# Public Eval Result Registry

This registry publishes public-safe eval result summaries. It distinguishes sample rows from reviewed operator-provided rows and excludes raw transcripts, secrets, live URLs, and copy-ready exploit payloads.

## Counts

- Runs: `1`
- Sample runs: `1`
- Reviewed operator runs: `0`
- Result rows: `5`

## Runs

| Run | Source | Model adapter | Defense adapter | Date | Case family | Verdict counts | Pass rate |
|---|---|---|---|---|---|---|---:|
| Offline harness sample | sample | `safe-offline` | `channel-isolation` | 2026-06-29 | Agent command execution or command-classification boundary, LLM-enabled agent security benchmark or survey surface, Prompt-bearing agent instruction boundary | pass: 5, fail: 0, inconclusive: 0, untested: 0, missing: 0 | 100.0% |

## Safety Boundary

- Operator-provided rows must use `approved_public_release` review status before publication.
- Registry generation fails on raw transcript fields, secrets, bearer tokens, SSH keys, and live URLs.
- Use `examples/eval-harness-results.sample.jsonl` for sample rows and replace it with reviewed operator rows only after scrub/sign-off.
- Use `scripts/ingest_eval_results.py` for operator-provided JSONL. Accepted rows must include `result_provenance`, `approved_public_release`, and any `redaction_report` entries created during ingestion.