EU AI Act · Article 12 · eIDAS

Qualified audit trails for AI Act Article 12

TrustBeat provides eIDAS-qualified tamper-proof anchoring for every AI decision. Bind model input, output, and decision metadata to a single qualified proof — privacy-safe, independently verifiable, legally presumed accurate under EU law.

AI Act Obligations

Which AI Act articles does this cover?

Art. 12 — Record keeping

High-risk AI systems must automatically log events throughout their lifetime

The challenge

Proving that a specific model version made a specific decision at a specific time, without retroactive alteration of logs.

TrustBeat solution

Anchor SHA-256 hashes of model input and output together with decision metadata to a qualified eIDAS timestamp — immutable evidence of the decision state at time T.

Art. 13 — Transparency

High-risk AI systems must be designed to enable users to interpret outputs

The challenge

Demonstrating to regulators that decision logs accurately reflect the model's actual operation, not a post-hoc reconstruction.

TrustBeat solution

The combined hash binds model ID, system name, decision type, and time envelope to the proof. Any retroactive change to metadata is cryptographically detectable.

Art. 14 — Human oversight

Appropriate human oversight measures must be implemented

The challenge

Evidencing that human oversight was present (or absent) for each high-risk decision at the time it was made.

TrustBeat solution

The human_oversight flag is bound into the combined hash. Its value at the time of anchoring is sealed — it cannot be changed after the fact without invalidating the proof.

Technical flow

How AI Decision Anchoring works

The combined hash binds input, output, and metadata deterministically. Any tampering — including retroactive metadata edits or post-hoc log reconstruction — is cryptographically detectable.

01

Hash input and output locally

Your application computes SHA-256 of the model input data and the model output/decision. Neither is ever sent to TrustBeat — only the 64-character hex digests.

input_hash  = SHA-256(model_input_bytes)
output_hash = SHA-256(model_output_bytes)
02

Submit to TrustBeat

POST both hashes plus decision metadata (model ID, system name, risk category, human oversight flag, inference time window) to /v1/ai/decisions/anchor.

POST /v1/ai/decisions/anchor
{
  "input_hash": "a665a4...",
  "output_hash": "b94f6f...",
  "metadata": {
    "model_id": "claude-3-5-sonnet",
    "risk_category": "employment",
    "human_oversight": true
  }
}
03

Combined hash is anchored

TrustBeat computes SHA-256(input_hash ‖ output_hash ‖ JCS(metadata)) and queues it for the next 10-minute Merkle batch. Both hashes and all metadata are cryptographically bound together.

combined = SHA-256(
  input_hash_bytes
  || output_hash_bytes
  || UTF-8(JCS(metadata))
)
04

Retrieve qualified proof

After batching, retrieve an inclusion proof linking the combined hash to an eIDAS-qualified timestamp from an EU Trusted List QTSP. The proof bundle is self-contained and independently verifiable.

GET /v1/ai/decisions/{id}/export
→ {
  "verification_status": "VERIFIED",
  "anchored_at": "2026-04-15T10:10:00Z",
  "token": "MIID..."
}

Why TrustBeat for AI Act compliance

Decision integrity

Not just 'this model ran' — 'this model produced this output from this input at this exact time, with human oversight in place.'

Privacy-safe by design

Only SHA-256 hashes are sent to TrustBeat. Raw model inputs and outputs never leave your system. Fully GDPR-compatible.

30-year archive chain

Archive stamping runs automatically before each TSA certificate expires. Evidence remains valid for 30+ years under eIDAS.

Portable proof bundle

Export a self-contained JSON bundle per decision — includes Merkle proof, RFC 3161 token, and all LTV archive stamps. Independently verifiable offline.

Annex III — High-risk AI systems

If your AI system falls into any of these categories, Article 12 record keeping obligations apply from August 2026.

Biometric identification and categorisation
Critical infrastructure management
Education and vocational training
Employment, worker management, and self-employment
Access to essential private and public services
Law enforcement
Migration, asylum and border control
Administration of justice

Legal basis

EU AI Act (EU) 2024/1689 — Article 12

High-risk AI systems shall be designed to automatically record events (logs) throughout their lifetime. Record-keeping obligations apply from August 2026. TrustBeat provides the qualified cryptographic evidence layer proving these records existed and were unmodified at the time of creation.

eIDAS Regulation (EU) 910/2014 — Article 42

A qualified electronic timestamp carries the legal presumption of accuracy of the date and time it indicates and of the integrity of the data to which it relates. Issued by an eIDAS-qualified QTSP on the EU Trusted List. The presumption is reversed — your counterpart must prove the timestamp invalid.

Simple, flat pricing

One plan covers up to 50,000 AI decision anchors per month. No per-decision metering.

  • 14-day free trial — no credit card
  • 50,000 AI decision anchors / month included
  • Standalone or alongside other TrustBeat products
  • Cancel any time

€49

/ month

Start free trial
TB

Ready to anchor your first document?

Start free. No credit card required. Your first 100 hashes are on us.

eIDAS Regulation (EU) No 910/2014 · Art. 41 legal presumption · EU Trusted List QTSP

Qualified Audit Trails for AI Act Article 12 | TrustBeat