# Held-out retrieval and evidence-span audit

Run: heldout-retrieval-audit-20261028T141818Z
Corpus: 52 installed public system-documentation files; no user/project data.
Queries: 21 held-out paraphrases, evaluated against the previous fixed source-based qrels.
Independent top-five order agreement: 42/42
Independent metric agreement: 42/42
Citation/source-span checks: 210; failures: 0
No-evidence controls: 5; failures: 0
Valid quote accepted: True
Fabricated quote rejected: True
Unknown citation rejected: True
Automated gate: PASS

## Held-out paraphrase hit@5

| Ranker | Baseline hits | Paraphrase hits | Minimum required | Gate |
|---|---:|---:|---:|---|
| BM25 | 20/21 | 20/21 | 18/21 | PASS |
| TF-IDF | 18/21 | 19/21 | 16/21 | PASS |

## Interpretation boundary

PASS only confirms the stated test contracts on this small local corpus. Citation span checks validate source provenance and quote integrity, not entailment of all possible generated claims. The no-evidence controls intentionally use zero-overlap vocabulary and do not certify broad out-of-domain detection.

No production default, service, index, or model weights were changed. No network access or package installation was used.