Article 12 requires high-risk AI systems to automatically record events — logs — across the system’s lifetime, in a form that can identify risk situations and survive six months of Article 26 retention. In practice, that means append-only, tamper-evident event logs, not a database table an admin could quietly edit. This is what the requirement means technically, and the audit-logging layer VeriSwarm ships to help satisfy it.
Article 12 of the EU AI Act (Regulation (EU) 2024/1689) requires high-risk AI systems to “technically allow for the automatic recording of events (‘logs’) over the duration of the system’s lifetime.” The logs must be capable of enabling the provider to identify situations that may result in the system presenting a risk, and of facilitating post-market monitoring. Article 26 layers on top of that: deployers of high-risk systems must retain those logs for at least six months, unless other applicable law — GDPR, sector-specific record-keeping rules — requires longer.
The high-risk obligations, Article 12 included, became enforceable August 2, 2026, the latest milestone in the Act’s staged rollout. If an agentic system falls under a high-risk use case, this is a live requirement, not a roadmap item.
Most teams building agent logging reach for the obvious tool: a Postgres table with an occurred_atcolumn. It technically satisfies “automatic recording.” It does not satisfy the question an auditor, a regulator, or your own incident-response team will eventually ask: how do you know this record hasn’t been changed?
None of that requires malice — ordinary operational access is enough. A mutable log is an assertion. What Article 12 evidence actually needs is a record that’s independently verifiable: proof the log wasn’t altered, not just a claim that it wasn’t.
VeriSwarm Vaultis an append-only, hash-chained event ledger. Every event — trust decisions, Guard scan findings, Passport verifications, agent lifecycle changes — is written automatically once Vault is enabled, with each entry’s hash incorporating the previous entry’s hash. That chain structure is what makes tampering detectable instead of merely discouraged: alter, delete, or insert a record anywhere in the chain and every hash after it stops matching.
Once enabled, Vault requires no separate logging call. Any event processed through Gate, Guard, or Passport is written to the ledger as a side effect of normal operation — the “technically allow for automatic recording” language in Article 12, satisfied by default.
GET /v1/suite/vault/verify walks the hash chain and returns ok: true/falsewith a count of entries checked. A broken chain is immediately, mechanically detectable — not something you find out during a breach investigation months later.
Ledger exports (JSON or CSV, filterable by event type, actor, or agent) support the Article 26 retention window and give an auditor a structured, checksummed artifact rather than a raw database dump.
VeriSwarm’s compliance API maps this directly: the Article 12 control in GET /v1/compliance/eu-ai-act evaluates against Vault event volume and reports pass/warn/fail with evidence counts. Full technical detail on the ledger structure, event schema, and export formats lives in the Vault documentation.
Being direct about the boundary here matters more than the sales pitch. Vault is the logging and evidence layer for Article 12. It is nota certification, a legal opinion, or a conformity assessment. Turning it on does not make a system EU AI Act compliant on its own. The Act requires a risk management system (Article 9), technical documentation (Article 11), human oversight design (Article 14), and — for many high-risk systems — a formal conformity assessment before market placement. Those are organizational and legal processes, not a feature flag.
What Vault does is remove the weakest link in the record-keeping chain: a mutable log nobody can prove wasn’t edited. Everything else in an Article 12 compliance story — scope, retention policy, incident-response procedure — still has to be built by the team deploying the system, ideally with counsel involved for anything that touches a formal filing.
For an interactive, framework-specific self-assessment of exposure across the Act’s full obligation set, use the EU AI Act readiness checker — this page explains the logging control in depth; the checker scores a specific system against it.
Article 12 is the regulatory expression of a more general problem: proving an AI agent’s actions after the fact, in a way that can’t be quietly rewritten. The mechanism — a hash-chained, independently verifiable ledger — predates and outlives any specific regulation. For the underlying concept, how it works, and why it applies beyond EU compliance, see What Is an Immutable Audit Trail for AI Agents?
Article 12 requires high-risk AI systems to “technically allow for the automatic recording of events (‘logs’) over the duration of the system’s lifetime.” The logs need to be capable of identifying situations that may present a risk, and they need to support the log-retention obligation in Article 26 — deployers of high-risk systems must keep those logs for at least six months, longer if other applicable law requires it. The requirement isn’t just “have logs somewhere” — it’s automatic recording that stands up as usable evidence later.
It satisfies the letter of “automatic recording” but not the spirit auditors will apply. A conventional log table proves an event was written; it doesn’t prove the record wasn’t altered afterward. Anyone with write access — an engineer, a compromised credential, a rushed incident-response edit — can change a row and the table looks exactly the same as before. Article 12 doesn’t use the word “tamper-evident,” but a reviewer asking “how do you know this log is accurate” deserves a better answer than “we trust our admins.”
In a hash-chained ledger, each event’s cryptographic hash incorporates the hash of the event immediately before it, so every entry links to the one before it in an unbroken chain. Modify, delete, or insert an entry anywhere in that chain and every hash after it stops matching — tampering becomes mathematically detectable instead of merely procedurally discouraged. VeriSwarm Vault writes every agent-related event — trust decisions, Guard scan findings, Passport verifications, agent lifecycle events — into exactly this kind of append-only, hash-chained ledger, and exposes a chain-verification endpoint (GET /v1/suite/vault/verify) that walks the full chain and reports whether it’s intact.
No, and any vendor implying a single feature does that is overselling. Vault is the logging and evidence layer for Article 12’s record-keeping requirement — automatic, tamper-evident, exportable event logs on every agent action, with independently verifiable integrity. The Act contains many other obligations — risk management (Article 9), technical documentation (Article 11), human oversight (Article 14), transparency (Article 13), and a conformity assessment — that require organizational processes VeriSwarm doesn’t perform on a customer’s behalf. Vault gives a compliance program the evidence it needs for Article 12; it isn’t the whole program.
The EU AI Act rolled out in stages: prohibited-practice bans took effect February 2025, general-purpose AI model obligations in August 2025, and the high-risk system obligations — including Article 12’s logging requirement — became enforceable August 2, 2026. If an organization is deploying or providing a high-risk AI system in the EU market, the logging requirement is live now, not a future planning item.
No. Vault is a Max-plan feature. Gate’s free tier includes the full trust-scoring engine and unlimited event ingestion, but the hash-chained immutable ledger, the chain-verification endpoint, and audit exports that back Article 12 evidence require a Max plan. See /pricing for current plan details.
Vault is a Max-plan feature — the hash-chained ledger, chain verification, and audit exports aren’t on the free tier. Read how it’s structured, or check where a specific system stands against the full Act.