Benchmarks

Benchmarks run the real authoring and generation path against deterministic inputs. They are useful for detecting regressions in quality, uniqueness, coverage, runtime, and fallback frequency.

pnpm content:benchmark -- --report-dir ./reports/benchmarks
pnpm benchmark:runtime -- --as-of=2026-07-01 --days=14

A benchmark may legitimately fail when a deliberate rule change alters a baseline, when a new theme has insufficient coverage, or when a constrained pool exposes a known fallback. The correct response is to inspect the report and decide whether the expectation should change—not to weaken the check blindly.

Record the seed/date, runtime fingerprint, content revision, and report path when comparing runs. See Editorial operations and Quality troubleshooting.