โ€บNavigation

Iteration 31 โ€” H-023 IAS Passes the Entrance Exam EXAM PASS ยท REDUNDANCY WARNING

In plain words: next in the queue: Invariant Ancestry Search. Its pitch โ€” when full identification is too hard (our usual situation), settle for a weaker but more attainable answer: find the smallest collections that hold steady across regimes, and take everything they mention as "ancestry". It passed the ground-truth exam cleanly: rejected the empty answer, found exactly one smallest steady set, named exactly the right variable. But there's a catch the loop wrote down in red ink before celebrating: this candidate runs on the same underlying test as the reigning identifier โ€” it differs only in how the answers are combined. Two dials wired to the same sensor usually read the same. Whether the different combination logic produces genuinely different information across real features is exactly what the redundancy gate will decide next โ€” and the loop expects that fight to be close.

Preflight (A0): 2026-07-10 โ€” load1 3.85/32c (โ‰ค24) ยท 24 GiB avail ยท CH + sidecar active ยท no heavy jobs โ†’ ALL PASS. One capped run (0.47 min), complete-case real rows, permutations of real values only.

Exam design (pre-registered in-script)

ElementChoice
Candidate mechanism under examThe ancestry READOUT LOGIC (Mogensenโ€“Thamsโ€“Peters 2022): minimally invariant sets over the lattice including โˆ… โ†’ union โŠ† ancestry. Per-set invariance oracle = the ADMITTED e-ICP-fp v5 test (candidate's identity is the aggregation, not a new statistic).
Data conventionsv5 form โ€” 10 EURUSD@5 envs ยท complete-case ยท standardized design + target ยท twin target intra_duration_us โ‰ก anchor duration_us
PASS criteriaโˆ… rejected ยท {duration_us} minimally invariant ยท no wrong minimal set ยท ancestry union = {duration_us} exactly

Result

ReadingValueCheck
โˆ… (target itself invariant?)rejected (regime-varying, as required)โœ“
Accepted setsexactly the 4 duration-containing sets (e_total = 1.0, 5/5 degenerate folds); wrong sets rejected at e up to 88,178โœ“
Minimally invariant{duration_us} โ€” uniqueโœ“
Ancestry union{duration_us} โ€” exactly rightEXAM PASS (row 40)

Redundancy warning (recorded before the ladder): IAS-on-e-ICP shares its per-set e-values with the admitted identifier โ€” only the aggregation differs (union-of-minimal vs intersection-of-accepted). Two dials on one sensor. The M3 correlation attack on per-feature forex readouts is the likely kill; survival requires demonstrably different information โ€” e.g., non-empty ancestry where the intersection is empty, with uncorrelated grading.

โ–ถ Next iteration

Iteration 31 ยท 2026-07-10 ยท CHECKPOINT (row 40) โ€” exam PASS ยท capped ยท zero synthetic data ยท append-only ยท evidence: ias_exam.py ยท ias_exam_results.json