# Second Opinion

A paid micro-API for autonomous agents. Send a claim, plan, or draft output;
get a structured critique back before you act on it.

## Paid endpoint
POST /critique
Price: $0.19 USDC per call, settled via x402 on Base mainnet (eip155:8453).
Paid to: 0x0b9B8a32AD1E821af6E470A4De97ba7188A130a4

An unpaid request returns HTTP 402 with the payment requirements, including an
`extensions.bazaar` block describing the exact input and output shape. Sign and
retry with the `X-PAYMENT` header; a settled call returns HTTP 200 plus an
`X-PAYMENT-RESPONSE` header.

Request body (JSON):
  content  string  required  the text to critique
  context  string  optional  what the content is for (improves the critique)
  lenses   array   optional  any of: correctness, risk, completeness,
                             security, assumptions
                             (default: correctness, risk, completeness)

Response body (JSON):
  verdict          "pass" or "flagged"
  confidence       0-1, confidence the content is safe to act on as written
  issues           array of {lens, severity, note, suggestion, evidence?, question?}
                   severity is one of high, medium, low; sorted most severe first
  summary          one-line human/agent readable summary
  severity_counts  {high, medium, low}
  lenses_applied   array
  questions        array of specific things to resolve first
  engine           engine version string

## Free endpoints (no payment)
POST /sample          same engine, truncated to the top 2 issues,
                      20 requests/hour/IP. Use this to check output
                      quality before paying.
GET  /                service info
GET  /.well-known/x402 machine-readable manifest
GET  /openapi.json    OpenAPI schema
GET  /stats           settlement counters since last cold start

## Notes for calling agents
- The engine is deterministic and heuristic (engine: heuristic-v2): no model
  call, no per-call cost, no rate limit on the paid path, sub-100ms responses.
- Payment settles before any work is done, so a paid call cannot fail because
  of downstream model spend.
- Idempotent: same input, same output.