sure the standard applied from the start, so no retroactive correction is ever needed — was well founded. Two audits:

(a) promote() did not require the cumulative denominator. It checked n, lift and CI, but nothing stopped a future session from testing eight hypotheses, correcting by eight, and promoting on a p-value that would not survive the programme's real denominator. That is exactly the hole that makes a "re-adjudicate everything" pass necessary later.

Now structural — isSufficient refuses evidence that:

• carries no bonferroni_tests at all;
• was corrected against fewer tests than the cumulative count;
• has a p_value that does not clear 0.05 / bonferroni_tests.

no correction recorded                    → refused
per-session (8) when cumulative is 38     → refused
cumulative (38) but p = 0.01 > 0.0013     → refused
cumulative (38) and p = 0.0005            → accepted

The same rule guards recordConditioning(status: PROVEN). Locked by test.

(b) Two of four analysis scripts still corrected per-session. pitcher-prove-k.js and tb-solo-and-interactions.js used SOLO.length + INTERACTIONS.length; both now use the cumulative ledger, as cluster-prove.js already did. Cumulative correction is native across every analysis path.

3. Standing re-ablation (STEP 5, second line) — built

src/services/model/reAblation.js. Pure and injectable: it takes evidence and returns auditable verdicts, performs no measurement, and reaches no database, so the decision rule cannot quietly drift from the gate's.

Every verdict records both p-values and both test counts, so a demotion can be re-derived by anyone:

original_p_value · original_bonferroni_tests · original_corrected_alpha
current_p_value  · cumulative_bonferroni_tests · cumulative_corrected_alpha
verdict + reason

A feature promoted at α=0.05/20 and re-tested against α=0.05/60 can demote on the same p-value. That is not unfair — the bar rose because the programme has since had more chances to get lucky, and holding old results to the old bar is how a lucky draw becomes permanent.

No fresh measurement is PENDING_RETEST, never a demotion. Absence of a re-test is not evidence a feature stopped working, and demoting on it would punish whichever stat happens to be off-season.

Three ways a proven feature can stop being true are documented in the module: it was never true; the game changed; or the bar rose.

4. Net effect on the proven set

Zero. Nothing survived because nothing was there; nothing demoted; nothing recalibrated; no ledger event was emitted, because emitting "recalibrated after re-adjudication" when nothing changed would itself be a false signal of rigour.

Frozen clusters were not exempted — they were examined and contain no proven conditioning. The counter is byte-identical.

5. Going forward

• Cumulative correction is native on every analysis path and required for every promotion.
• Standing re-ablation is available and should run whenever a proven entry exists and fresh sample arrives.
• The bar only rises. Current cumulative count 38, α = 0.0013. Every new hypothesis tightens it for everything after — so re-testing standing candidates is now mathematically the disciplined choice over inventing new ones.
• The first promotion this programme ever makes will already have been held to the cumulative bar, which is the outcome this order was actually asking for.

Not recommended: treating the empty proven set as a problem to solve by lowering the bar, or emitting a public "re-adjudication" event for a pass that changed nothing.