โ€บNavigation

Iteration 64 โ€” ADMIT #12: the Calibration Gap Meter STRONGEST M4 LIFT OF THE CAMPAIGN

In plain words: the candidate that survived where three whole schools fell finished with the campaign's best result. Plugged into our verdict machinery on strictly held-out later regimes, its confidence-honesty readings improved forecasts by nearly TWICE the previous record โ€” with the entire uncertainty band positive and far beyond anything lucky pairing produces. The era-stability question that executed the gap meter answered itself: a reading whose meaning flips between eras cannot produce an all-positive held-out result, and this one did. Twelfth certificate issued, for a channel nothing else on the shelf reads: does a model's stated confidence stay honest in every regime? One more scientific deposit: for the FOURTH time through a fourth lens, features that look worse by a stability-adjacent measure keep their edge more often โ€” the iteration-9 inverted mechanism is now corroborated four ways, begging for its own dedicated investigation. One candidate pair remains before the batch closes.

Preflight (A0): 2026-07-10 11:45 โ€” load1 6.16/32c (โ‰ค24) ยท one 1-core external python job (under bar) ยท 26 GiB avail, si trickle/so=0 ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. One capped run, 3.03 min, readonly=2 loader, watchdog clean, leakage-safe expanding PIT (ฯƒ recomputed within windows 0..t).

Finale results (frozen disjunction, calibration_gap_m4_m5b.py)

LegReadingRuling
M4 held-out incremental valueฮ”AUC = +0.01406, CI95 [+0.00742, +0.01948] โ€” entirely positive, far outside the pairing-null band [โˆ’0.0014, +0.0008], and nearly 2ร— the previous best (Causal Dantzig and Andersonโ€“Rubin at +0.0073). The temporal-stability question answered by the leg itself.PASS โ€” campaign record
M5โ€ฒb directionฯ = โˆ’0.159, CI entirely negative โ€” calibration-divergent features KEEP orthogonality more: the inverted mechanism's fourth sighting (verdict-stability ยท anchor lenses ยท divergence lens ยท now calibration). Fails the one-sided leg; the disjunction admits on M4.LOGGED (discovery #4)
Noveltymax |ฯ| vs base gates = 0.244 โ€” not a repackagingCLEAR
Verdict (row 75)ADMIT(forex, 8-window EURUSD@5 regime grid, role = regime-calibration divergence meter โ€” the theorem-linked confidence-honesty channel, expiry = next completed forex regime slice) โ€” ADMIT #12; full trail rows 73 โ†’ 74 โ†’ 75

โ–ถ Next iteration

Iteration 64 ยท 2026-07-10 ยท ADMIT #12 (row 75) ยท capped (3.03 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: calibration_gap_m4_m5b.py ยท calibration_gap_m4_m5b_results.json