โ€บNavigation

Iteration 60 โ€” H-063 V-REx Passes the Entrance Exam EXAM PASS ยท GAP-AFFINITY CHECK DECISIVE NEXT

In plain words: the regime-variance candidate asks the softest version of the worst-case question: not "how bad is the worst regime" but "how UNEVEN are all the regimes together?" Its exam passed cleanly โ€” exactly zero unevenness on the perfect-by-construction model, and on the deliberately wrong model an unevenness reading twice as large as anything label-shuffling of the same real losses can produce (that shuffle test was this candidate's paperwork requirement, wired in on schedule; the honest 2ร— margin โ€” not astronomic โ€” is on the record, a consequence of heavy-tailed losses). But a shadow hangs over the ladder: unevenness (variance) and the just-executed worst-regime gap are two summaries of the SAME per-regime numbers. If the field test shows they rank features nearly identically, the gap's fate โ€” actively harmful at the final hurdle โ€” transfers, and this candidate dies cheaply at the census rather than expensively at the finale. That affinity check is the next firing's first question.

Preflight (A0): 2026-07-10 10:25 โ€” load1 2.36/32c (โ‰ค24) ยท no heavy jobs ยท 31 GiB avail, swap fully drained (si/so=0) ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. One capped run, 0.48 min, readonly=2 loader, real rows only (permutations of real values per ยง0). Provenance note: reference repo lacks a license file โ€” moot, the readout is clean-room three-line arithmetic; recorded.

Exam results (pre-registered form, verex_exam.py)

LegReadingRuling
E1 โ€” true model reads perfectly evenvar over envs = 0 exactly (every per-env loss < 10โปยนยฒ, degenerate guard)PASS
E2 โ€” wrong model's unevenness realvar = 0.865 > the 97.5th percentile (0.435) of K=199 env-label permutations of the REAL per-obs losses โ€” the dossier's REQUIRED readable-loss null, wired in-exam. Margin: an honest 2.0ร— (heavy-tailed losses make the shuffle null high โ€” recorded, not hidden)PASS
Dial auditNone โ€” variance has no dials; K=199 is the campaign-standard null sizeZERO DIALS
Named ladder risk (decisive next)Variance and the executed H-060 gap are spread-measures of the SAME per-env losses. The census measures ฯ(var, gap) using row-69's stored per-feature gaps: near-identical โ‡’ the H-060 M4 fate (actively harmful, temporally unstable, row 70) transfers โ‡’ cheap census kill instead of an expensive doomed finaleGAP-AFFINITY CHECK NEXT
Verdict (row 71)EXAM PASS โ€” checkpoint; ladder walk unlocked with the affinity question front-loaded

โ–ถ Next iteration

Iteration 60 ยท 2026-07-10 ยท CHECKPOINT (row 71) โ€” exam PASS ยท capped (0.48 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: verex_exam.py ยท verex_exam_results.json