AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

agent-decision-receipts

Use when an autonomous agent takes a side-effecting action (deploy, delete, pay, grant access) that may need proving later, or for EU AI Act Art. 12 record-keeping. Builds a tamper-evident signed receipt with build_action_manifest.py and verifies it.

Original source
alirezarezvani
License
MIT
Pinned revision
e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read alirezarezvani--agent-decision-receipts at revision e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 2

references/receipt-fields.md

Receipt fields + the post-quantum rationale

Detailed reference for the agent-decision-receipts skill. The receipt is produced by the OpenAgentOntology mint_receipt primitive (Apache-2.0); this documents what comes back and why.

The receipt schema

A minted receipt is an ASCII-only, JSON-safe dict:

| Field | Meaning | |-------|---------| | atom_id | stable id derived from the action source + the evidence-hash prefix | | type | AgentGovernanceReceipt | | decision | the verdict label passed in (e.g. ACTION_GOVERNED, DEPLOY_APPROVED) | | evidence_hash | sha256(canonical(evidence)) — the commitment to the full action manifest | | signed_at | ISO-8601 timestamp (in the signed body, never in the hashed evidence) | | evidence | {summary, ontology} — the full action manifest is under evidence.ontology | | signature_b64 | Ed25519 signature over canonical(body) | | verify_pubkey_b64 | the Ed25519 public key, so the receipt verifies offline | | signed | bool — false + an explicit flag if the crypto backend is absent (never faked) | | ml_dsa_signature_b64 | ML-DSA-65 (FIPS 204) leg, present when [pq] is installed | | slh_dsa_signature_b64 | SLH-DSA (FIPS 205) leg, present when [pq] is installed | | signature_alg | names every leg actually on the receipt | | kid | sha256(verify_pubkey_b64)[:32] — 128-bit key identifier. Lets a verifier comparing two receipts answer "signed by the same key?" offline, without any registry. |

Canonicalization (why the hash reproduces anywhere)

canonical(obj) = json.dumps(obj, sort_keys=True, separators=(",",":"), ensure_ascii=True, allow_nan=False).

Because the evidence is sorted, separator-fixed, ASCII, and finite, a verifier in any language recomputes the identical sha256 byte-for-byte. There are no timestamps inside the hashed evidence, so the hash is deterministic; the timestamp lives in the signed body instead.

Verification result

verify_receipt(receipt) returns:

| Field | Meaning | |-------|---------| | hash_ok | recomputed sha256(canonical(evidence)) equals evidence_hash | | legs | per-leg status: ok / fail (tamper) / absent / unverifiable (no backend) | | sig_ok | at least one signature leg verified | | ok | hash_ok AND no leg failed AND (sig_ok OR the receipt is honestly unsigned) | | reason | human-readable verdict |

A one-byte edit to the action breaks hash_ok. A forged signature flips a leg to fail. An older Ed25519-only receipt still verifies; the post-quantum legs are simply absent.

Why post-quantum, by default

A receipt is long-lived evidence — an insurer, a regulator, or an auditor may verify it years after the action. A cryptographically-relevant quantum computer could forge Ed25519 signatures, so a signature that must stay trustworthy for years needs an algorithm that resists that. The post-quantum legs are NIST-standardized signatures designed for exactly this:

  • ML-DSA-65 (FIPS 204) — lattice-based; the size/performance default for hot paths.
  • SLH-DSA (FIPS 205) — hash-based, stateless; smaller assumptions, good for long archival.

Each leg signs the same canonical body, so a verifier with only one backend can still prove authenticity from whichever leg it can check. This is post-quantum cryptography (quantum- resistant signatures on classical computers), not quantum computing — label it precisely.

Install both legs: pip install "openagentontology[pq]".

PQ-required verification (strict mode, recommended)

A verifier that accepts Ed25519-only receipts can be fooled by an attacker who strips the post-quantum legs from a hybrid-signed receipt — the residual Ed25519 leg still verifies, but the long-lived guarantee is gone. For long-lived evidence, require the post-quantum legs to be present and valid, and only accept Ed25519-only receipts when you must verify legacy archival data:

import os
import openagentontology.receipt as oao

def verify_strict(receipt, *, require_pq: bool | None = None):
    must = require_pq if require_pq is not None else (
        os.environ.get("RECEIPT_REQUIRE_PQ", "true").lower() in ("1", "true", "yes", "on"))
    out = oao.verify_receipt(receipt)
    if must and out.get("ok") and out.get("signed"):
        legs = out.get("legs", {})
        missing = [n for n in ("ml_dsa", "slh_dsa") if legs.get(n) != "ok"]
        if missing:
            out["ok"] = False
            out["reason"] = f"PQ-required: missing {missing}"
    return out

Default the gate to on; set it off only for legacy Ed25519-only archival receipts.