Every call below costs less than the tokens your agent would burn doing the same work itself — and ours is verified, not improvised. Don't trust us, check us: the assumptions are stated, the math is yours to redo.
| Endpoint | Our price | Self-serve* | What self-serve burns | What you get here |
|---|---|---|---|---|
| /judge-batch | $0.01/series | ~$0.05/series | 1,000 backtests validated one-by-one in tokens: 1,000 round trips of selection math in prose, errors compounding. | One call triages up to 1,000 series, survivors ranked best-first. Feeds judge-lite for the full verdict. |
| /judge-nano | $0.01 | ~$0.05 | 2–4k output tokens reasoning through the selection correction, with a real chance of botching the arithmetic. | Deterministic pure-Python math. An LLM doing statistics in prose can hallucinate the formula; this can't. |
| /judge-lite | $0.05 | ~$0.15 | Three statistical gates reasoned through in tokens, ~8–12k output. Errors compound across gates. | Three gates, one call, one verdict — selection, deflated Sharpe, clustering — with killed_by reasons. |
| /factcheck | $0.03 | ~$0.35 | 3–4 page fetches at ~20k tokens each plus comparison reasoning. Four-figure token burn before the answer. | We fetch the sources, extract cited passages, and report agreement — verdicts include insufficient_evidence. |
| /v1/extract | $0.02 | ~$0.10 | One page fetch (~20k tokens) plus extraction reasoning. Three-method cross-check would be ~3x that. | Three independent extraction methods cross-checked, confidence-scored, tamper-evident attestation. |
* Assumptions, so you can check the math: frontier-class model (~$3/1M input, ~$15/1M output tokens), doing the work carefully — fetching, cross-checking, verifying — in a single attempt. A budget model doing a sloppy single pass costs less in tokens; what you're buying here is the verification, not just the answer.
How to buy: no account, no key, no invoice. POST the endpoint, receive HTTP 402 with the payment requirements, pay USDC on Base (x402 v2), retry with the payment signature. Machine-readable manifest at /.well-known/x402, full specs at /shop, plaintext catalogue at /llms.txt.