# Subcento

> Pre-flight transaction verification for AI agents. Pay with x402 (0.002 USDC on
> Base mainnet), get a signed verdict — SAFE | CAUTION | BLOCK — with evidence in
> /regime under 150 ms; /check ~0.3 s warm / ~1 s cold (GoPlus + Jev run live per
> request — only GoPlus is cached). Fail-closed: if trust layers are unreachable you
> get BLOCK, never a false SAFE.

## Why an agent should call this before signing
Four layers: address risk (GoPlus), OFAC blocklist, semantic intent analysis of the
calldata (Jev / TypeSafe AI), Ed25519-signed verdict verifiable with `verify_key`.

## Endpoints

### POST /check (x402 paywall — returns HTTP 402 + Payment-Required if unpaid)
Body: `{"to": "0x…", "calldata": "0x…", "value": 0, "chain_id": 8453, "intent": "optional declared intent"}`
x402 terms: scheme exact, network eip155:8453, asset USDC, amount 2000 (= $0.002),
facilitator https://facilitator.goplausible.xyz, maxTimeout 60s.

### GET / — free machine-readable discovery card
### GET /health — free liveness + verify_key

### POST /regime (x402 paywall — 0.0001 USDC)
Body: `{}` or `{"live_jev": true}` to refresh the Jev classification now.
Returns a signed market-stress verdict: deterministic cross-asset seismograph
score (0-100, SAFE <45 ≤ CAUTION <75 ≤ BLOCK), Jev regime label
(calm|volatile|stress|crisis) and 48h escalation noul, snapshot freshness
(`age_seconds`, always <1800). x402 terms: network eip155:8453, amount 100
(micro-USDC = $0.0001), same facilitator. This is aggregate market stress —
not a price prediction, not advice.

## Docs
- Root: https://subcento.alterfield.com/ (full OpenAPI-ish discovery JSON)

## Notes
- Verdicts include `latency_ms`, `reasons[]`, `signature`, `verify_key`.
- `intent` mismatch alone ⇒ CAUTION (anti false-positive), drainer probability > 0.7 ⇒ BLOCK.
- v1.2: send the ERC-3009 `nonce` you intend to spend and the signed verdict is `nonce-bound`
  (opens against exactly one settlement; a shape-bound verdict never upgrades).
- `evaluation_status: not-measured` on a BLOCK means a trust layer could not run — gate
  conservatism, not evidence. Never cite it as a negative signal about the agent.
- Signature covers: verdict, to, function, value, reasons, timestamp, chain_id, binding,
  evaluation_status, checks_failed, limits. Everything else in the response is advisory detail.