Real candidates, bounded authority.

Codex generated each candidate inside the validated isolated runner. Candidate prose never decided acceptance; the independent evaluation report did.

Accepted candidates

2 repair repositories were selected from 2 affected consumers.

Human review

github.acme/job-insights-skill-holdout

05fa8eb0c1f6399bff80

Real codex · attempt 1

Base commit: da3c0f5015e638a8ef34a2ef9e9a7a822892ef8b
Candidate tree: 68300fd8290b36bf50475f68fb2ed1b946cff903
Changed files: src/job_insights/adapters/status_client.py, src/job_insights/services/summary.py
Diff summary: 2 files, +6 / -3 lines
Evaluation checksum: 11eae4f681e8bf6079a32644773ead29f87033c13e59a6d19c01a36a68d2b553
Model selection: Not observable in validated evidence

View committed diff Open draft PR #1

Human review

github.acme/operations-dashboard-agent-holdout

9c7f3aaf1dc9788d3f14

Real codex · attempt 1

Base commit: 5efe44677d3a5957945d8892c1301797b6f9dedf
Candidate tree: 3bee655476f9a07ae90a283ea7bf12126879725e
Changed files: src/acme_ops/integrations/job_gateway.py, src/acme_ops/services/dashboard_service.py
Diff summary: 2 files, +3 / -3 lines
Evaluation checksum: 3ff3acd6fa7c31f38c9cb2abb701351bd8d306eb5b2b3909c6ce1c0fa3b9ad9e
Model selection: Not observable in validated evidence

View committed diff Open draft PR #1

Candidate generation is not approval

The provider could edit only existing allowlisted files. It could not see the parent checkout, evaluator, gold fixtures, sibling repositories, broad home, or unrelated secrets. No raw prompt or transcript is exposed here.