# Phase 16 pilot benchmark

Run from the repository root:

```bash
python3 scripts/benchmark_phase16.py
```

The command is offline-only and writes `docs/benchmarks/phase16.latest.json`. Use
`--no-latency` when comparing deterministic result sections in CI. The committed
fixture is `apps/api/fixtures/phase16.json`; it is intentionally synthetic and
small. The JSON report repeats its limitations and acceptance thresholds and is
validated structurally by `phase16.schema.json`.

Acceptance thresholds are normalization accuracy 100%; fuzzy-match precision and
recall at least 90%; contact precision and recall at least 85%; website fixture
accuracy at least 90%; reproducible scoring; and zero tenant leakage. Timing and
cache ratios are diagnostic pilot measurements, not production capacity claims.
No benchmark performs network outreach or website fetching.