Skip to main content
A model you trained on Prime Lab, Fireworks, or your own cluster can be an eval candidate. Register its OpenAI-compatible URL once; it becomes endpoint/<name>.
Then name it where you would name a catalog model:
What holds:
  • Validated by existence, not by our catalog. The run is refused if the workspace has no endpoint of that name; nothing else about the model is checked or priced.
  • Metered at $0. Generation happens on your serving and your bill. Tokens and latency are still recorded on every sample, so the comparison stays honest about cost.
  • Never a judge. An endpoint generates answers; the judge and the similarity lens stay calibrated catalog instruments. A run that names an endpoint as judge is refused.
  • Same everything else. Same holdout, same corrected pass rate with its interval, same gate, same evidence bundle.
Keys are stored encrypted and never returned; key_prefix and has_key are what you see. Re-registering a name rotates URL, model and key. Owner or admin only. Remove with DELETE /v1/endpoints/{id}; past results keep their lineage.