ReportCheck

Submit an engineering report (API 510 pressure vessel, API 570 piping, or PSM 1910.119 compliance audit) as JSON claims tagged with code + section. ReportCheck runs a deterministic paragraph-match against the canonical citation index and returns a structured audit response identifying every missing_citation, wrong_paragraph, and unit_mismatch — the three most common engineering-report hallucinations produced by general-purpose LLMs.

• Report text is held in process memory only. Retention < 60 seconds; never written to disk.
• NIM-only path. Customer report content is never streamed to OpenRouter, Sonar, or any external LLM.
• Only the response, audit ID, and ledger metadata are persisted — not the input report.
• Source repository is private.

First call returns 402 Payment Required with the x402 accepts envelope. Second call with X-PAYMENT returns 200 and the audit JSON.

POST /api/audit
Content-Type: application/json
X-PAYMENT: <base64 payment proof>

{
  "report_type": "API_510",
  "vessel_id": "V-101",
  "claims": [
    { "id": "c1",
      "claim_text": "MAWP shall not exceed nameplate rating...",
      "code": "API 510", "section": "5.4.1",
      "value": "150", "unit": "PSI" }
  ]
}

{
  "report_type": "API_510",
  "findings": [ { "error_type": "missing_citation" | "wrong_paragraph" | "unit_mismatch" | null, ... } ],
  "citations_checked": 12,
  "hallucinations_detected": [...],
  "unit_errors": [...],
  "paragraph_mismatches": [
    { "cited_paragraph_text": "...", "expected_paragraph_text": "...", "jaccard_score": 0.04 }
  ],
  "confidence_score": 67,
  "privacy": { "mode": "NIM_ONLY", "retention_seconds": 60 }
}

Hit the live endpoint with a standard x402 payment header. First call returns 402 with the accepts envelope; pay $49 USDC on Base and replay with X-PAYMENT to get the audit JSON. Settlement is verified + settled through the Coinbase CDP facilitator (Base mainnet, 2s finality).

# 1. Discover the payment requirement (returns 402 with x402 accepts envelope)
curl -s -X POST https://reportcheck.adametherzlab.com/api/audit

# 2. Pay $49 USDC on Base via any x402 client (e.g. @x402/hono client,
#    Coinbase Agentic Wallet Skills, ag402, x402-pay), then replay with the proof:
curl -s -X POST https://reportcheck.adametherzlab.com/api/audit \
  -H 'content-type: application/json' \
  -H 'X-PAYMENT: <base64-encoded x402 payment payload>' \
  --data-binary @your-report.json

Sandbox (no payment, fixture-only) available on request: adametherzlab@gmail.com. Real customer report text is never used for the sandbox — only synthetic fixtures.

Block below is fetched from /sample-audit.json — a static deterministic snapshot of one real audit run against sample-api510-with-3-known-errors.json. The audit engine is deterministic, so re-running the same fixture always produces these same findings. No live free-tier endpoint; nothing to abuse.

Loading /sample-audit.json …

{
  "report_type": "API_510",
  "citations_checked": 12,
  "hallucinations_detected": [
    { "claim_id": "c7",  "code": "API 510", "section": "8.1.3",
      "reason": "Claim text overlap with canonical paragraph is 0 (threshold 0.15). The report cited API 510 8.1.3 but the prose does not match the canonical content." },
    { "claim_id": "c10", "code": "API 510", "section": "5.6",     "reason": "..." },
    { "claim_id": "c12", "code": "API 510", "section": "99.99.99","reason": "Code/section API 510 99.99.99 not found in canonical citation index." }
  ],
  "unit_errors": [
    { "claim_id": "c1", "code": "API 510", "section": "5.4.1",
      "detail": "Report uses unit \"BAR\" but canonical section 5.4.1 expects one of: PSI, PSIG, kPa, MPa." }
  ],
  "paragraph_mismatches": [
    { "claim_id": "c7", "code": "API 510", "section": "8.1.3", "jaccard_score": 0,
      "cited_paragraph_text":    "Garden tomatoes require full sun exposure of at least six hours daily...",
      "expected_paragraph_text": "All repairs and alterations to pressure vessels shall be performed by an organization holding a valid National Board R Certificate..." }
  ],
  "confidence_score": 67,
  "privacy": { "mode": "NIM_ONLY", "retention_seconds": 60,
               "report_text_persisted": false, "external_llm_used": false }
}