Public read API / codesota.v1

Evidence that agents can read.

Public HTML is rendered on the server. No JavaScript, account or API key is required for data reads.

Start with the discovery manifest → · OpenAPI 3.1 → · Agent guide →

Find models and evidence

curl 'https://beta.codesota.com/api/v1/benchmarks?preset=liquid-reference&limit=100'
curl 'https://beta.codesota.com/api/v1/models?limit=100'
curl 'https://beta.codesota.com/api/v1/results?current=true&limit=100'
curl 'https://beta.codesota.com/api/v1/queue/eta'

List responses contain api_version, data, pagination and generated_at. Follow pagination.next until null. Default limit 20; maximum 100. Offset starts at zero. Invalid parameters return 400; missing records return 404.

Resources

• Tasks: canonical inputs → outputs.
• Evaluation profiles: executable registry entries and pinned configurations.
• Benchmarks: 50 selection plans, distinct existing benchmarks and dated BPB snapshots. Filter by preset, family, status or q.
• Model versions: exact revisions and known parameter counts.
• Protocols: scoring rules and comparison scope IDs.
• Results: origin, coverage, metric, scope, model and immutable evidence links. Filter by model or scope_id. Current scopes are the default; current=false includes history.
• Queue, ETA, traffic, factory, batching profiles and serial validation.

Follow record IDs from lists to /api/v1/{resource}/{id}. IDs distinguish model revisions, protocols and benchmark variants.

Read the qualifiers

Compare identical scope_id and metric only. CodeSOTA origin does not imply full-benchmark coverage. Current measurements are subsets. A planned catalogue entry has no measured score or cost. Raw-base loss is separate from instruct models and the future fixed-SFT track. Vendor benchmark names do not imply matching vendor protocols. No global aggregate SOTA score is published.

Measurement timestamps use Unix seconds; generated_at is ISO 8601 UTC. GPU-slot time includes reserved evaluator wall time, loading and CPU work; it is not GPU kernel-active time. Queue estimates assume available GPUs and exclude adapter development or new arrivals.

Access

These endpoints support public GET reads and cross-origin browser access. Worker mutations require private credentials; accounts and execution traces are separate protected services. Existing unversioned read APIs remain available.