Iteration 20 — Candidate A Extraction Checkpoint: E-Values Fit the Gap Better Than Hoped

Iteration 20 — Candidate A Extraction Checkpoint CHECKPOINT

In plain words: the first candidate paper was pulled and read (via the project's own scraper). It fits the campaign's open gap better than hoped, for three reasons: (1) its tests are computed by reshuffling real data — exactly the style the no-synthetic rule mandates; (2) they work for any test statistic and never divide by a residual variance — so the perfect-fit breakdown that killed the old instrument mathematically cannot happen; (3) the same framework yields always-valid sequential monitors — which is what the expiry alarm (the other starved boundary) needs. On-paper reasoning even suggests it passes the perfect-duplicate entrance exam where the old instrument failed. That claim now has to survive contact with real data: the entrance-exam harness is the next slice.

Preflight (A0): 2026-07-03 23:18 UTC — load1 3.15/32c · CH + sidecar active → ALL PASS. Rebased onto moved main (22 commits, clean). Scrape via firecrawl (one transient failure, retry OK). Zero ClickHouse; zero synthetic data.

Extraction highlights (arXiv 2310.01153 v5)

Property	Why it matters here
E-values against group invariance, arbitrary test statistics, Monte Carlo over the group	Environment-permutation invariance = finite group → permutations of REAL rows (03c-compliant by construction)
No residual-variance division anywhere	The causalicp killer (zero-residual degeneracy) cannot occur; under the ground-truth duplicate the true subset should read e≈1 (accepted) while wrong subsets accumulate evidence
E-processes / test martingales, arbitrary filtrations	Ville-bounded sequential monitoring — the Expiry boundary (Frontier #4 family) served by the same framework
M1 anchors	(a) matched permutation null AND (b) anytime-valid error bound — the strongest combination in the campaign's taxonomy

Property

Why it matters here

E-values against group invariance, arbitrary test statistics, Monte Carlo over the group

Environment-permutation invariance = finite group → permutations of REAL rows (03c-compliant by construction)

No residual-variance division anywhere

The causalicp killer (zero-residual degeneracy) cannot occur; under the ground-truth duplicate the true subset should read e≈1 (accepted) while wrong subsets accumulate evidence

E-processes / test martingales, arbitrary filtrations

Ville-bounded sequential monitoring — the Expiry boundary (Frontier #4 family) served by the same framework

M1 anchors

(a) matched permutation null AND (b) anytime-valid error bound — the strongest combination in the campaign's taxonomy

▶ Next iteration

Next firing: the entrance-exam harness — an e-value invariance test per anchor-subset, permutation-calibrated, run on the committed perfect-duplicate case (must identify {ofi}) and on the 4 STABLE-candidate features. Pass → full ladder walk begins for candidate A. Standing for the operator: the 73 grid-corroborated draft statuses and the Frontier #4 precondition review.

Iteration 20 · 2026-07-03 · candidate A extraction checkpoint (LEDGER row 26) · append-only