What Regulator-Ready Evidence Means for AI Agents

Compliance teams evaluating AI agents face a terminology problem. Logs, traces, findings, evidence, artifacts, packs — each word means something different to an engineer than it does to an auditor. The gap between what developers export and what regulators need is where compliance risk lives.

This post defines what "regulator-ready evidence" actually means in practice: the structure, the guarantees, and the format that make an evidence pack defensible in an audit.

The four layers of evidence

RADAR structures compliance evidence into four layers, each independently reviewable and each answering a specific auditor question:

What makes evidence regulator-ready

A regulator-ready evidence pack has four properties that distinguish it from a log export or a dashboard screenshot:

1. Tamper-evident structure

Every event in a RADAR audit trail is cryptographically linked to the previous event using a Merkle chain. Modifying a single event invalidates all subsequent hashes. The chain root is signed with an Ed25519 key that can be verified offline — no network access, no vendor API, no trust required. An auditor can verify evidence integrity with a single command-line tool, independent of RADAR.

2. Retention attestation

Evidence is useless if the retention policy cannot be proven. Each export bundle includes a signed attestation of the retention period, the purge policy, and any legal holds that were active during the collection window. This prevents the auditor question: "What happened to the records from before your retention window?"

3. Framework-specific control mappings

A generic audit log forces the compliance team to map events to regulations manually. RADAR evidence packs include pre-built control mappings: a PII detection finding is automatically linked to GDPR Article 5 (data minimization), SOC 2 CC6.1 (logical and physical access controls), and EU AI Act Article 13 (transparency). The mapping is explicit and reviewable.

4. Self-hosted chain of custody

Evidence that was exported from a cloud service and delivered via email has a weak chain of custody. RADAR evidence is generated and stored inside the customer's infrastructure. The export bundle includes the Merkle proof, the signing key identifier, and the generation timestamp — all produced inside the customer's security boundary. No third-party infrastructure touched the evidence.

Real-world evidence pack structure

A typical RADAR evidence pack contains the following sections:

Producing and verifying evidence in practice

The workflow from agent activity to auditor-ready evidence is a single CLI command. RADAR collects trace records continuously, then packages them into a self-contained, cryptographically signed evidence pack:

$ radar pack --last 24h --format pdf --output eu-ai-act-audit.pdf
  Evidence pack ready: eu-ai-act-audit.pdf
  Finding register: 12 findings (2 PII, 3 policy, 7 review)
  Control mappings: EU AI Act Art. 12-14, Art. 17, SOC 2 CC6.1
  Signed: Ed25519 key 7f42...a1b2
  Hash: SHA-256 a1b2c3d4e5f6...7890abcdef

  $ radar verify eu-ai-act-audit.pdf
  Signature: VALID
  Hash:       MATCH (SHA-256 a1b2c3d4e5f6...7890abcdef)
  Contents:   12 findings, 4 control mappings, 247 trace records
  Verdict:    Evidence pack integrity confirmed

The auditor receives a single PDF file containing every trace, finding, control mapping, and cryptographic signature. The audit trail is self-contained — no additional tools, no access to the production environment, no engineer-written summaries required. The compliance team can produce evidence packs on demand, and the auditor can verify them independently.

Why this matters for procurement

A compliance team evaluating RADAR does not need to trust the product's marketing claims. They can deploy the trial, connect a representative agent workflow, and inspect the evidence packs directly. The question is not "Does the vendor promise compliance?" but "Does the evidence pack satisfy our auditor's requirements?"

That distinction — verifiable evidence over vendor claims — is what makes evidence regulator-ready. Not a certification badge. Not a SOC 2 report. The evidence itself.