AI Audit Exchange
Go back

Guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026

What Evidence Does an SB 1119 Audit Need?

What is SB 1119 evidence?

SB 1119 evidence is the artifact set filed in an SB 1119-tagged evidence room on AI Audit Exchange for one audit period: suite imports produced by ReviewDesk — ADMT Decision Review, CompanionGuard, ProvenanceGuard and CRAClock, stored read-only and unchanged with source badges and SHA-256 seals at intake, plus direct uploads for organisational records. The Exchange hosts evidence and matches auditors; it never generates the evidence types themselves and makes no legal determination about any customer's duties.

The audit reads records, so the room is made of them

An independent audit under SB 1119 is an examination of what your organisation's AI did and what your processes produced around it. The auditor's confidence comes from records that were created in the ordinary course of operation — not prepared for the visit. That principle decides the room's contents: file the ordinary course. Decision records as the review system produced them, notices as they were generated, safety evidence as the guardrails logged it, provenance as it was attested. Evidence created retrospectively for the auditor carries its birthdate in every field; evidence filed monthly from live systems carries the period's own texture, which is what examiners are trained to trust.

Decision records and notices: the ReviewDesk layer

For systems making or shaping consequential decisions, the decision trail is the audit's centre of gravity: what was decided, on what basis, with what human review, and what the affected person was told. ReviewDesk — ADMT Decision Review produces exactly this layer, and its exports import into the room unchanged. File them completely for the period — the full run, not a selection. Selection is the auditor's job, exercised by sampling the manifest; an auditee who pre-selects has answered the audit's central question themselves, and no examiner accepts that answer. Completeness here is not volume for its own sake; it is the difference between a population and a showcase.

Safety, provenance and conformity layers

CompanionGuard's safety and guardrail evidence shows how companion-style systems were constrained and monitored during the period. ProvenanceGuard's records show where generated content came from and under what attestations it circulated. CRAClock contributes conformity-timing material. Together with the decision layer, these four sources form the suite's complete account of the period — each produced by the product responsible for that domain, each badged accordingly, none of them generated by the Exchange itself. If one of these systems did not operate in your organisation during the period, the room simply lacks that source; the manifest's silence is honest, and the thread can note why.

The organisational layer, as direct uploads

Around the system records sits the human layer: governance minutes that adopted a policy, approvals that authorised a deployment, attestations signed by accountable people, correspondence that belongs to the record. These arrive as direct uploads, badged as direct uploads — the organisation speaking in its own voice, distinguishable at a glance from system output. File them with the same period discipline as the imports: the adoption minute belongs in the period of its meeting, not in a year-end bundle. A room whose organisational layer is as well-dated as its system layer reads as a governed organisation, which is, not coincidentally, what the audit is trying to determine.

Depth across the period beats depth in one month

Examiners sample across time because behaviour drifts: configurations change, staff rotate, volumes spike. A room whose artifacts cluster in two months testifies about two months. Monthly filing spreads the record across the period naturally; the manifest's import-date column then demonstrates continuity without a word of argument. Pair depth with the freshness lens — some SB 1119 questions legitimately reach back, so older artifacts may be substance rather than staleness — and read validation flags accordingly, as prompts to check that each stretch of the period has its evidence, not as orders to purge the old. See the freshness rule for the distinction.

Retention: the evidence must outlive the engagement

SB 1119's framework sits in a registry world reported to carry 10-year audit-record retention. Build the room's afterlife accordingly: retention set to up to 10 years on paid plans at close-out, manifest exported as the record's inventory, legal hold available the moment a period is touched by dispute. Evidence that evaporates at the free plan's 12-month default cannot serve a framework measured in decades; the retention step in the room's completeness view exists to make sure the long setting is a decision taken, not a default suffered. The afterlife planning is covered in the 10-year retention programme.

Five actions to take

  1. File the full run of decision records and notices from ReviewDesk — never a pre-selected sample.
  2. Import CompanionGuard, ProvenanceGuard and CRAClock evidence for every period in which those systems operated.
  3. Add the organisational layer as dated direct uploads: governance minutes, approvals, attestations.
  4. File monthly so the manifest's import dates demonstrate continuity across the whole period.
  5. Close out with the manifest exported and retention set toward 10 years, hold ready if the period is disputed.

Worked example

A fictional lending-technology firm, Copper State Lending Tech, builds its SB 1119 room for the 2028 period. The decision layer dominates, as it should for its business: 88 ReviewDesk artifacts across twelve monthly imports. CompanionGuard adds 9, ProvenanceGuard 23, CRAClock 6. Eleven direct uploads carry the human layer — four governance minutes, three deployment approvals, two attestations, two correspondence records — each filed in its own month.

The auditor samples 20 artifacts stratified across quarters and sources; 20 hashes verify. Her one coverage question — a six-week gap in ProvenanceGuard imports in spring — is answered from the thread, where the team had noted the provenance tool's migration outage in the week it happened, with the post-migration exports filed on resumption. The room's completeness was not an accident of storage; it was the visible result of filing the ordinary course, monthly, for a year.

Frequently asked questions

Does the Exchange create any of this evidence for us?

No. AI Audit Exchange never generates notices, decision traces, crisis-runtime evidence, SBOMs or conformity packs. Those are produced by ReviewDesk, CompanionGuard, ProvenanceGuard and CRAClock and imported unchanged.

Should I file every decision record or the important ones?

Every one the system produced for the period. Sampling is the auditor's professional act, performed on a complete population. A pre-filtered set undermines the examination's logic and the room's credibility.

What if our organisation does not use one of the suite products?

Then that source is absent, honestly. The manifest shows the sources that operated; the room's thread can note the scope. Absence with an explanation is evidence practice; absence disguised is a finding.

How long should SB 1119 evidence be retained?

Plan toward the long horizon: the registry framework is reported with 10-year audit-record retention, and paid plans allow per-room retention up to 10 years. Set it at close-out, not by default.

Who decides whether our evidence is sufficient?

The auditor, exercising independent judgement on the record — informed by your counsel's view of the regime. The platform's checklists are organisational aids and make no legal determination.

Keep reading

All AI Audit Exchange guides · Auditor directory · About AI Audit Exchange · Security · Pricing · AI Audit Exchange home · Previous: What Evidence Does a CPPA Cybersecurity Audit Need? · Next: How to Export an Audit Room Manifest (and When)

Immutable Evidence Storage: Why It Matters for Audits · How to Verify an Artifact Hash in an Evidence Room · AI Auditor Vetting: How the Owner-Run Workflow Works