Asked whether two more sophisticated dependence detectors (GCM and O-information) deserve to join the four measures already used to judge if a feature is genuinely new. Both proved correct and stable, but neither was shown to catch anything the existing four miss, so neither was adopted.
| Quantity | Value | Meaning |
|---|---|---|
| Mutation kill rate | 201 of 201 non-equivalent mutants (0 boundary survivors, 45 equivalents excluded) | no deliberate bug in the decision code slipped through |
| Effect-size separation | orthogonal xi-max 0.182 vs redundant xi-min 0.968 | the 0.80 cutoff sits in an empty gap |
| Out-of-sample Jaccard | GCM 0.635; O-information 0.772 | selections repeat on unseen folds |
| Value-add instances | 0 of 330 construction pairs (GCM); 0 of 110 triplets (O-info), 110 valid cells | the decisive gate returned empty |
| Jackknife 95% CI over 6 symbols | GCM [0.569, 0.701]; O-info [0.689, 0.854] | both lower bounds above the 0.20 floor |
| Firing | What it did |
|---|---|
| 01 | Broke the decision code 201 ways; every break was caught |
| 02 | Set the cutoff from a null built on our own data, not borrowed convention |
| 03 | Re-ran selection with both leaks removed; checked picks held on unseen folds |
| 04 | Decisive test: did either detector find anything the baseline four missed |
| 05 | Error bars by dropping one symbol at a time |
| 06 | Decision recorded; neither detector added |
Flipping either to promoted requires a non-empty FDR-survived value-add test via a nonlinear/ratio construction identity, or a Pearson-independent forward-return answer key. Deployment is Terry-gated. Nothing is blocked meanwhile. Data-hygiene note: manifest.json cites audit_twin candidates/gcm-oinfo-selection-s0.md, which does not exist.
manifest verdict {GCM: WATCH, O-information: WATCH}; iteration 6 titled 'Final verdict — both WATCH (not promoted)'. A decision was reached and recorded. (LOOP_LEDGER.json _status_evidence)
Lifecycle status is process state, not a judgement of the findings. Results are stated as numbers with their uncertainty. Generated between CLOSING-RECONCILIATION markers — regenerate rather than edit.
Can GCM (A2 conditional) and O-information (A3 synergy) be promoted on top of the production orthogonality baseline {Pearson, Spearman, h_norm, MI}? Not yet — both WATCH. Correct, calibrated, reproducible, leakage-free, stable — but value-add over the baseline is not demonstrable on this panel's only non-circular ground truth.
| # | Iteration | Verdict |
|---|---|---|
| 01 | S1 — Code correctness (mutation oracle) | PASS |
| 02 | S2 — Calibration / effect-size separation | PASS |
| 03 | S3 — De-leaked out-of-sample stability | PASS |
| 04 | S4 — Value-add over baseline (DECISIVE) | WATCH |
| 05 | S5 — Confidence intervals (jackknife) | reported |
| 06 | Final verdict — both WATCH (not promoted) | WATCH |
Audit twin: candidates/gcm-oinfo-selection-s0.md · Full verdict page: instruments/gcm-oinfo-selection-verdict.html