Technical mapping - draft-zambo-aer1-26

EU AI Act evidence: where execution records fit.

Article 12 logging requirements, and what a receipt trail records

Article 12 of Regulation (EU) 2024/1689, official EUR-Lex text requires high-risk AI systems to technically allow automatic event logging over the lifetime of the system, enabling traceability appropriate to the system's intended purpose. It identifies relevant monitoring and risk events, and additional logging requirements for the systems named in Article 12(3). This paragraph is a summary, not legal advice.

  • Event capture: each call through Zambo tools returns a receipt. Work outside that boundary is not automatically captured.
  • Change detection: compare the digest over the saved result using the public verifier.
  • Retrieval: public receipt URLs can be opened without the producer's account. Private receipt bytes may be withheld; review the individual record.
  • Retention: the live documentation does not publish a retention duration applicable to this use. Do not infer a lifetime storage commitment. Retain the records your own review requires and check their availability.

Zambo provides a logging layer. Whether it satisfies your Article 12 obligations is a legal determination for your own compliance review, and this page does not make it.

A receipt proves the saved result was not changed since it was saved. It never proves the answer was right and never proves an external event happened.

A technical mapping of AER-1 execution records against Articles 12, 19, and 26 of Regulation (EU) 2024/1689. The full table names each requirement, status, and gap.

A technical mapping between the AER-1 receipt format (draft-zambo-aer1-26) and the record-keeping provisions of Regulation (EU) 2024/1689 (the EU AI Act): Article 12 (record-keeping), Article 19 (automatically generated logs), and Article 26 (obligations of deployers), as those provisions concern records of what an AI system did.

This document is a technical mapping, not legal advice and not a conformity assessment. A record kept in AER-1 form does not by itself show that a system or an organisation complies with the Regulation. The duties in these articles fall on providers and deployers of high-risk AI systems. AER-1 is a record format. It can supply the execution records those duties call for, where the system's relevant events are tool calls recorded by an AER-1 producer. It cannot operate a system, keep records on anyone's behalf, or make the judgements the articles assign to providers and deployers.

Every row below names its gap. A mapping that only lists strengths would not be worth reading.

How to read the status column

rule that produces it. Operating the producer and keeping the records remain the operator's duties in every row.

The gap column says which part it does not supply.

so the gap is on the record.

Article text is quoted from Regulation (EU) 2024/1689 as published in the Official Journal (EUR-Lex). AER-1 references are to draft-zambo-aer1-26 by section.

Article 12: Record-keeping

ProvisionWhat it requiresAER-1 mechanismStatusGap
Art. 12(1)"High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system."Section 3: a producer MUST issue a receipt for every tool call it receives, including calls it rejects before execution. "A call that leaves no receipt is a conformance failure: the absence of a record MUST NOT be allowed to hide an attempted call." Recording is therefore automatic and exception-free at the tool-call layer.PartialOnly events that reach a recording producer as tool calls are recorded. Events inside a model that never become tool calls leave no receipt, and whether recording spans the system's lifetime is a property of the deployment, not of the format.
Art. 12(2)Logging capabilities shall enable the recording of events relevant for (a) identifying situations that may result in the system presenting a risk under Article 79(1) or in a substantial modification, (b) facilitating post-market monitoring under Article 72, and (c) monitoring the operation of the system under Article 26(5).Each receipt binds an identifier, a creation time, the tool name, version and caller scope, a provenance class (Section 5), and a hash commitment over the exact output bytes (Sections 3 and 4). Optional members add an input commitment (Section 3.3), execution context (model, source system, framework, environment; Section 3.2), checks not performed and actions not taken (negative_record, Section 3.4), reversal state, and fresh-read mismatch outcomes. Sections 7 and 8 order a job's receipts into a hash-chained timeline and bind multi-step workflows under one root, so a reviewer can reconstruct what ran, in what order, with what recorded result. Records are checkable by any party without an account (Section 9).PartialAER-1 records events; it does not judge which events signal a risk, and it performs none of the monitoring itself. Identifying risk situations, running post-market monitoring, and monitoring operation remain the provider's and deployer's work. The receipts are evidence those processes can read.
Art. 12(3)(a)For the systems in Annex III point 1(a) (remote biometric identification): recording of the period of each use (start date and time and end date and time of each use).Every receipt carries created_at, the time the record was created. In a Section 7 timeline, the first and last entries bound the period over which a job's recorded calls occurred.PartialAER-1 defines no start-of-use and end-of-use members. A record creation time is not necessarily the start of a use, and bounding a period from a chain is an inference by the reader, not a recorded fact.
Art. 12(3)(b)The reference database against which input data has been checked by the system.execution_context.source names the source system the tool acted upon; the draft's own examples include a database name (Section 3.2).PartialThe member is optional, specific to no kind of database, and producer-asserted. Section 3.2 places execution_context outside the hash-bound core: it is not part of the metadata_hash computation, and a verifier must surface it as asserted context, never as an attested fact.
Art. 12(3)(c)The input data for which the search has led to a match.The exact recorded input can be committed with input_hash (Section 3.3) under a published byte layout, and the reference payload shape carries the inputs themselves, so what the input was can be checked later.PartialAER-1 has no concept of a match. Nothing in the format designates which inputs produced one. The input can be recorded and committed; the match designation cannot.
Art. 12(3)(d)The identification of the natural persons involved in the verification of the results, as referred to in Article 14(5).None in the interoperable core. field_provenance (Section 3.4) names which party wrote each field, by role: actor, tool, or counterparty.Not coveredAER-1 defines no member that identifies a natural person. Caller identity appears only in deployed records outside the interoperable core (Section 3), which verifiers must not require. A deployment that needs Article 12(3)(d) must record the persons' identities outside AER-1 or in deployment-specific members, and accept that verifiers will not check them.

Article 19: Automatically generated logs (providers)

ProvisionWhat it requiresAER-1 mechanismStatusGap
Art. 19(1)Providers shall keep the logs referred to in Article 12(1), to the extent such logs are under their control, "for a period appropriate to the intended purpose of the high-risk AI system, of at least six months, unless provided otherwise in the applicable Union or national law, in particular in Union law on the protection of personal data."Section 9: every receipt is retrievable at a stable public URL, and "the record at that URL MUST NOT be altered." Corrections are published as new receipts carrying the superseded id in the supersedes member; they never rewrite the original. Because each kept record carries its output commitment and metadata commitment, a later change to the committed bytes or to the bound metadata of a kept record is detectable by recomputation (Sections 4, 3.1, and 6). Section 10 anchors can additionally fix a record's digest at a point in time on infrastructure outside the keeper's store.PartialAER-1 sets no retention period. The six-month floor, and the judgement of what period is appropriate to the intended purpose, belong to the provider. The format also cannot keep anything by itself: storage, availability, and deletion policy are the deployment's. What the format contributes is that records kept under it can be shown, later, to be the records that were issued.

Article 26: Obligations of deployers

Only the paragraphs that concern execution records are mapped. Paragraphs on worker information, registration, and data protection impact assessments do not turn on record format and are out of scope here.

ProvisionWhat it requiresAER-1 mechanismStatusGap
Art. 26(1)Deployers shall take appropriate technical and organisational measures to ensure they use the systems in accordance with the instructions for use.None.Not coveredThis is an organisational duty. Receipts can afterwards evidence how a system was in fact used, which is evidence about the measure, not the measure.
Art. 26(2)Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.None in the core record.Not coveredAER-1 records executions, not oversight assignments, and approval events are not a core member. A deployment may record approvals in deployment-specific members; verifiers will not check them.
Art. 26(4)To the extent the deployer exercises control over the input data, the deployer shall ensure that input data is relevant and sufficiently representative in view of the intended purpose.input_hash (Section 3.3) commits to the exact input bytes the producer recorded, under a published byte layout, so the input a decision rested on can be produced and checked later.PartialRelevance and representativeness are judgements about data. A commitment proves what the recorded input was; it says nothing about whether the input was representative.
Art. 26(5)Deployers shall monitor the operation of the system on the basis of the instructions for use; where use may result in a risk under Article 79(1), inform the provider or distributor and the relevant market surveillance authority without undue delay and suspend use; where a serious incident is identified, inform the provider, then the importer or distributor and the authorities immediately.The deployer's operation record exists in checkable form: receipts for executed and rejected calls alike, hash-chained job timelines whose breaks are detectable by recomputation (Section 7.2), workflow records, and verification outcomes under Section 6. A monitoring function, human or automated, can read the same records an auditor will later read, because resolution is public and needs no account (Section 9).PartialMonitoring is supported; the duties are not discharged by records. Informing providers and authorities, suspending use, and reporting incidents are acts of the deployer that no record format performs.
Art. 26(6)Deployers shall keep the logs automatically generated by the system, to the extent under their control, for a period appropriate to the intended purpose, of at least six months, unless provided otherwise in applicable Union or national law, in particular on the protection of personal data.The same mechanics as Article 19(1): stable public URLs, no alteration in place, corrections by supersession, commitments that make later alteration of kept records detectable, optional external anchors over digests only (Section 10, privacy bound: an anchor MUST NOT contain step content, tool inputs, or tool outputs).PartialThe duration and the control judgement are the deployer's, exactly as in Article 19(1). One further tension to plan around: this paragraph is expressly subject to Union data-protection law, and AER-1's default is public resolution. A deployment whose logs contain personal data must decide what may resolve publicly before it issues receipts, because the format will not make that decision for it.

What AER-1 provides that these articles do not ask for

Three properties go beyond the articles' floor. They are listed because deployers evaluating evidence formats will ask, and because none of them should be mistaken for compliance.

AER-1 additionally lets a party who did not produce a record check that the record was not changed after it was saved, by recomputing its commitments, without trusting the keeper's storage and without an account (Sections 6 and 9). The check covers the saved result. It does not show the result was correct.

recording system executed the call, observed it through a gateway, or received it as an agent's report, and verification must not upgrade one class into another (Section 5). The articles require traceability appropriate to the purpose; they do not prescribe a provenance vocabulary.

(126 vectors, 129 checks) and independent implementations, so the claim "this producer emits AER-1 records" is itself testable against fixed vectors rather than asserted.

Scope notes

is high-risk, and which of its events are tool calls, are classification and deployment questions this mapping does not answer.

(EU) 2026/1744, the Chapter III obligations that include Articles 12, 19, and 26 apply from 2 December 2027 for systems under Article 6(2) (Annex III) and from 2 August 2028 for systems under Article 6(1) (Annex I). Preparation time is the point of publishing a mapping now; nothing in this document suggests the duties are already in force for a given reader.

systems of Annex III point 1(a). Its rows are mapped because they are the Act's most specific statement of log content, not because most agent systems fall under them.

Sources

Official Journal text, EUR-Lex: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng

to 10, IETF Datatracker: https://datatracker.ietf.org/doc/draft-zambo-aer1/26/

v1.5.7): https://zambo.dev/aer1/corpus.json

Canonical source text: EU AI Act mapping - kit document at commit 5beff84

Questions and answers

What does Article 12 require?

Article 12 requires high-risk AI systems to technically allow automatic recording of events over the system lifetime. Its logging provisions cover traceability and monitoring, and Article 12(3) names minimum details for specified remote biometric identification systems.

Does AER-1 fully cover the EU AI Act rows in this mapping?

No. Of the 12 mapped rows, none is fully covered by the format alone: 9 are Partial and 3 are Not covered. The gap column identifies what the format does not supply.

Does an AER-1 record establish compliance?

No. This is a technical mapping, not legal advice or a conformity assessment. A record kept in AER-1 form does not by itself show that a system or organisation complies.

Who has the duties described in Articles 12, 19, and 26?

The duties in these articles fall on providers and deployers of high-risk AI systems. AER-1 is a record format and cannot operate a system, keep records on anyone’s behalf, or make the judgements assigned to providers and deployers.

Questions and answers

What are the EU AI Act Article 12 logging requirements?

Article 12 addresses automatic recording of events over a high-risk AI system's lifetime, with traceability appropriate to its intended purpose, relevant risk and monitoring events, and additional requirements for the systems it names. Consult the official text and your own legal review.

Can execution receipts serve as AI system logs?

Per-call receipts can be one input to a logging program: call-time records, saved-byte checks and public retrieval. They do not cover all system events, provide a stated lifetime retention term, or decide whether your legal obligations are satisfied.

Add Zambo to your AI

Connect the free MCP at https://zambo.dev/mcp in Claude, ChatGPT, Gemini, Cursor, Grok, Windsurf, or VS Code. No key is needed. The first receipt arrives with the first call through Zambo. Adding a URL does not automatically install a connector; use your client's MCP settings.

More AI agent answers · Try the demo · Current pricing and tool-specific limits

Keep going with Zambo in your AI

Get receipts for your own AI work. Add one link, then ask in plain English.

Add Zambo to your AI