Audit and Attestations

Signed proof for every decision.

Every run becomes a signed record: its AI BOM (AI Bill of Materials), the rules it was checked against and how each one came out. Hand it to your auditor as is.

AISquare · Reasoning

RUN 8823 · SUPPORT AGENT · REFUND.CREATE

Why the $1,250 refund was blocked

4 claims · 1 unsupported
CLAIMEVIDENCESTATUS
Customer identity verifiedID check, session 8823Supported
Outage fell inside the SLA windowStatus incident INC-2291, Sep 12Supported
Customer is owed $1,250Assumed from 3 days of downtimeAssumption
$1,250 is within the refund capRule 4.2 caps refunds at $500Unsupported

Agent decision

$1,250 refund blocked. Sent for manager approval.

Rule 4.2 · refund cap $500 · Why?

Attestation

RUN 8823 · REFUND.CREATE

Signed by AISquare RuntimeSigned and tamper-evident
Key ID
ais-runtime-key-01
Signed at
Sep 12, 2026 · 14:32 UTC
Decision
Blocked, sent for approval
Rules
4.2, sla.window, privacy.pii
Evidence
Ticket 8823, INC-2291

AI BOM entry sealed

The problem

When an auditor asks why, logs can't answer.

Logs show what an agent did. They can't prove why, or that nothing changed since.

Audit requestRefund 8823 · Sep 12
Which model made the decision?Unknown
Which prompt was live?Unknown
Which rule applied?Unknown
Has the record changed since?Can't tell

Rebuilt by hand from logs, weeks later. Editable, so not evidence.

Inside a record

Four layers, sealed together.

AI BOM

gpt-4o-mini · v7 · 3 tools

Verdicts

3 rules · 1 blocked

Reasoning

4 claims · 1 unsupported

Signature

Signed and tamper-evident

  1. 01

    AI BOM

    What the agent used: model, system prompt version, tools and tool calls.

  2. 02

    Verdicts

    Each rule it was checked against, the version that was live, and the result.

    How rules are enforced →
  3. 03

    Reasoning

    The claims the agent made and the evidence behind them.

    See RML →
  4. 04

    Signature

    Every rule verdict and the run itself are signed and tamper-evident. Change anything and verification fails.

What's inside an AI BOM

Run 8823 · Support agent

Export · CycloneDXJSON
How exports work →

Model

gpt-4o-mini

Provider and version recorded

System prompt

v7

Version recorded

Tools called

crm.lookupstatus.incidentsrefund.create

Data used

Refund policy 4.2Ticket 8823INC-2291

Rulebook

Support Rulebook v4

Continuous evidence

No audit project. The run documents itself.

Time from decision to evidence

Compiled after the fact

Weeks pass
Questionnaires
Evidence hunt
Report
weeks later

AISquare

Sealed with the run
Signed

Evidence is written while the agent works, not reconstructed weeks later.

Compliance packs

See each record against the frameworks you answer to.

Map the same signed records to the frameworks your auditors ask about.

EU AI Act

High-risk AI obligations

OWASP Top 10 for LLMs

LLM application risks

NIST AI RMF

AI risk management

ISO/IEC 42001

AI management system

SR 11-7

Model risk management

DoD AI principles

Responsible, traceable, governable

Your internal controls

Your own policies and SOPs

And more, mapped to the same signed records.

Talk to us about your framework →

Example · EU AI Act pack

EU AI Act

Support agent · last 30 days
RequirementWhere the evidence comes fromStatus
Art. 12 Record-keepingEvery step of every run, recorded and signedCovered
Art. 13 TransparencyReasoning for each decision in RMLCovered
Art. 14 Human oversightRules, corrections and escalations to your teamCovered
Art. 15 Accuracy and robustnessRule verdicts and failures across runsCovered
Incident reconstructionReopen and replay any past runCovered
Art. 9 Risk managementYour risk process, with run evidence attachedNeeds your input

The evidence layer, not a certificate.

No tool makes you compliant on its own. AISquare gives your auditors the records, mapped to the frameworks they check.

Export and share

Hand it over without a fire drill.

Export one or many

Export a single record or a whole set of runs at once.

Share with auditors

Give your auditor access to the records they need to review.

Verify anywhere

Anyone can check that a record hasn't changed since it was signed.

Export formatsPDFJSONDSSE / in-totoCycloneDX AI BOM

Built for every reviewer

One record. Every reviewer.

Risk and compliance

Sign off with evidence, not assurances.

See every decision against your compliance packs.

Internal audit

Sample any decision and read it end to end.

Check the signature yourself.

Engineering

Reconstruct what happened, step by step.

Fix what keeps failing with Learnings →

For developers

Pull your compliance summary over the API.

curl -s -H "X-API-KEY: $EXPLAINABILITY_API_KEY" "$EXPLAINABILITY_GATEWAY_URL/v1/studios/$STUDIO_ID/policy/compliance"

Returns the recent compliance rate, violations and the rules that fail most.

FAQ

Questions, answered.

What is a decision record?

A signed record of one agent run: what it used, which rules it was checked against, the result of each, and the reasoning behind it.

What exactly is signed?

Each rule verdict and the run itself. Both are tamper-evident, so any later change fails verification.

What is an AI BOM?

The record of the model, system prompt version, tools and tool calls a run used. Exportable as CycloneDX.

Which formats can I export?

PDF, JSON, DSSE / in-toto envelopes and a CycloneDX AI BOM.

Can I export many runs at once?

Yes. Export one record or a whole set of runs.

Can I share records with an auditor?

Yes. Give your auditor access to the records they need, and they can verify each one.

Which frameworks can I map to?

Compliance packs include the EU AI Act, OWASP Top 10 for LLM applications, NIST AI RMF, ISO/IEC 42001, SR 11-7 and the DoD AI ethical principles, plus your own internal controls.

Does this make us compliant?

No tool does that on its own. AISquare gives you the evidence layer: signed records mapped to the frameworks your auditors check.

Give your auditors proof, not logs.

Every decision your agents make, signed and ready to hand over.