Article 12, mapped.
Automatic event logging for high-risk AI systems. What Runback supplies, and what stays yours.
EU AI Act · Article 12 · Annex III
390 days.
Until automatic logging is mandatory for high-risk AI. 2 December 2027.
Article 12 applies to high-risk AI systems from 2 December 2027 (Annex III) and 2 August 2028 (Annex I) — automatic event logging, retained at least six months, sufficient to trace what the system did and why. Penalties reach €15,000,000 or 3% of worldwide annual turnover. Systems being designed now are the ones that have to carry it.
5 supplied. 2 partial. 1 yours.
Each requirement, paired with exactly what answers it.
- Art. 12(1)SDK capture, automatic, from inside your processSupplied
The SDK records every model call, tool call, and agent step from inside your process — automatic, not manual documentation. Capture is a single import, or OpenTelemetry if you already emit spans.
npx @runback/verify <your-export>.json - Art. 12(2)(a)Every gated decision recorded, with the rule that blocked itSupplied
Policy rules evaluate on every gated tool call and record the decision — allowed or blocked — as part of the run. A blocked action carries the rule that blocked it, so a risk situation is a queryable event rather than something inferred from prose.
GET /api/compliance/report?start=YYYY-MM-DD&end=YYYY-MM-DD - Art. 12(2)(b)Period report computed from the data, not sampledSupplied
Runs aggregate into a period report — volumes, error rates, policy evaluations and blocks, redaction counts — computed from the database, not sampled, so the report's numbers are the data's numbers.
GET /api/compliance/report - Art. 12(2)(c)Per-agent coverage against your declared registerSupplied
Per-agent coverage reconciles the agents you declared against the ones actually reporting: declared but never instrumented, reporting then gone silent, or running while absent from your register.
GET /api/app/coverage - Art. 12(3)Timestamps, inputs, context — human identity only via an approval gatePartial
Start/end timestamps, exact inputs, retrieved context, and tool results are captured per run. Human verification is captured where an approval gate is used. If your process has a reviewer step outside Runback, that identity isn't in our record — record it yourself.
- Art. 19 / 26(6)Per-plan retention and legal hold; six months needs Enterprise or self-hostPartial
Retention is per plan, enforced by a scheduled sweep: Enterprise unlimited, Pro 90 days, Scale and Growth 60, Starter 30, Community 7, with regulated verticals raising the floor (healthcare 180 days). Meeting the six-month obligation therefore means Enterprise, a vertical floor that covers it, or self-hosting with your own retention. Legal hold suspends deletion for runs under a preservation obligation, and if the holds table cannot be read the sweep is skipped — the failure favours keeping data.
- Art. 12 — impliedHash chain, signatures, RFC 3161 timestamps, public logSupplied
The Act does not use the word 'tamper-proof', but a record that can be edited without anyone noticing is not evidence of anything.
Every run is hash-chained and signed. Each hour, each org's runs are checkpointed into an append-only audit log, and each checkpoint is signed (Ed25519, or HMAC when no keypair is configured). Checkpoints are also sent, best-effort, to independent RFC 3161 timestamp authorities and published to a public append-only log — so the record is not merely signed by us, it is anchored outside our control.
openssl ts -reply -in <checkpoint>.tsr -token_in -text - Art. 9, 11, 17Evidence feeds them; the process is yoursYour obligation
Runback supplies evidence that feeds these; it does not produce them. A conformity assessment is an organisational process. Runback's Enterprise regulatory dashboard does list Art. 9(1), 9(7) and 17(1), and shows one as "Evidenced" when the evidence it maps to is present — read that as evidence on file, not as the article being met.
Assessed by hand, requirement by requirement.1
This page maps Runback's capabilities to named controls. It is not legal advice and not an assessment of your compliance — whether an obligation applies to you, and whether you meet it, is your assessor's or regulator's determination.
Check every claim yourself. No account.
A compliance claim you cannot check is a marketing claim. Every row above resolves to something you can run.
# 1. Verify a record's integrity and origin — no account required
npx @runback/verify audit-export.json
# exit 0 = intact AND signed by our published key
# exit 2 = intact, but origin unproven
# exit 1 = a check failed
# 2. Check our published signing key out of band
curl https://runback.dev/.well-known/runback-audit-key.pem
# 3. Confirm a checkpoint was time-stamped by an authority we do not control
openssl ts -reply -in runback-checkpoint-4-freetsa.org.tsr -token_in -text
# 4. Read the public transparency log and archive it
curl https://runback.dev/api/transparencySteps 2–4 use no Runback software at all: an auditor should not have to run our verifier to check our tamper-evidence.
- No regulator certifies a file format. Whether your deployment satisfies Article 12 is a determination for your assessor. What we supply is the evidence that argument needs.
- The harmonised standards are not final. ISO/IEC 24970 (AI system logging) is at FDIS and prEN 18229-1 is at DIS ballot. Nobody can claim conformance to an unpublished standard, including us. We track both and will publish a clause map when they land.
- Runback is not certified. No SOC 2 report, no ISO 42001, no third-party penetration test yet. The security page states exactly what is shipped and what is not.
- Article 12 is one clause. Articles 9, 11 and 17 are organisational processes. Evidence helps; it is not the same thing.
- This deadline is twelve months away. Another one is not. If you are APRA-regulated, CPS 230 has applied since 1 July 2025 and the transition for existing service-provider arrangements closed on 1 July 2026. It never mentions AI, which is exactly why a model API inside a critical operation is caught by it.