Glossary
- BYOK: Bring‑your‑own‑keys for provider credentials; stored encrypted per org.
- Run: A single LLM request/response with metrics (tokens, cost, latency, status).
- Experiment: A group of runs tied to a prompt version and optional dataset.
- Prompt version: A specific prompt template (and optional system) under a prompt.
- Matrix sweep: Running combinations of models/temperatures/params.
- Gate: A rule on metrics (e.g., error rate ≤ X, cost ≤ Y, score ≥ Z).
- p95: 95th percentile; used for latency.
- KPI: Key performance indicator (cost, error rate, p95 latency).
- Hash‑only: Storing only prompt hashes, not plaintext previews.