A benchmark gate that catches small regressions
How a seeded 50,000-row corpus, a recall baseline and a p95 latency baseline turned a search benchmark into a pass or fail check that runs on every pull request.
Posts tagged sqlite
Local-first memory for AI agents: system design
How Aletheia is built: hooks designed to fail open, redaction before storage, one SQLite file, a leased job queue and a bounded session-start block. Choices and trade-offs.