agent-decision-receipts – skill Agent & Tự động hóa cho agent AI

Agent Decision Receipts

Overview

A log says an action happened. A receipt is tamper-evident: it records who, what, and under which policy, and it is signed, so any later edit breaks the signature. This skill mints one for a consequential agent action and verifies it later from the certificate alone: no database, no network, no trusting the issuer.

The crypto is not in this skill. It is the open-source OpenAgentOntology receipt primitive (Apache-2.0), which signs every receipt with Ed25519 and the post-quantum legs ML-DSA-65 (FIPS 204) + SLH-DSA (FIPS 205) when the post-quantum backend is installed. This skill is the decision layer: when to mint, what to put in, how to verify. One install, no per-skill crypto.

Three decisions, nothing else:

1. Does this action need a receipt? — side-effecting + consequential + later-provable = yes.

2. Mint the receipt — build the action manifest, sign it with the OAO primitive.

3. Verify it — recompute the hash, check each signature leg, from the cert alone.

This skill is NOT log analysis. Logs describe what happened and can be silently edited. A receipt is minted before/at execution and breaks if edited. Use logs for debugging; use receipts for evidence.

This skill is NOT a hosted notary. It mints a LOCAL, self-signed receipt anyone can verify offline. Cross-organization verification (one org proving to another) is a separate hosted service, out of scope here.

This skill is NOT a legal opinion. It produces evidence shaped to support FRE 902(13)/(14)-style certification and EU AI Act Article 12 record-keeping. Whether a given receipt is admitted is a question for counsel.

Quick Start

# Install the open-source receipt primitive (Apache-2.0). Add [pq] for the post-quantum legs.
pip install "openagentontology[pq]"

# 1. Build + validate an action manifest (stdlib only, no crypto, no network)
python scripts/build_action_manifest.py --agent my-deploy-agent --operation deploy \
    --target prod/api --policy "EU AI Act Art 12" --out action.json

# 2. Mint the receipt over it (Ed25519 + post-quantum legs)
python -c "import json,openagentontology.receipt as r; \
  print(json.dumps(r.mint_receipt(json.load(open('action.json')), decision='ACTION_GOVERNED')))" > receipt.json

# 3. Verify from the cert alone (no DB, no network)
python -c "import json,openagentontology.receipt as r; \
  print(r.verify_receipt(json.load(open('receipt.json'))))"
# -> {'ok': True, 'sig_ok': True, ... 'reason': 'verified from the cert alone via: ed25519, ml_dsa, slh_dsa'}

> Dependency note. This skill delegates the signing to openagentontology (Apache-2.0, opt-in pip install). The script shipped here is stdlib-only and adds no repo dependency; the package is installed by the operator (BYO-library pattern). If it is not installed, the build step still works — only minting/verifying require it.

Core Workflow

The three decisions below are the skill: decide whether to receipt, mint, then verify.

Decision 1: Does this action need a receipt?

Mint a receipt when the action is all three of:

| Test | Mint if... |

|------|-----------|

| Side-effecting | it writes, sends, deploys, deletes, pays, grants access, or changes external state |

| Consequential | a wrong call costs money, breaks compliance, or harms a person |

| Later-provable | someone (auditor, insurer, regulator, court, counterparty) may ask "what did the agent do and why?" |

Read-only, reversible, trivial actions do not need a receipt. Receipt everything and the signal drowns; receipt nothing and the one call that mattered cannot be proven.

High-signal triggers (mint by default): deploy, delete, pay/wire/refund, grant_access, export/egress, approve/deny a claim, any model decision that affects a person under a high-risk AI system.

Decision 2: Mint the receipt

The action manifest is any ASCII-safe dict describing what the agent did. Four keys are required — build_action_manifest.py rejects the manifest (exit 2) if any is missing. Two more are added automatically:

| Key | Required? | What it carries |

|-----|-----------|-----------------|

| agent_id | required | the acting agent |

| operation | required | the verb (deploy / delete / pay / decide / ...) |

| target | required | what it acted on |

| policy | required | the rule that governs it (e.g. "EU AI Act Art 12", "internal change-control") |

| inputs_hash | auto-added | a hash of --inputs, so the full payload need not be stored in the clear (defaults to the hash of empty when --inputs is omitted) |

| decision_label | auto-added | the receipt decision label (defaults to ACTION_GOVERNED) |

mint_receipt(manifest, decision=...) hashes the full manifest into the receipt evidence, signs the canonical body, and returns a receipt that carries: evidence_hash, signature_b64 (Ed25519), and — when [pq] is installed — ml_dsa_signature_b64 + slh_dsa_signature_b64. Each leg signs the same bytes; any one verifying proves authenticity.

> See [references/receipt-fields.md](references/receipt-fields.md) for the full receipt schema and the post-quantum rationale.

Decision 3: Verify it

verify_receipt(receipt) recomputes sha256(canonical(evidence)), compares it to evidence_hash, then checks every signature leg it has a backend for. It returns {ok, hash_ok, sig_ok, legs, reason}. A single edited byte anywhere in the action breaks hash_ok; a forged signature breaks the leg. Verification needs only the receipt — no call back to the issuer.

This is the property that makes it evidence: a reviewer who distrusts the issuer can still confirm the receipt is intact and authentic, entirely offline.

Anti-Patterns

• Receipt the log, not the decision. Minting a receipt over a log line written after the fact proves nothing. Mint at the point of action, over the action.
• Storing the signing key next to the receipts. If the key is compromised, signatures mean nothing. Treat the key like any signing secret; never commit it.
• Ed25519-only when the post-quantum legs are available. A receipt is long-lived evidence. Sign it once with the post-quantum legs (ML-DSA-65 + SLH-DSA) so it stays verifiable if a future quantum computer could break Ed25519. Install [pq].
• Putting raw secrets or PII in the manifest. The manifest is hashed into evidence and is recoverable from the receipt. Carry hashes (inputs_hash), not the cleartext.
• Calling it "admissible." It is evidence shaped to *support* FRE 902(13)/(14)-style certification. Admissibility is a court's decision, not the tool's claim.
• Faking a signature when crypto is missing. The primitive emits an explicit unsigned flag instead. Never present an unsigned receipt as signed.

Cross-References

• ra-qm-team/skills/eu-ai-act-specialist/ — decide the AI system's risk tier and Article 12 obligations; this skill mints the per-action record those obligations require.
• ra-qm-team/skills/iso42001-specialist/ — the AI management-system controls; receipts are the per-decision evidence those controls call for.
• OpenAgentOntology (Apache-2.0): the open receipt primitive this skill drives — pip install "openagentontology[pq]".