Resources

Human approval and execution lineage

Human approval is part of an execution record. This playbook explains what to capture before an action, when a reviewer decides, and after execution.

Lineage Explorer
Invoice INV-2048

Payment execution

Submission accepted

  1. Payment requested

    Submit supplier payment · €420,000

    Treasury payments agent
  2. Approval required

    Amount exceeds the €250,000 threshold

    Payment limits · version 3
  3. Treasury approved

    Invoice and approval scope confirmed

    Treasury reviewer
  4. Tool accepted submission

    Submission accepted; settlement not shown

    Payment tool
What to capture at each stage

An agent requests a payment above the review threshold. The policy requires approval. The evidence needs to connect that request to the review and the eventual execution result.

  1. 1

    Capture the request

    Record the agent, proposed action, relevant inputs, and request identifier.

  2. 2

    Keep the policy decision

    Record the policy version, evaluated context, outcome, and reason.

  3. 3

    Attach the human decision

    Keep the reviewer identity, decision, rationale, time, and relationship to the original request.

  4. 4

    Record what executed

    Connect the permitted action and its result to the same request. A rejected request must remain distinguishable from a completed action.

  5. 5

    Prepare the review artifact

    Collect the records and supporting material with a manifest and integrity information.

The useful unit is the full decision: request, rule, reviewer, result, and supporting evidence. Lineage Explorer and Evidence Room provide the investigation and review surfaces for that work.

Verification checks and limits

A Sealed Evidence Bundle gives the reviewer an artifact to inspect. Verification checks its cryptographic structure against the supplied material and trust configuration.

Authenticity needs an out-of-band key check. The SEK, TEK, and receipt signatures are all checked against the keys the bundle embeds in keys/jwks.json, so offline the bundle proves only that it is internally consistent. To prove it came from KLA, pin those keys against a KLA-published JWKS or key fingerprint.

The OpenTimestamps Bitcoin anchor is trusted offline. Confirming its block height compares the proof against the live Bitcoin chain, and a pending calendar receipt stays pending until a calendar upgrade.

The immudb ledger anchor's proof is present and parses offline. Confirming its inclusion against the ledger's signed state needs a networked check.

Manifest signature
The manifest canonicalizes (RFC 8785 / JCS) and its digest recomputes. The detached JWS carries valid SEK and TEK ES256 signatures over that digest, checked against the public keys the bundle embeds in keys/jwks.json.
Merkle inclusion
Every artifact re-hashes (SHA-256) to the digest the manifest declares, each artifact's inclusion proof reaches the sealed root, and the recomputed Merkle root equals the manifest value.
Receipt signatures
Every governance receipt chain verifies: each receipt's ed25519 signature checks against a key embedded in the bundle, and each step's prevReceiptHash links to the prior step, genesis to last.
Ledger hash chain
Every exported ledger record re-hashes to its stated hash, and consecutive records link through previousHash end to end.
OpenTimestamps anchor
timestamp.ots parses under the strict OpenTimestamps rules and commits to the recomputed manifest digest. When the receipt is Bitcoin-attested the verifier reports its block height; a pending calendar receipt is accepted as pending Bitcoin confirmation.

Regulatory precedents

These references describe different recordkeeping contexts. Use them to examine the evidence question in each regime and follow the source for its scope.

SOX

Original requirement

"Adequate internal control" (no definition)

Read the source

Recordkeeping practice

COSO framework: 17 principles, 87 focus points, 40% Big 4 audit deficiency rates

MiFID II

Original requirement

"Accurate time source" for trading records

Read the source

Recordkeeping practice

100 microseconds UTC divergence (HFT), WORM storage, 5-7 year retention, 72-hour reconstruction

GDPR

Original requirement

"Appropriate technical measures"

Read the source

Recordkeeping practice

2FA now expected (Haga Hospital fine), documented access procedures required

MDR

Original requirement

"Technical documentation" requirements

Read the source

Recordkeeping practice

IEC 62304 audit trails, complete version control, traceability from requirements to tests

Source: Regulation (EU) 2024/1689. Review the requirements applicable to your system with the people responsible for its legal and operational assessment.

Questions and details

Does Article 12 explicitly require cryptographic integrity?

No, but the pattern from MiFID II, SOX, GDPR, and MDR is unambiguous: principles-based text becomes prescriptive practice within 2-4 years. Auditors and notified bodies will interpret "automatic recording" to mean tamper-evident logging.

What retention periods apply to AI logs?

Article 12 specifies 6 months minimum for automated logs, but sector-specific requirements often override this. Financial institutions should expect 5-7 years. Technical documentation must be retained for 10 years after the system is placed on the market.

When will harmonized standards be published?

prEN ISO/IEC 24970 (logging standard) is in Draft International Standard ballot. prEN 18286 (QMS standard) entered public enquiry in October 2025. Final standards expected Q4 2026, but auditor expectations are forming now.

What access will notified bodies have?

Under Annex VII, notified bodies can require full API access to training, validation, and testing datasets. They may conduct direct tests if unsatisfied with provider evidence. Dataset access logs will themselves become audit targets.

How should we handle ML model auditability?

Thresholds for extensive logging remain unclear, but auditors will likely expect: model version tracking, hyperparameter changes, training run metadata, and decision traceability for high-risk outputs.

Human Approval and Execution Lineage Playbook | KLA