EU AI Act - capability crosswalk
If your AI program needs to evidence EU AI Act Article 12 record-keeping, Article 14 human oversight, or Article 73 serious-incident reconstruction, here is how the obligation maps to the AGLedger record, the Signed Statement chain, and the audit-export surface.
The crosswalk below is a capability mapping, not a compliance certificate. AGLedger provides the evidence pattern; your program provides the policy, the methodology, and the decisions.
Article references reviewed against Regulation (EU) 2024/1689 (OJ L, 12 July 2024) and the Digital Omnibus, Regulation (EU) 2026/1744
Where enforcement stands
The AI Act entered into force on 1 August 2024. Prohibitions and AI-literacy duties have applied since 2 February 2025, and general-purpose AI obligations since 2 August 2025. Under the Digital Omnibus, formally adopted by the Parliament (16 June 2026) and the Council (29 June 2026), the high-risk obligations this page maps - including Article 12 record-keeping - apply from 2 December 2027 for standalone Annex III systems and 2 August 2028 for high-risk AI embedded in regulated products. The records you will need in 2027 have to start accumulating before 2027.
The same evidence pattern - signed records, hash-chained, append-only - also supports non-AI automated work in any control family that expects tamper-evident audit trails for RPA, CI pipelines, and microservice calls. AGLedger is software you self-host; the regulations are AI-framed, the underlying evidence pattern is not.
Article-by-article mapping
| Article | Requires | AGLedger provides | You own |
|---|---|---|---|
| Art. 6 + Annex III | Classification rules for high-risk AI systems | Field to record the classification (high, limited, minimal, unclassified) and Annex III domain tag per record, applied consistently across the chain | The classification decision and methodology |
| Art. 12 | Record-keeping (event logging) | Append-only audit vault: every state change, attestation, and tolerance check result recorded automatically, Ed25519-signed and hash-chained, with signed Merkle checkpoints written about every six hours. External anchoring of those checkpoints to write-once storage (S3 Object Lock) is available and ships disabled | Determining which events are in scope, and deciding whether to turn external anchoring on |
| Art. 13 | Transparency and provision of information to deployers | Full chain exportable and machine-readable (JSON, CSV, NDJSON, COSE), so deployers can see what the system intended and did | The instructions for use, and deciding what to disclose and to whom |
| Art. 14 | Human oversight | Structured record of the designated overseer (name, role, authority scope, designation date) plus the signed Gate verdict when a human renders a decision. AGLedger records oversight; it does not control the AI runtime, and there is no stop button | Designating the overseer, defining their authority, and providing the mechanism that can actually interrupt the system |
| Art. 15 | Accuracy, robustness and cybersecurity | A partial fit, on the integrity of the record: signing, hash-chaining, database-enforced immutability, payload-drift detection, and offline verification. Tolerance bands hold numeric bounds on record criteria and deviations land as signed events. AGLedger does not measure the AI system’s accuracy | Defining the tolerance thresholds and acceptance criteria, and measuring the accuracy and robustness of the AI system itself |
| Art. 19 | Automatically generated logs (retention of at least six months) | Append-only retention: the software deletes and overwrites nothing in the vault, so how long the log survives is set by the retention policy you operate on your own database; offline-verifiable export at any point in that window | The retention schedule and its enforcement, and the lawful-basis and erasure analysis under GDPR |
| Art. 26 | Deployer obligations, including Art. 26(6) log retention | 4 attestation record types (workplace notification, affected persons, input data quality, FRIA) plus the same retained, tamper-evident log store on the deployer side | Performing the attestation (we record it, you do it) and the deployer’s retention schedule |
| Art. 27 | Fundamental rights impact assessment | Structured record per assessment with risk level, domain, mitigation measures | Conducting the assessment itself |
| Art. 72 | Post-market monitoring by providers | A continuous, signed record stream of production behaviour (intents, actions, outcomes, verdicts) as a data source the monitoring plan can rely on without questioning its integrity | The monitoring plan, the analysis, and the conclusions |
| Art. 73 | Reporting of serious incidents | The audit vault contains timestamped, hash-chained records of every agent action, record, completion, and verdict: export a complete incident reconstruction from a single signed source rather than correlating across multiple unsigned systems | Incident determination, authority notification, and the filing obligation |
Where AGLedger sits, and where it stops
AGLedger is not an AI system under Article 3(1). It is an accountability-evidence layer for automated work, and its relevance under the Act is as deployer- and provider-facing tooling: it produces the records, logs, and oversight evidence the high-risk obligations call for, in a form a regulator or a counterparty can verify offline.
Three obligations are out of scope, and the table above does not map them. Article 9 asks for a risk management system and Article 17 for a quality management system; AGLedger is a record of operation, so its tolerance bands, verdicts, and dispute trail are evidence that those processes ran, not the processes themselves. Article 50 content marking is out of scope because AGLedger neither generates nor marks content. Classification is yours as well: AGLedger stores the risk classification you set and does not determine high-risk status.
The stop control stays with the systems that run the agent. Article 14(4)(e) contemplates a human who can interrupt the system; that interrupt lives in your orchestration. AGLedger records that a halt was requested and holds the record. The product has no halt state and no halt endpoint, so the Article 14 control is met elsewhere in the deployment.
Accuracy under Article 15 is measured by your evaluation; verdicts record acceptance against the criteria you set. The evidence layer validates the structure of the evidence and leaves business data, prompts, and model outputs unread, and an encrypted record's payload is unreadable to it by construction.
These obligations exist because automated work needs structurally durable evidence. Whether or not your jurisdiction enforces them on schedule, the engineering requirement is real today.
Record-keeping under the Act also intersects data-protection law: Article 12 logs are themselves records to which GDPR residency and international-transfer analysis applies. Self-hosting keeps that analysis short - the logs live in your PostgreSQL, in the jurisdiction you choose, with no processor relationship and no transfer introduced by the evidence layer. Data sovereignty is a property of the deployment model, not a clause to negotiate.
Frequently asked
- Is this crosswalk a compliance certificate?
- No. The crosswalk is a capability mapping, not a compliance certificate. AGLedger provides the evidence pattern; your program provides the policy, the methodology, and the decisions.
- How does AGLedger map to EU AI Act Article 12 event logging?
- Article 12 requires automatic recording of events (logs). AGLedger provides an append-only audit vault: every state change, attestation, and tolerance check result is recorded automatically. You own determining which events are in scope.
- How does AGLedger support Article 73 serious-incident reporting?
- Article 73 requires providers to report serious incidents to market surveillance authorities within fixed deadlines. The audit vault contains timestamped, hash-chained records of every agent action, record, completion, and verdict. You can export a complete incident reconstruction from a single signed source rather than correlating across multiple unsigned systems. You own incident determination, authority notification, and the filing obligation.
- How long must EU AI Act logs be retained?
- At least six months. Article 19 requires providers to keep the Article 12 logs under their control for at least six months; Article 26(6) places a parallel duty on deployers. AGLedger holds records in an append-only audit vault, so the software deletes and overwrites nothing. How long the log is kept is set by the retention policy you operate on your own database, and an export verifies offline at any point in that window.
- When do the EU AI Act high-risk obligations apply?
- Under the Digital Omnibus, formally adopted by the Parliament (16 June 2026) and the Council (29 June 2026), the high-risk obligations - including Article 12 record-keeping - apply from 2 December 2027 for standalone Annex III systems and 2 August 2028 for high-risk AI embedded in regulated products. Prohibitions and AI-literacy duties have applied since 2 February 2025, and general-purpose AI obligations since 2 August 2025.
- Is AGLedger itself an AI system under the EU AI Act?
- No. AGLedger is not an AI system under Article 3(1). It is an accountability-evidence layer for automated work, and its relevance under the Act is as deployer- and provider-facing tooling: it produces the records, logs, and oversight evidence the high-risk obligations call for, in a form a regulator or counterparty can verify offline.
- Can AGLedger stop an AI system under Article 14?
- No. AGLedger records oversight; the AI runtime and its stop control stay with the systems that run the agent. It records that a halt was requested and holds the record. The product has no halt state and no halt endpoint, so the mechanism that interrupts the system is yours to provide.
- How does self-hosting affect data-protection analysis for Article 12 logs?
- Article 12 logs are themselves records to which GDPR residency and international-transfer analysis applies. Self-hosting keeps that analysis short: the logs live in your PostgreSQL, in the jurisdiction you choose, with no processor relationship and no transfer introduced by the evidence layer.
Primary sources
- Regulation (EU) 2024/1689, the AI Act - the official consolidated text on EUR-Lex, and the source of every article cited above
- Article 12, record-keeping - the article this page maps most directly, in a navigable form
- European Commission: regulatory framework for AI - the Commission's own timeline, which is where date changes land first