Skip to content

Build Week 2026.1: Adversarial contract evidence

Choose a tag to compare

@Kkkakania Kkkakania released this 18 Jul 07:48
dc2df7a

This incremental Build Week release hardens the Codex-to-MATLAB boundary with a reproducible adversarial benchmark.

  • Accepts one checked GPT-5.6 control review
  • Rejects 14/14 hostile mutations fail closed
  • Covers executable-action injection, candidate/model spoofing, field smuggling, invalid scores, contradictory verdicts, oversized findings, and malformed JSON
  • Commits human-readable and machine-readable evidence
  • Runs the benchmark in the release gate

Reproduce:

./scripts/run_review_contract_benchmark.py \
  --validator scripts/validate_plot_review.py \
  --manifest examples/review/multi_series_manifest.json \
  --review examples/review/multi_series_review.json \
  --out /tmp/plot-review-contract-benchmark

Verification: PR #43 passed both Quality checks, and main Quality run 29636319169 completed successfully. All inputs are synthetic and reports contain no private paths.