For frontier-lab release gates
One auditable result, signed and timestamped. Drop the receipt into your model card. Verifiable forever — no test-set leakage.
Cognitive evaluation · per-run · tamper-evident
The benchmark
that doesn't leak.
Most benchmarks stop measuring anything once they hit the open web. OCA keeps the challenges private, scores them server-side, and ships you a signed receipt you can publish without exposing the test set.
For small-model reasoning teams
A cheap competitive signal you can cite in a paper or product launch. Single probe from $50.
For academic labs
Reproducible private evals without writing IRB paperwork for a held-out test set you can't legally redistribute.
What a receipt looks like
Below is the shape of a sealed receipt. The signature is HMAC-SHA256 over the canonical-JSON payload and proves the result was scored against the held-out challenges we never expose.
{
"run_id": "9c1a4e…",
"customer_id": "acme-research",
"seed": 1738801,
"modules": ["RFNN", "HVM", "QRS", "CLA"],
"aggregated": {
"overall": 0.71,
"per_module": { "RFNN": 0.83, "HVM": 0.65, "QRS": 0.74, "CLA": 0.62 }
},
"scores": { ... },
"receipt_signature": "fa9c…(64 hex)…0d2b"
}