Files
MeltMind/docs/phase2-gate-outcome-results-2026-11-18.md
T

3.3 KiB
Raw Blame History

Gate v2 outcome comparison: complete, fixed equivalence gate not passed

The one approved continuation and postfit completed. Decision FAIL_OUTCOME_EQUIVALENCE: none of the four confidence intervals is wholly inside its frozen equivalence band. This fails the specified acceptance gate; it is not proof of non-equivalence or a rejection of all gate-dependent methods. Original exact seed-set FAIL remains unchanged.

Original333 feature results were reused by hash and160 never-attempted parts completed PASS. Both preparation batches exited before custody. Original frozen A/G models produced493 new predictions;52 exact-gate predictions and metrics were reused. All545 predictions were frozen before493 existing target-cache reads. Seven invalid cases remain in the552 eligibility denominator (7/552). No new fit, decode, retry or clipping.

Model Metric Label − LOG mean 95% CI Fixed band Gate
A relative_l2 0.000826489915 [0.000389475815, 0.001336301591] [-0.001, 0.001] FAIL
A last_location 0.004182083689 [-0.004048948093, 0.013011176499] [-0.01, 0.01] FAIL
G relative_l2 0.000768036482 [0.000347445584, 0.001245813598] [-0.001, 0.001] FAIL
G last_location -0.014814399934 [-0.045660884822, 0.006501384511] [-0.01, 0.01] FAIL

Each comparison has545 paired parts /202 equally weighted components. Two components have no defined paired parts and are excluded; partial components are retained under the unchanged paired-defined-parts rule. All2000 bootstrap replicates are defined, seed2026111408. A/G field differences are positive with intervals above zero for these fixed models; neither last-location interval establishes a sign. These observations do not change the pre-specified four-band acceptance rule.

Both field upper bounds exceed+.001. A last-location upper bound exceeds+.010; G last-location lower bound is below−.010. No threshold adjustment or follow-on experiment is authorized by this result.

New independent PASS1f3a4778 was fixed as implementation-review8e2b3a07 and actually validated before prediction/target access. Implementation5ad512e9, contractf70de9b9 and comparison6b79ace1 remain unchanged. Five prediction guards and one score guard all PASS. Actual service is inactive/dead, MainPID0/exit0; coordinator790438 and workers exited. Original S2 training701941 remains active and untouched.

Formal feature post-freeze prior is94090.936844268CPU-s, already including old66936.020178167 once. Postfit through terminal freeze is170.203873842CPU-s; cumulative94261.14071811 /104639.3107342874 and postfit≤3600 both PASS. Separate metadata/hash/reporting overhead is recorded additively in report-post-cost.json without rewriting original execution receipts.

Immutable prediction freeze: e2aaa60bf6e780b69465eee0de28c0c3703123ed; execution terminal: 25bce707d691bdfd2f5eee4af03307fc45754804; score SHA256: de5ec5e111c8142a3a79600ca953326a1087fa0cf3e01ba4e38bf169534fc058. Full545 prediction bytes have been independently hash-checked in this terminal audit; no metrics or targets were recomputed. Evidence root: /home/frank14f/MeltMind-runs/phase2-gate-outcome-v2-continuation-postfit-2026-11-17.

The result is submitted to Claude for the existing collaborative stage review. No new research stage, retraining, geometry rerun or threshold change is launched.