# CodeSOTA beta

> Independent, inspectable AI evaluation pilots and a planning catalogue for small models.

All public pages are server-rendered HTML. No JavaScript, account or API key is needed to read benchmark data. Sign-in is optional. Private traces, credentials and worker writes are not public APIs.

## Discover
- [Public API manifest](https://beta.codesota.com/api/v1/manifest)
- [OpenAPI 3.1 specification](https://beta.codesota.com/openapi.json)
- [API documentation](https://beta.codesota.com/api/docs)
- [Under-1B leaderboard and 50-entry catalogue](https://beta.codesota.com/sub-1b)
- [Ontology and evaluation loop](https://beta.codesota.com/ontology)
- [Runnable queue and conditional ETA](https://beta.codesota.com/queue)

## Read
- /api/v1/tasks
- /api/v1/evaluation-profiles
- /api/v1/benchmarks?preset=liquid-reference&limit=100
- /api/v1/models
- /api/v1/protocols
- /api/v1/results?current=true&limit=100
- /api/v1/queue/eta

List responses include pagination links; follow next until null. Default limit 20, maximum 100. Read endpoints support GET only. The OpenAPI document specifies filters and detail routes.

## Interpretation
Compare identical scope_id and metric only. All current results are CodeSOTA-measured subsets, not full benchmark rankings. Planned entries and planning cost labels are unmeasured. Base/adapted raw-text loss is separate from instruct models and from the future fixed-SFT track. Source manifests and item-level evidence are linked from each result. Fresh text is not proof of contamination immunity or practical usefulness. GPU-slot wall time is not GPU kernel-active time. Do not infer global SOTA from these pilots.

Agent usage and sanitized traces: /agents and /api/v1/agents. Hourly UTC token counts include cache categories; raw prompts and tool payloads remain private.