โ€บNavigation
MHC6 selection sweep ยท S0โ€“S5 ยท 2026-06-07
BOTH WATCH

GCM + O-information โ€” Promotion Verdict

Can these two detectors be promoted on top of the production orthogonality baseline {Pearson, Spearman, h_norm, MI}? Not yet โ€” both WATCH.

S1 kill rate
1.00
0 boundary survivors
S2 separation
0.18 / 0.97
orthogonal vs redundant ฮพ
S3 OOS Jaccard
0.635 / 0.772
GCM / O-info, leakage-free
S4 value-add
0 / 110
instances on non-circular GT

Decisive gate S4 = empty.

The only non-circular ground truth โ€” the exact identity volume = buy_volume + sell_volume (verified to 2.2e-16) โ€” is 100% Pearson-visible across all 110 cells (buyโ†”sell |ฯ|โ‰ฅ0.41, never <0.30). So no marginally-orthogonal redundant structure exists for GCM/O-info to flag that the baseline does not already catch. GCM flags 232/330 construction pairs โ€” all baseline-visible. Not a failure of the detectors; the test lacks a valid positive case. Confirmed by a 3-lens adversarial review (0 material flaws).

Gate scoreboard (per candidate)

Gate What it proves GCM (A2) O-info (A3)
S0Pre-registration freezefrozenfrozen
S1Code correctness (mutation oracle, kill 1.0)PASSPASS
S2Calibration / effect-size separation 0.18 vs 0.97PASSPASS
S3De-leaked OOS stability (Jaccard 0.635 / 0.772)PASSPASS
S4Value-add over baseline (0 instances)WATCHWATCH
S5Confidence intervals (jackknife)reportedreported

What would flip to PROMOTE

A non-empty, FDR-survived S4 on valid ground truth: (1) a nonlinear/ratio construction identity that is redundant-by-arithmetic yet marginally Pearson-orthogonal (verify the Rust feature definitions first), or (2) a Pearson-independent forward-return answer key. Deployment is Terry-gated regardless.