Evaluate the active Phase 1 reference on the human-benchmark pool

[NeoLorenzo]

Audit Priority: P1
Audit Type: VALIDATE

Problem / Opportunity

The repository's complete 17-map PPO comparison on the human-benchmark split uses the superseded Phase 1 v3 Seed3 16M model. The active frozen reference is Phase 1 v5 PARITY002 Seed3 16M, so the model side of the current human-relative evidence is stale.

A valid current human/model comparison requires the active frozen reference to be evaluated on exactly the same 17 benchmark maps under the same policy-visible environment/interface contract and recorded episode configuration.

Why This Matters

Issue #15 cannot compute a meaningful current human-relative gate from complete human evidence if the paired model evidence comes from an older observation/interface contract. The repository explicitly treats the v5 PARITY002 run as the active frozen Phase 1 reference.

Evidence

• README.md names the Phase 1 v5 PARITY002 Seed3 16M Terminal-SPT run as the current frozen reference benchmark.
• docs/human-benchmark.md still reports the full 17-map PPO human-benchmark baseline from the superseded v3 Seed3 16M model (mean 16.71 SPT, median 17.00 SPT).
• The same document notes that the current official human interface exposes the v5 human-information-parity observation/action contract.
• tools/evaluate_phase1.py already accepts --pool human_benchmark and verifies map hash, legal-action validity, checkpoint interface compatibility, and the Turn-10 contract for every episode.
• Issue Add canonical human-benchmark evaluation provenance #14 separately owns classifying a full human-benchmark model run as canonical human-benchmark evidence; this issue owns producing the active-reference result itself.

Proposed Outcome

Run the active frozen Phase 1 v5 PARITY002 Seed3 16M reference checkpoint across all 17 human-benchmark maps and preserve a complete model-side artifact suitable for later paired comparison with canonical human first-attempt results.

Definition of Done

• The active frozen Phase 1 v5 PARITY002 Seed3 16M reference checkpoint is evaluated on all 17 human-benchmark maps.
• Evaluation uses the exact 17 manifest-assigned human-benchmark maps and verifies each CSV/canonical identity.
• Evaluation uses the benchmark's policy-visible interface and episode configuration required for human/model parity.
• All 17 model episodes satisfy the Turn-10 contract and action/interface validity checks.
• The saved artifact contains checkpoint/sidecar hashes, environment/interface metadata, map identities, episode seeds, and enough provenance to join each model result to the corresponding canonical human result.
• Aggregate model metrics for the 17-map pool are recorded, including mean and median Turn-10 SPT and uncertainty appropriate for the 17-map sample.
• Documentation reports the active v5 human-benchmark comparison while preserving the historical v3 result as superseded evidence.

Constraints / Non-Goals

This issue does not collect human attempts (#12), change evaluator provenance/classification (#14), or define the statistical pass/fail gate (#15). Do not replace or rewrite historical v3 artifacts.

Validation

Verify the completed artifact contains exactly 17 unique benchmark map identities, all drawn from the authoritative human_benchmark split, and that every episode records contract_valid: true. Confirm the checkpoint and action-interface sidecar hashes match the frozen v5 reference used in the current reference documentation.