AI Audit Exchange
Go back

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

Importing ReviewDesk — ADMT Decision Review Evidence Into an Audit Room

What is decision evidence?

Decision evidence is the ADMT decision-review record produced by ReviewDesk and imported into AI Audit Exchange rooms: decision records and notice artifacts as the review system generated them, held read-only and unchanged, SHA-256 sealed at intake and badged with the ReviewDesk source label in the room manifest, with import dates spread across the audit period they evidence.

Why the decision layer carries the audit's weight

For systems that make or shape consequential decisions about people, the examination converges on one population of records: what was decided, on what basis the review system recorded, and what the affected person was told. ReviewDesk — ADMT Decision Review produces exactly that population as its native output. In the room, its artifacts are typically the largest and most-sampled source, and their treatment sets the engagement's tone: imported complete and current, they tell the auditor the organisation understands what an examination is. Imported partially or late, they tell a different story before a single record is read. Give this layer your best filing discipline — monthly cadence, full runs, zero editing — because the audit will.

The full-run rule

Import every export ReviewDesk produces for the period, not the exports that present well. The rule rests on sampling logic: an auditor's sample means something only against a complete population, and completeness is the auditee's contribution to the examination's validity. The moment an auditee filters the population, every conclusion drawn from sampling inherits the filter — and a competent examiner, discovering the filter, must either rebuild the population or discount the engagement. Full runs also protect the auditee: a decision record that looks awkward in isolation often reads correctly inside its population's context, and the room is the context. If a record is wrong, the remedy is correction at source and a superseding import, narrated in the thread — the versioning pattern in evidence version control.

The import routine for decision records

Monthly: export the period's ReviewDesk output, import untouched into the room tagged with the right regime and period, confirm the manifest entries — badge, format, size, hash, import date — and glance at the growing shape of the population. Volume is expected here; decision systems produce records at operational tempo, and the manifest will show it. Where exports arrive in parts (per month, per decision family), import each part as its own artifact rather than merging parts into a bundle: merging creates a derivative file that no product produced, while parts preserve the product's own grain. Merging is the kind of helpfulness the import rule — unchanged, always — exists to prevent.

Notices: the other half of the layer

Decision records show what the system did; notices show what the person was told — and the audit reads them as a pair. Ensure the notice artifacts ReviewDesk produces travel with the decision records through the same monthly rhythm, so any sampled decision can be matched to its notice story inside the room. Where notices are generated by a distinct flow in your stack, their exports still arrive as ReviewDesk evidence if ReviewDesk produced them, or as clearly-badged direct uploads if they are the organisation's own documents. What must not happen is a decision layer with its notice half silently thinner: the manifest's source and format columns will show it, first to you in the monthly review, then to the auditor if you let it.

Sampling the decision layer: what to expect

Auditors sample this layer stratified: across quarters, across decision families, across outcomes. They verify seals on the sample, then read for the record's internal coherence — basis recorded, review trail present, notice consistent with decision. Your rehearsal should mirror theirs: in the readiness review, draw your own stratified sample, verify its hashes, and read five pairs of decision-and-notice as a stranger would. Questions you cannot answer from the room become thread questions to resolve before the engagement. Expect the auditor's questions in this layer to be the engagement's most specific; the room's thread, with its anchored answers, is where that specificity gets honoured rather than feared.

Decision evidence across regimes and time

Both regimes read this layer, differently: SB 1119 directly, as the substance of the independent audit; the CPPA cybersecurity examination for what decision-control says about governed processing. File the same exports into each room that needs them, separately sealed per room. Across time, resist reorganising history: decision records stay in their period rooms with their intake dates, retained toward the 10-year horizon, because later questions about a period's decisions are answered from that period's room or not answered credibly at all. The decision layer is the audit's memory of what your systems did to people; treat its preservation with the seriousness that sentence implies.

Five actions to take

  1. Import the full ReviewDesk run for the period, monthly — never a selection, never edited.
  2. Import export parts as separate artifacts; never merge parts into a derivative bundle.
  3. Keep notice artifacts flowing with decision records through the same rhythm and room.
  4. Rehearse stratified sampling in the readiness review: verify seals, read decision-and-notice pairs as a stranger.
  5. File the same exports into both regime rooms where both examinations will read the decision layer.

Worked example

A fictional hiring-platform firm, Open Range Talent Systems, files ReviewDesk exports monthly into its SB 1119 room: 12 rounds, 214 artifacts, decision records and notices together, import dates keeping operational tempo. The manifest's population is large enough that the auditor's stratified sample of 25 — across four quarters and three decision families — is a genuine sample, with 25 seals verified in her first week.

Her one structural question targets a thin month: April's import count dips by half. The thread already holds the auditee's April note — a decision family was paused for a model change, volume genuinely halved, and the resumption exports appear in May. The dip is data, explained contemporaneously, inside a complete population. Had April been 'completed' retrospectively to look smooth, the same dip would have been invisible and the population's credibility — all 214 artifacts of it — would have rested on a cosmetic edit. The full-run rule exists for exactly this fork in the road.

Frequently asked questions

Can we import only the decisions the auditor will care about?

No — that is pre-selection, and it invalidates sampling. Import the full run; the auditor chooses the sample. Your credibility is built on the population's completeness.

A decision record contains an error. What do we do?

Correct it in ReviewDesk, re-export, import the corrected artifact as a new entry, and post a supersession note in the thread. The original stays, sealed — the pair tells the quality-control story honestly.

ReviewDesk exports in monthly parts. Merge them for tidiness?

Never merge. Import each part as produced; merging creates a file no product made and breaks the unchanged rule that gives imports their authority.

Do notices really need their own filing attention?

Yes. The audit reads decisions and notices as a pair. Keep both halves flowing monthly so any sampled decision's notice story is present in the same room.

How long should decision evidence be retained?

Toward the long horizon — per-room retention up to 10 years on paid plans, set at close-out. Decision records are the layer later questions reach for first.

Keep reading

All AI Audit Exchange guides · Auditor directory · About AI Audit Exchange · Security · Pricing · AI Audit Exchange home · Previous: Importing ProvenanceGuard Evidence Into an Audit Room · Next: Importing CompanionGuard Evidence Into an Audit Room

Importing CRAClock Evidence Into an Audit Room · How to Answer Auditor Evidence Requests Faster · Common AI Audit Evidence Mistakes (and Their Fixes)