a2a cloud
eu ai act · timeline checked 11 july 2026

Evidence is a query, not a quarter.

Article 12 asks for automatic, lifetime event records. Article 19 asks you to keep them six months. This runtime already signs receipts for Agent API, public /invoke, and standard MCP tools/call executions; authorization decision — compliance is querying records you already have, not instrumenting code you haven't written.

receipts for ai systems · explicit authority · no ambient trust

0
day retention floor
0
audit surfaces, one query
0
receipt / invocation
current stand-alone high-risk timeline
2027-12-02T00:00:00Z
··days··hrs··min··sec

Commission guidance currently lists 2 December 2027 for certain stand-alone high-risk systems and 2 August 2028 for systems embedded in regulated products. Confirm your category and the law in force.

what the act asks · what the runtime emits

Four articles. Four primitives. Zero instrumentation.

Article 12Record-keeping

asks · High-risk AI systems must technically allow automatic recording of events over their lifetime.

runtime · Agent API, public /invoke, and standard MCP tools/call executions seal Ed25519 receipts: caller, skill, input evidence, verified grants, outcome, and timing.

Article 19Log retention

asks · Provider-controlled automatic logs are kept for a purpose-appropriate period of at least six months, unless other applicable law provides otherwise.

runtime · Retention policy floored at 180 days while the EU AI Act framework is active — enforced at write time and re-checked inside the purge path. Legal hold blocks deletion outright.

Article 14Human oversight

asks · Systems are designed so natural persons can effectively oversee them.

runtime · Platform-managed approval and decision workflows record who decided—auto, policy, or human. Those records remain separate from execution receipts unless explicitly correlated.

Article 11Technical documentation

asks · Documentation is drawn up before market placement and kept current.

runtime · Agent card, version, image digest, risk tier, intended purpose, and oversight statement export as a per-system dossier inside the evidence pack.

decision records

One query across every decision your agents made.

Skill executions, authorization decisions, and administrative actions normalize into a single record shape. Filter by agent, outcome, kind, or date range. Open a record and the platform re-verifies its Ed25519 signature in front of you — proof, not a promise.

No second store to sync, no ETL to a compliance warehouse. The records the auditor reads are the records the runtime wrote.

See replay & receipts →
decision record · shape
example
{
  "record_id": "rcpt_<id>",
  "kind": "skill_execution",
  "agent_name": "<agent>",
  "actor": "user:<caller>",
  "action": "skill:<skill>",
  "outcome": "ok",
  "occurred_at": "<iso-8601>",
  "verifiable": true,
  "signature_status": "valid",
  "details": {
    "grant_ids": ["gr_<id>"],
    "tool_calls": 3,
    "eval_score": null
  }
}

Field shape only — illustrative, not a real customer run. Records the platform sealed are served from the API and rendered at /replay.

one export · three frameworks

The evidence pack answers the questionnaire.

A single JSON export: decision records, agent dossiers, signature verification results, and the control mapping inlined — an auditor who has never seen this platform can read it cold.

EU AI Act

Articles 11, 12, 14, 19

Queryable decision records, 180-day floored retention, human-oversight attribution, and per-system documentation. The Commission's current timeline distinguishes December 2027 stand-alone and August 2028 product-embedded high-risk systems.

NIST AI RMF 1.1

Govern · Map · Measure · Manage

Org policy with named owners, agent inventory with risk tiers, separately recorded eval scores, scoped authority and budget caps — mapped function by function in the export.

ISO/IEC 42001

Procurement dossier

The evidence pack answers the AI-management-system questionnaire your buyer's procurement team sends: traceability, operational control, performance evaluation.

how it works

Governed agent execution creates the evidence layer.

01

Run

Supported delegation and workspace operations can execute under scoped, signed grants. Agent-local filesystem, network, database, and configured-secret access use separate controls.

02

Record

The control plane seals authenticated Agent API calls and trusted public /invoke or standard MCP tools/call executions into signed receipts. Zero agent-code instrumentation.

03

Query

Decision records are one API: filter by agent, outcome, kind, date range. Signature re-verified when you open a record.

04

Export

One click produces a self-describing evidence pack — records, agent dossiers, control mapping — ready for auditors and procurement.

build the evidence before the deadline

Deploy governed. Query the evidence.

Retention windows only protect the records that exist. Every day agents run ungoverned is a day of evidence you can't produce later.