Open world-state outcome-verification + portable reputation for inter-agent calls.
When one agent hires another (over A2A/MCP), Warrant independently checks the world state to prove the work actually happened, then issues a signed warrant (intent → claimed outcome → independent verification → verdict → signature). Verified warrants roll into a portable reputation the agent carries everywhere.
To warrant something = to certify it's true. No claim without a warrant. Your agent says it booked the flight — did it? Get the warrant.
Live board: https://warrant-gold.vercel.app · Playground: https://warrant-gold.vercel.app/playground · npm: npm i @avee1234/warrant-verify
Every inter-agent standard today (MCP, A2A, AGNTCY, AP2, ERC-8004) answers "is this agent who it claims, and was it authorized?" — none answers "did the agent actually do what it claimed?" And the whole field verifies the trace (the agent's self-reported log — exactly what a silently-failing agent lies in), not the world-state outcome. That distinction is the moat.
- Non-crypto, enterprise-friendly (ERC-8004 owns the on-chain path).
- An A2A extension, not a new transport (A2A/MCP own transport).
- Outcome, not trace — verify against ground truth, not the self-reported log.
| Layer | Source |
|---|---|
| Evidence / format substrate | OpenTrajectory (shipped) — a warrant's body is an OpenTrajectory record |
| Verifier engine | TrueCall — runtime outcome / post-condition checking |
| Verification brain | RL Trajectory Auditor / Inspector (shipped, validated judge) — reused as-is |
| New on top | Warrant schema · portable verified reputation · public reputation board |
v1 (spec + verifier + demo + board) and v2 (cross-harness portable reputation) shipped;
v3 in progress (real probes, hosted registry, live A2A, SDK/DX). See STATUS.md
and docs/quickstart.md. Full spec + hard constraints in CLAUDE.md.
Live: board https://warrant-gold.vercel.app and playground https://warrant-gold.vercel.app/playground (verify any warrant in your browser).
cd packages/verify && npm test # zero-dep verifier, 29/29
cd packages/demo && npm run github # REAL verification against the live GitHub API
cd packages/demo && npm run judge # judged verification of a FUZZY outcome
cd packages/demo && npm run demo # live "did it book the flight?" clip
cd packages/demo && npm run cross-harness# one agent, 2 harnesses -> one portable reputation
cd packages/demo && npm run a2a # full HTTP end-to-end: A2A + verifier + registry
cd packages/demo && npm run board # -> site/index.html (public reputation board)
cd packages/demo && npm run playground # -> site/playground.html (in-browser warrant verifier)
cd packages/registry && npm test # registry trust gate (valid accepted / forged rejected)docs/ # gap-analysis, warrant-spec, positioning, trust-model, reputation-integrity, quickstart, deploy
schema/ # warrant-0.schema.json (machine-readable spec companion)
examples/ # hand-authored warrants (green + red) + a2a-extension manifest
packages/verify/ # zero-dep verifier — world-state probes, verdict engine, signing, schema, reputation, CLI
packages/demo/ # two-agent A2A demo (one lies), cross-harness, live A2A end-to-end, board generator
packages/registry/ # optional hosted reputation registry (re-verifies + sig-checks every warrant)
site/ # generated public reputation board (index.html, openable via file://)