MichelLiao/foresightcoder-qwen-eval-artifacts

ForesightCoder Qwen evaluation artifacts

Private snapshot updated 2026-09-24. The latest results are the reconciled 212-task runs below. Historical folders are retained, but are superseded for this comparison.

Latest results

Model	PR-1 F2P	PR-1 P2P	Task Score	PR-2 F2P	PR-1 Resolved	PR-2 Fully Anticipated	resolved & anticipated
Qwen3.5-27B	29.00%	89.01%	59.01%	26.00%	39/212 (18.40%)	48/212 (22.64%)	13/212 (6.13%)
Qwen3.8-27B	36.86%	89.74%	63.30%	28.37%	54/212 (25.47%)	53/212 (25.00%)	16/212 (7.55%)

All metrics use the same 212-task denominator. Per-test percentages are macro averages across tasks. Task Score = (PR-1 F2P + PR-1 P2P)/2. PR-1 Resolved requires 100% F2P and all regression tests passing. PR-2 Fully Anticipated requires all held-out anticipation tests to pass; these are anticipation scores, not PR2-direct inference. The last column is the intersection of resolved and fully anticipated. Qwen3.8's five invalid and three empty submissions receive zero in this table. The original harness aggregates preserve report-only means (204 Qwen3.8 reports), so their means differ from this full-212 comparison. No cases are excluded. F2P-only full-pass counts, if needed, are 43/212 for Qwen3.5 and 60/212 for Qwen3.8.

Authoritative run directories

• Qwen3.5-27B
• Qwen3.8-27B
• Machine-readable scores and CSV
• Per-task validation evidence
• 424 task grades

The run directories include trajectories, selected patches, predictions, test logs, JUnit, aggregate reports, validation summaries, and retry-selection provenance. The frozen dataset has 212 tasks at revision 0e44f45d7df7671f4ce01c5da3608905526156aa of shaswat123/foresightcoder-anticipation. These are reconciled authorized-attempt results, not a first-attempt-only benchmark. Normal empty-patch retries were limited to one; six confirmed 250-turn cutoffs received an explicitly authorized extra attempt at 350 turns. No new inference was performed during final grading validation.

Validation and limitations

All 424 grading outcomes were checked against current expected-test lists. There are 416 test reports plus eight provenance-verified bad submissions. A model's compilation/import/early assertion failure can prevent tests from executing; main scores those expected MISSING tests as zero. This is distinct from a grading infrastructure failure. No grading infrastructure failures remain unresolved. Gold controls were audited for all 212 tasks; stale/failing controls were refreshed. The gold-audit JSON rescored historical XML with current expectations and takes precedence over historical control report percentages. Shared grader ba90945; 58 harness regression tests pass. Earlier coverage files flag missing test records; use the new grading_validation.json and evidence for final adjudication.

Dataset caveat: svelte__9826-10786's gold PR1 solution already passes both anticipation tests. It remains included identically for both models. This is an unresolved benchmark-definition/execution discrepancy, not proof of genuine anticipation on that case. Both models receive anticipation credit there.

Some grading runtimes were repaired consistently for gold and both models, including Node versions, localhost/IPv6, and canonical workspace test metadata. Axios's unchanged outbound HTTPS protocol test uses an isolated local TLS fixture because compute-node egress is blocked. Certificate verification remains enabled; the negative/positive TLS control is included. No test assertions were relaxed.

Privacy and integrity

The repository remains private. Known credentials and recognized token strings are redacted in uploaded copies; local originals remain unchanged. The snapshot manifest hashes describe this update's sanitized artifacts; historical files remain in the repository. Local paths/process metadata remain. Upstream excerpts retain their original licenses. No credentials, container images, model weights, or TLS fixture private keys are part of this update. There is no automatic synchronization.