# SENTINEL — Pre-Transaction Safety Oracle for AI Agents
The trust layer for agent-to-agent commerce: before an autonomous agent moves money, SENTINEL answers "is it safe to proceed?" in two dimensions — the transaction and the counterparty. Reference implementation of the x402 trust-provider extension (issue #2299). This is sentinel-agent.dev (on-chain oracle 0x6841496c3e7eDF9eEB02bE64ab98CF5D5c5aF813), distinct from unrelated projects sharing the SENTINEL name.

## Capability 1 — Counterparty trust (0-100 score, signed, on-chain)
Is this payment counterparty trustworthy? Behavioral scoring of x402 buyers and sellers on Base, backed by an append-only Default Registry of verified incidents.
  GET  /v1/attestation?subject=0xADDR&direction=buyer|counterparty  -> free cached ES256 attestation (JWS)
  GET  /v1/attestation/deep?subject=0xADDR                          -> paid fresh attestation ($0.03 USDC, x402)
  GET  /.well-known/jwks.json   -> public key for offline verification (ES256, kid sentinel-2026-01)
  GET  /defaults               -> append-only, hash-anchored incident registry
  On-chain: SentinelOracle 0x6841496c3e7eDF9eEB02bE64ab98CF5D5c5aF813 (Base) — free getScore/meetsThreshold
  Methodology: /methodology (public v1.0.0) · /methodology.json

## Capability 2 — Transaction safety (pre-execution verdict, signed)
Is this specific on-chain transaction safe to sign? Contract-level risk check before execution.
  POST /v1/guard
    Input:  {"chain": "base", "from": "0xAGENT", "tx": {"to": "0xCONTRACT", "data": "0x", "value": "0x0"}}
    Output: {"verdict": "SAFE|UNSAFE|UNKNOWN", "risks": [...], "reasons": [...], "signature": "...", "txDigest": "sha256:..."}
    Payment: $0.005 USDC per call (x402, EIP-3009, facilitator-free).
    Risk data: GoPlus token security (honeypot, owner abuse, taxes, mintable, proxy) + Alchemy eth_call simulation (reverts).

## Discovery
  - Agent card: /.well-known/agent.json
  - For agents (decision & integration guide): /for-agents (human) and /for-agents.txt (machine)
  - Pricing: /pricing

## Methodology
  - Public methodology v1.0.0 (how SENTINEL judges): /methodology
  - Machine-readable: /methodology.json

## Compliance
  - Regulatory mapping (IMDA, EU AI Act Art.12, NIST): /compliance

## Writing (positioning)
  - Whitepaper: Recomputable Trust (why behavioral evidence beats reputation): /whitepaper
  - Identity Is Not Trust: /writing/identity-is-not-trust
  - The Delivery Gap: /writing/the-delivery-gap
  - The 86% Problem: /writing/the-86-percent-problem

## Siblings
  VeraData (verified LATAM data), Intelica (competitive intel), TrustBoost (PII sanitization).