GWAYA v3.1.1: Fail-Closed Gate & Real Toolchain Verification for LLM Code Generation (corrected)

Correction: the first 3 October 2026 revision of this report (DOI 10.5281/zenodo.23121788, version 3.1.0) claimed that the low-tier optimizer improved pass rate by at least 8 points. That claim rested on a results file that was written by hand instead of being produced by the benchmark. It is withdrawn. This version reports the benchmark that was actually run.

Scope-limited report restricted to executable, reproducible evidence. GWAYA accepts a candidate only if a deterministic check succeeds (AST stub audit, Python parser with sandboxed test execution, rustc check, Lean 4 kernel with #print axioms); missing toolchains and generators yield UNVERIFIED.

Model qwen2.5-coder:1.5b on MBPP-sanitized test. Primary pre-registered run (local CPU): n=60, pass@1 A0 (raw)=68.33%, A3 (best-of-3 + repair)=75.0%, delta +6.67 points (95% bootstrap CI [+1.67, +13.33], A3 gained 4 / lost 0 problems, exact McNemar p=0.125); pre-registered gate (A3-A0 >= +8.0 points): FAIL. Replication on the full split (NVIDIA L4 (ollama ps: 100% GPU) / 8 vCPU): n=257, pass@1 A0 (raw)=64.2%, A3 (best-of-3 + repair)=71.98%, delta +7.78 points (95% bootstrap CI [+4.67, +11.28], A3 gained 20 / lost 0 problems, exact McNemar p=0.0); pre-registered gate (A3-A0 >= +8.0 points): FAIL. Both gate verdicts are reported; the threshold was not changed after the results.

Audit workflow status: APPROVED. Regression suite: 132 passed, 1 skipped, 1 warning in 124.31s (0:02:04). The bundle holds sources, tests, raw per-problem results, the pre-registration, the generation cache and SHA-256 checksums.