{"schema":"sentinel.evaluation.v1","role":"ERC-8183 Evaluator (off-chain, signed verdict)","chain_default":"eip155:8453","verifies":"provider behavioral reliability (payment/delivery track record)","does_not_verify":"correctness of the specific deliverable","scope":"This verdict evaluates the behavioral reliability of the provider (payment and delivery track record on record with SENTINEL), not the correctness of the specific deliverable. It answers 'is this provider trustworthy?', not 'is this work correct?'.","reason_hash":"keccak256 over JCS-canonical evidence block; use directly as bytes32 reason in complete()/reject()","signature":"ES256 JWS, verifiable offline against /.well-known/jwks.json","methodology":"https://sentinel-agent.dev/methodology","payment":{"required":true,"scheme":"x402","base_price_usdc":"0.020","note":"Scales with declared escrow_value_usd (same tiers as /v1/guard, x4). The underlying score is available free via /v1/attestation; this endpoint charges for the signed, threshold-applied verdict and reason_hash ready to submit on-chain."},"usage":"GET /v1/evaluation?jobId=..&provider=0x..&chain=..&escrow_value_usd=.. → pay x402, verify jws offline, then call complete(jobId, reason_hash) or reject(jobId, reason_hash) on the ERC-8183 AgenticCommerce contract. SENTINEL does not move funds.","verdicts":{"complete":"provider trustworthy (score >= threshold, no verified incidents)","reject":"below threshold, has a verified incident, or insufficient evidence (fail-closed)"},"threshold":{"requested_default":60,"note":"The pass bar tightens with declared escrow_value_usd, not just the price (>=1,000 -> +5, >=10,000 -> +10, >=100,000 -> +15, capped at 100). Rationale: research on long-horizon coding-agent evaluation (SpecBench, arXiv:2605.21384, 2026) finds the gap between passing validation and actually satisfying the spec widens as task complexity/stakes grow -- a provider just above threshold is a reasonable bet for a small job and a worse one for a large job, even at an identical score. When escrow_value_usd is supplied, the response includes both requested_threshold and the effective threshold actually applied inside the signed verdict, so the adjustment is disclosed and recomputable, never silent."}}