Skip to content

Relation Governance Benchmark

The 1.2.0 relation-governance gate uses fresh SQLite stores containing only public synthetic fixtures:

npm run benchmark:governance

It runs the same report ten times against a sparse 50-memory store and a relation-dense 200-memory store with 2,000 explicit relations. The gate fails unless every run:

  • returns one or more candidate pairs;
  • has byte-identical JSON ordering;
  • contains no synthetic memory content;
  • leaves memory, relation, and audit-event counts unchanged; and
  • completes below a deliberately generous 1,000 ms p95 bound.

Use --json for machine-readable evidence. Latency is host-specific and is reported as min, p50, p95, and max; determinism, content-free output, unchanged state, and the upper bound are the portable pass/fail contract.

1.2.0 release-host evidence

The first release-host run on August 14, 2026 passed:

Profile Memories Explicit relations Candidates p50 p95
Sparse 50 0 25 84.76 ms 129.68 ms
Relation-dense 200 2,000 100 323.95 ms 342.34 ms

All ten runs per profile had byte-identical ordering, content-free output, and unchanged memory, relation, and audit-event counts. Latency remains descriptive; the final release run repeats the same gate.