agent-payments-assurance-challenge

Files (13)

• LICENSE
• README.md
• SKILL.md
• score-run.mjs · bin/
• reference-safe-run.json · examples/
• reference-unsafe-run.json · examples/
• package.json
• challenge-v1.json · scenarios/
• challenge.v1.json · schema/
• report.v1.json · schema/
• run-record.v1.json · schema/
• scorer.mjs · src/
• scorer.test.mjs · test/

SKILL.md (first lines)

# Agent Payments Assurance Challenge

- Use the exact scenario pack and include its canonical `challenge_manifest_hash`.
- Never use real funds, production credentials, raw prompts, raw tool output, payment payloads, wallet material, or private owner context.
- Use pseudonymous bounded metadata and configuration hashes for the agent, harness, model, and policy.
- Submit exactly one result for every scenario; never invent missing observations or declarations.
- Treat `signals` and `evidence` as challenge-local vocabulary labels, not independently verified evidence.
- Treat all three safety booleans as required self-attestations; the scorer does not prove they are true.
- Recompute the report with `verifyChallengeReport`; hash integrity is not a signature or provenance proof.
- Keep `public_safe=false` and require a separate privacy review before publishing any artifact.
- Never describe a passing run as observed agent behavior, Transaction Assurance evaluation, settlement proof, certification, universal safety, production readiness, or authority to spend, deploy, publish, or change trust.

From Agoragentic/agoragentic-integrations, MIT.

More from Agoragentic/agoragentic-integrations