AI Decision Anchoring
EU AI Act Article 12 requires immutable records of every decision made by high-risk AI systems. Prove what your model decided, when it decided it, and that no one altered the record afterward — for 30+ years.
Obligations apply from August 2026. A database row is not sufficient — it can be edited, deleted, or lost. An eIDAS-qualified proof cannot.
14-day free trial · no credit card required
When the record is all that matters
Three moments where an unalterable decision log changes the outcome.
The regulator investigates a decision
Your credit-scoring model rejected a loan application. The applicant filed a complaint. The supervisory authority wants the complete, unaltered record: what the model received as input and exactly what it output — at that moment.
Every decision is cryptographically sealed within 10 minutes. The proof shows input hash, output hash, and model version — immutable, independently verifiable.
The model has been updated
You retrained your model six months ago. A decision made before the update is now disputed. You need to prove which model version decided and what it saw at that moment — not what the current version would decide.
The model ID and version are cryptographically bound to each decision record at the time it was made. A retrained model cannot retroactively alter past records.
Human oversight is questioned
Under Art. 14, certain high-risk decisions require human review before taking effect. A claimant alleges the decision was fully automated with no human in the loop.
The human oversight flag is recorded and cryptographically bound per decision. You can prove whether a human reviewed and confirmed or overrode the output — at that exact time.
How it works
Four steps from every AI decision to an irrefutable, regulation-ready record.
Hash inputs and outputs locally
Your application computes SHA-256 hashes of the model input and output on your own infrastructure. Raw data — prompts, documents, personal data — never leaves your system.
This makes the record privacy-safe by design. You comply with GDPR data minimisation while meeting the AI Act's immutability requirement simultaneously.
Combine with model metadata
The two hashes are combined with metadata — model ID, version, risk category, human oversight flag — using RFC 8785 JCS canonical JSON. The result is a single combined hash.
Canonical JSON ensures the combined hash is deterministic: the same inputs always produce the same hash, making the proof reproducible and independently verifiable.
eIDAS-qualified anchor every 10 minutes
The combined hash is queued for the next Merkle batch. Every 10 minutes, all pending decision hashes are anchored with a single qualified eIDAS timestamp from an EU Trusted List QTSP.
Your application receives a 202 Accepted instantly — the anchor cycle never blocks your inference pipeline.
Per-decision proof — valid 30+ years
Each decision gets its own Merkle inclusion proof. It binds: input hash, output hash, model ID, human oversight flag, and the qualified timestamp. Verifiable offline, forever.
Recursive LTV archive stamping ensures the proof chain remains valid well beyond any EU AI Act retention requirement.
metadata = model_id · version · risk_category · human_override · timestamp
RFC 8785 (JCS) · RFC 3161 (TSA) · EU AI Act Art. 12 · eIDAS Art. 41
Article-by-article coverage
EU AI Act obligations mapped to exactly what each decision record contains.
| Article | Covered |
|---|---|
| Art. 12(1) | |
| Art. 12(2) | |
| Art. 14 | |
| Art. 9 | |
| Annex XII |
EU AI Act Annex III — High-risk AI system categories
Art. 12 record-keeping obligations apply to all systems in these categories. Obligations begin August 2026.
Anchor a decision in five lines
Five zero-dependency SDKs. One call after each inference. Art. 12 covered.
$ npm install trustbeatimport { TrustBeatClient } from "trustbeat";
const tb = new TrustBeatClient("YOUR_API_KEY");
const { decisionId } = await tb.ai.anchorDecision({
inputHash: sha256(modelInput), outputHash: sha256(modelOutput),
modelId: "my-model-v2", riskCategory: "high",
});
console.log(decisionId); // → 01JQ...Privacy by design: The sha256() calls above run locally in your infrastructure. Raw model inputs, outputs, and any personal data never reach TrustBeat servers. Only the combined hash is transmitted — compliant with GDPR data minimisation.
Common questions
Which AI systems require Art. 12 record-keeping?
All high-risk AI systems listed in Annex III of the EU AI Act — covering biometric identification, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice/democratic processes. If your system falls under any of these categories, Art. 12 logging is mandatory from August 2026.
From when exactly do Article 12 obligations apply?
For high-risk AI systems first placed on the market or put into service after 2 August 2026, obligations apply immediately. Systems already deployed before that date have an additional transition period (generally 36 months from the Act's entry into force). Starting implementation now avoids a compliance sprint in 2026.
Do my model inputs or outputs get stored on TrustBeat servers?
No. You compute the SHA-256 hash of your input and output data locally, on your own infrastructure. Only the combined hash — derived from both data hashes plus model metadata — is sent to TrustBeat. Raw prompts, documents, personal data, and model outputs never leave your system.
What exactly is the human oversight flag (Art. 14)?
A boolean field (human_override) recorded per decision: did a human review and either confirm or override the AI output before it took effect? This flag is cryptographically bound to the decision record at the time of anchoring — it cannot be changed retroactively without invalidating the proof.
How long must records be kept, and does your archive cover it?
EU AI Act Annex XII requires at least 6 months for most high-risk systems; sector-specific regulations (e.g. financial services) may require longer. Our plan tiers cover 1 year, 3 years, 5 years, and 7+ years of event retention. Regardless of the plan, proofs are kept forever — and our LTV archive stamping ensures they remain cryptographically valid for 30+ years as algorithms evolve.
Valmis ankkuroimaan ensimmäisen asiakirjasi?
Aloita ilmaiseksi. Ei luottokorttia tarvita. Ensimmäiset 100 tiivistettä ovat meiltä.
eIDAS Regulation (EU) No 910/2014 · Art. 41 legal presumption · EU Trusted List QTSP