Record-keeping and oversight: mapping to EU AI Act Art. 12 / 26¶
Not legal advice
This page is an architecture mapping, not legal advice and not a claim of compliance. Whether a particular system is a "high-risk AI system", who is its provider or deployer, and what its logs must contain are legal questions that depend on the system, its intended purpose and its jurisdiction. Note also that Regulation (EU) 2024/1689 does not define "agent": it regulates AI systems, and how an agentic deployment decomposes into systems, providers and deployers is for counsel to decide. What follows only shows which Lodestar artifacts have the shape the provisions describe.
The provisions, briefly¶
Wording below is from Regulation (EU) 2024/1689 as published in the Official Journal (EUR-Lex).
- Article 12(1): "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system."
- Article 12(2): the logging capabilities must enable recording of events relevant for identifying risk situations and substantial modifications, facilitating post-market monitoring, and monitoring the operation of the system by deployers.
- Article 12(3): minimum content (period of each use, reference database, matched input data, identification of verifiers) is specified for the biometric systems in Annex III point 1(a); it does not prescribe a schema for other systems.
- Article 14(4)(d)–(e): the system must let its overseers decide not to use it or to disregard or reverse its output, and to "intervene in the operation of the high-risk AI system or interrupt the system through a 'stop' button".
- Article 26(2): deployers "shall assign human oversight to natural persons who have the necessary competence, training and authority" (plus the necessary support).
- Article 26(6): deployers keep the logs automatically generated by the high-risk system, to the extent under their control, for a period "appropriate to the intended purpose" and at least six months, unless other Union or national law provides otherwise.
Read together: the system must record, the human must be able to intervene, and the deployer must keep the record. Article 12 says what the system must be capable of; it does not say a given product "satisfies" it.
Timeline¶
Regulation (EU) 2026/1744 (the "Digital Omnibus on AI") was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It moved the application of the high-risk requirements:
| Category | Originally | Now |
|---|---|---|
| Annex III high-risk systems (Art. 6(2)) | 2 Aug 2026 | 2 Dec 2027 |
| Annex I product-embedded systems (Art. 6(1)) | 2 Aug 2027 | 2 Aug 2028 |
The dates above are taken from the recitals of Regulation (EU) 2026/1744 as fetched from EUR-Lex; the amended text of Article 113 should be re-checked against the consolidated version before anyone relies on them. Other deadlines (for example Article 50 transparency obligations) were treated differently and are out of scope here.
The mapping¶
Every row names something that exists in this repository today.
| Provision shape | Lodestar artifact | Where it lives | Pinned by |
|---|---|---|---|
| Automatic recording of events over the system's lifetime (Art. 12(1)) | Every governed step is an EventEnvelope appended to an NDJSON log; seq, logical_clock, causal_parent_ids, actor_id and timestamp are on every record, along with a versions object whose fields (model, prompt hash, tool version, policy version, memory snapshot) are all optional; the guard, MCP-proxy and runtime-gate emitters currently populate only schema_registry_version, so a host that wants the rest must set them |
packages/core/src/schemas/event.ts, packages/event-log |
event-log-single-writer, event-log-canonical-hash |
| Integrity checking of recorded payloads | payload_hash = sha-256 of the canonical payload, a checksum a verifier can recompute. It covers the payload only (not actor_id, timestamp, type or causal metadata), the reader does not recompute it automatically, and the writer accepts a caller-supplied hash, so it detects payload edits only when someone explicitly verifies. It is not a forgery boundary; signatures are added where authority matters (approvals, policy versions, packs) |
EventEnvelopeSchema.payload_hash |
event-log-canonical-hash (round-trip), policy-version-signature-required |
| Reconstructing an individual decision after the fact (Art. 12(2)) | The envelope has a causal_parent_ids field for a causal DAG, but the guard, MCP proxy and runtime gate mostly leave it empty ([]), so links are currently sparse; the trace package instead derives relationships from event types and payload ids; lodestar report <session-id> renders a read-side markdown projection of the chain from observation through action and outcome. The current projection does not follow causal_parent_ids and has no Revision section, so revisions are read from the log itself |
epistemic chain, packages/trace |
documentation-evidence-provenance, contradicted-belief-flags-dependent-decisions |
| Recording events relevant to risk situations (Art. 12(2)(a)) | sentinel.alerted events when a sentinel (for example suspicious-memory-origin) fires; firewall.*@1 audit events for belief adoption and transitions |
packages/core/src/schemas/sentinel.ts, packages/core/src/schemas/firewall.ts |
guard-arbiter-gates-dependent-action, harvest-projection-surfaces-durable-lessons |
| Monitoring performance over time (Art. 12(2)(b)–(c)) | calibration.computed@1: a calibration pass recorded as a durable, schema-valid, replayable event (measure ≠ write) |
packages/core/src/schemas/calibration.ts, lodestar harness calibrate |
calibration-event-is-durable |
| A human can intervene on a consequential action (Art. 14(4), Art. 26(2)) | L4 actions that are not denied outright by a deny rule or arbitration park at pending_approval (a denied L4 goes straight to rejected with no approval request; L5 is prohibited and always denied, so it never reaches an approver) and execute nothing until resolved; approval.requested@1 carries the matched rule's reason and the required_authority an approver must hold |
trust ladder, Policy Kernel, packages/core/src/schemas/approval.ts |
l4-action-requires-approval, pending-approval-cannot-execute |
| The intervention itself is recorded, with a reason | approval.granted@1 / approval.denied@1 (approver id and optional note), approval.expired@1 on timeout, and action.rejected carrying the action's reason |
packages/core/src/schemas/approval.ts, packages/core/src/schemas/action.ts |
approval-timeout-denies, approval-via-side-channel |
| Oversight by someone with the authority to exercise it (Art. 26(2)) | On the signed side-channel and quorum paths, an approver's resolution is verified against operator-pinned keys, and quorum adjudication also checks eligibility against required_authority; a forged or unsigned grant cannot un-park an action there. This is not uniform across every single-approver path: guard.wrap() trusts the ApprovalOutcome returned by the resolver you inject, the proxy and runtime hosts do not themselves apply required_authority to a signed single-approver response, and a local channel can be configured with allow_unsigned |
packages/core/src/schemas/approval.ts, packages/policy-kernel |
forged-approval-cannot-execute, forged-approval-via-http-channel-cannot-execute |
| More than one overseer for the highest-stakes actions | M-of-N quorum: each approver's own signed vote is logged; the kernel, not a client, adjudicates and emits approval.quorum_reached@1 naming every vote |
ADR-0041, packages/core/src/schemas/approval.ts |
quorum-requires-distinct-eligible-approvers |
| Blocking an action before it runs (partial answer to Art. 14(4)(d)–(e)) | A pending hold stays parked; a deny or deadline expiry is a terminal soft denial of that action that cannot be resumed into execution. The agent can re-plan and propose something else. This is pre-execution blocking, not a stop button | packages/action-kernel, packages/runtime-core |
approval-timeout-denies, runtime-gate-enforces-two-phase |
| Keeping logs without over-collection (Art. 26(6), read alongside data-minimisation duties) | lodestar ship applies a sensitivity ceiling client-side before transfer: above-ceiling records ship as a redaction marker plus the original payload_hash, so a verifier can still check the payload without the content leaving; an invalid ceiling throws rather than failing open |
packages/ship (lodestar.session_ship@1) |
ship-respects-sensitivity-ceiling, ship-wire-roundtrip |
| Retention for at least six months (Art. 26(6)) | The log is an append-only NDJSON file under the operator's control; retention is an operator practice, not something Lodestar enforces (see below) | operator storage | not applicable |
What this does not cover¶
- Retention. Lodestar writes an append-only log and can ship it elsewhere; it does not enforce a minimum retention period, rotate, or protect the file system. Meeting a six-month floor (or a longer sectoral one) is the deployer's storage and backup policy.
- Log content adequacy. Article 12(3) lists specific fields only for
Annex III point 1(a) systems. Whether the fields on
EventEnvelopeand the chain payloads suffice for a given high-risk system is a conformity question, not something a schema can answer. - Competence, training and organisational measures. Article 26(2) is
largely about who is assigned and how they are trained. Lodestar can check
that an approver holds the
required_authorityan operator configured; it cannot show the person is competent. - Personal data. A log containing personal data is itself subject to data protection law. The sensitivity ceiling reduces what leaves the machine; it is not a lawful-basis analysis.
- A system-level stop. Nothing here interrupts an already-executing tool or halts the agent as a whole; a deny refuses one proposed action. An Art. 14(4)(e)-style stop is the host's to provide.
- The model. The log records what the agent observed, claimed, decided and did. It does not explain the internals of the model underneath.
- Trust in the host. The log is not tamper-proof, and its hashes only help when explicitly verified. An attacker with write access to the file and the ability to recompute hashes can rewrite history; shipping to an independent store narrows that window.
Using this page¶
If you are assessing a deployment, a reasonable way to use the mapping is to
take one real consequential action, run lodestar report for its session, and
check that it answers what the system observed, what it concluded, and (on the
action) who finally approved or rejected it and why. lodestar report does not
project approval.* or quorum events, so the required_authority of the
request and the individual votes must be read from the raw log events for completed
actions. The trace approval projections cover pending requests and quorum
records only, not a historical view of resolved single-approver approvals. If both reads give a reviewer what they need in your own cases, you have evidence of
the shape described above, which counsel can then weigh against the actual
obligations.