java-ai-golden eval

Latest reports from this GitHub Actions run on main. Fixture numbers are contract checks on recorded answers, not a live-model score. Live smoke and nightly stay in Actions artifacts unless that job wrote into the same build/eval tree.

Reports

• Development (fixtures)
  100.0% (8 / 8) · 2026-10-14T01-29-39 · PASS
• Holdout
  100.0% (8 / 8) · 2026-10-14T01-29-42 · PASS
• Judge calibration
  canned or live judge corpus; does not gate production scores

Each folder also has eval-report.md and run.json. Semantics: evaluation-methodology.md.