Panel · adversarial
RV — Reasoning Verification
Adversarial multi-model reasoning verification. Independent models evaluate, a red-team critic challenges the case, and a synthesizer returns ALLOW, BLOCK, or UNCERTAIN — with signed attestation when you need a receipt, not a vibe.
Call example
curl -X POST https://api.thoughtproof.ai/v1/verify \
-H "Content-Type: application/json" \
-d '{"claim":"Approve €500 payment to vendor-42 based on invoice #4421"}'Unpaid calls return 402 with the x402 payment descriptor. Fast / Standard / Deep: RV on /pricing/.
When to use it
- Consequential decision points Escalate here when the cost of a wrong ALLOW is high. Do not loop RV on every micro-step.
- Adversarial independence Multiple models and a critic — not a single oracle agreeing with itself.
- Attestable outcomes When downstream systems need a signed verification artifact, not just a UI badge.
Composition pattern
Common pattern: gate every step with Sentinel, escalate only the high-stakes decisions to RV. Verdicts merge conservatively — BLOCK > UNCERTAIN > ALLOW — so a clean result from one layer never overrides a stop from another.
Related
Plan/trace cascade: PLV. Five-axis scoring app: DQL. Behavior / latency: behavior reference.