ReviewDesk
← Back

ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026

How does one decision trace become Colorado, California and EU evidence packs?

Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing

The short answer. You record a decision once — decision id, timestamp, model version, factors used, outcome, and whether notice and human review were offered — through the workspace API. ReviewDesk projects that single immutable trace into three sealed evidence packs: Colorado SB 26-189, California CPPA ADMT, and an EU Annex III pack honestly labelled preparatory until those duties apply on 2 December 2027. Anything the trace does not show renders as 'not recorded' — never invented.

What is the One-Trace, Three-Regime evidence engine?

A decision trace is the structured, immutable record of one automated decision, filed per decision through the ReviewDesk workspace API on Pro and Team plans. The evidence engine is the projection of that trace into regime-specific packs: the same underlying facts, arranged and labelled for Colorado's, California's and the EU's respective frameworks, each pack sealed against the trace it came from.

Why record once and project three times?

Because the alternative — three evidence processes for three regimes — guarantees divergence. The Colorado file says the notice was version 3, the California file says version 2, and both were assembled by hand from the same underlying events. A single trace removes the assembly step from the evidence chain entirely: the facts exist once, immutably, and each regime's pack is a view. When a fact changes through a correction, the correction is a new trace event referencing the old — the packs regenerate from a history that includes its own repairs.

The model also matches how decisions actually happen. Your screening tool does not run a Colorado decision and then a California decision; it runs one decision that may need evidencing in several frameworks depending on where the person lives and which duties apply. Instrument the event, not the paperwork. The underlying record format is described in the hash-chain guide.

What does a trace contain?

The elements a reviewer in any regime will ask for. A decision id that ties the trace to your own system's record. The timestamp of the decision. The model or tool version that produced it. The factors used — the inputs that materially counted. The outcome. And two booleans with outsized legal weight: whether the pre-use notice was given, and whether human review was offered. A per-decision-type scope-test worksheet travels with the pack, recording your own determination of whether the system materially influences or substantially replaces human judgement.

Notice what is absent: no narrative, no reconstructed rationale, no fields the engine fills by inference. If the factor list was not supplied, the pack shows the gap. That restraint is the product's core evidence ethic — 'not recorded' is a finding, not a failure of the form — and it is what makes the packs credible to a reader who checks them against the inbox and the audit trail.

How do the three packs differ?

The Colorado pack is organised around SB 26-189's sequence: pre-use notice status, the adverse-action explanation elements, the correction path, human-review availability — the duties as they apply from 1 January 2027. The California pack follows the CPPA ADMT frame: pre-use notice, and the opt-out or qualifying human-appeal handling, as recorded. The EU pack frames the same trace for the Annex III file and is labelled preparatory evidence until the high-risk duties apply on 2 December 2027 — the label is part of the pack, not a footnote, so no reader can mistake preparation for compliance.

The differences are presentational and structural, not factual. Each pack draws exclusively on the trace and the workflow record; each renders missing elements identically as not recorded. A reviewer comparing packs from the same trace should find one story in three arrangements. Where your own scope worksheet records that a regime does not apply to a decision type, the pack says that too — scope honesty is evidence as well.

How does the scope-test worksheet fit in?

Every pack carries the worksheet for its decision type: your recorded determination, per regime, of whether the system materially influences or substantially replaces human decision-making, with the reasoning space and the timestamp. The determination is yours — the engine records and evidences it; it does not make legal determinations. That separation protects both sides: ReviewDesk never certifies your scope position, and your scope position is preserved exactly as you made it, when you made it.

Worksheets should be revisited when the system changes — a vendor model update that increases automation is a scope event, not just a technical one. The full method is in the scope-test guide. In the pack, the worksheet's date does quiet work: a determination refreshed after each material change reads as governance; a determination from launch day, untouched through three model versions, reads as wallpaper.

How do you integrate traces into a live pipeline?

Through the workspace API key: your decision system files a trace per decision as the decision is made, in the same transaction discipline as writing your own decision log. Retrofitting traces in nightly batches is possible but weaker — the timestamp then proves filing time, not decision time, and gaps appear exactly where pipelines failed overnight. File at decision time, monitor filing failures like payment failures, and reconcile counts daily: decisions in your system versus traces in ReviewDesk should match, and a mismatch is an operational incident, not a rounding issue.

The integration surface is deliberately narrow — one endpoint shape, one key, the defined fields. Teams evaluating the mechanics can read the wider integration notes in the API integration guide. What matters more than the mechanics is the ownership: name the engineer who owns trace completeness the way you name the owner of billing events, because evidence pipelines fail silently unless someone owns their silence.

What does the engine refuse to do?

It refuses to complete your record. It will not infer a notice from the fact that notices are usually given, will not synthesise factors from a model name, and will not backdate. It will not convert the EU pack's preparatory label into a compliance claim. And it will not let a corrected trace pretend the original never existed — corrections append, history persists. These refusals are the feature. Evidence tooling that flatters its inputs produces packs that collapse on first contact with an inbox record or an applicant's own copy of their letter.

Sell that honesty internally before you need it externally. Leadership accustomed to dashboards that always render green may find 'not recorded' confronting in the first pack. Reframe it once: every 'not recorded' is a work order with a location attached, and the second pack — after the gaps close — is the proof the programme works. The engine is available on Pro and Team; the plan contents are on pricing.

Your five actions for the One-Trace, Three-Regime evidence engine

  1. File a trace at decision time through the workspace API — decision id, timestamp, model version, factors, outcome, notice and review flags.
  2. Reconcile trace counts against your own decision log daily; treat a mismatch as an incident.
  3. Complete the scope-test worksheet per decision type and refresh it after every material system change.
  4. Read the first pack hunting for 'not recorded' entries — each one is a located gap with a fix.
  5. Keep the EU pack's preparatory label intact until the Annex III duties apply on 2 December 2027.

Worked example: one hiring decision, three packs, zero re-assembly

A fictional employer — Summit Staffing — screens applicants across three states with one tool. For one Colorado applicant declined on 9 February, the trace holds: decision id from Summit's own system, timestamp, model version 4.2, factors (licence status, availability window, safety-record flags), outcome, notice given under workflow version 3, review offered in letter RD-2041. The Colorado pack renders the SB 26-189 view of exactly those facts; the same trace pattern, for a California applicant the same week, renders the California pack with the opt-out handling recorded.

When the Colorado applicant corrects an availability error and is re-scored, a correction event joins the chain referencing the original trace — the pack now shows both the original decision and its repair, in order. Nothing was re-typed, nothing was quietly replaced, and a reviewer in either regime reads the same history Summit's own inbox shows. One record, honestly kept, did the work of three evidence processes.

Frequently asked questions

What is a decision trace?

The structured record of one automated decision — id, timestamp, model version, factors used, outcome, and whether notice and human review were offered — filed through the ReviewDesk workspace API and held immutably, hash-chained, on Pro and Team plans.

Which regimes do the packs cover?

Colorado SB 26-189 (effective 1 January 2027), California CPPA ADMT (regulations in effect 1 January 2026, compliance 1 January 2027), and an EU Annex III pack labelled preparatory until those duties apply on 2 December 2027.

What happens if a trace element is missing?

The pack renders it as 'not recorded'. The engine never invents factors, notice status or review offers — gaps appear as gaps, so they can be found and fixed.

Who makes the scope determination in the worksheet?

You do. The worksheet records your own determination per decision type and regime, timestamped and stored with the pack. ReviewDesk records and evidences the determination; it does not make legal determinations.

Is this guide legal advice?

No. ReviewDesk is templates and workflow tooling, not legal advice. The engine is described as built; consult counsel for your obligations.

Keep reading: What is a hash-chained decision record — and why does immutability matter as evidence? · How do you run the scope test: does your system materially influence or substantially replace human decision-making? · How should you prepare EU Annex III evidence for automated decisions — before the duties apply? · How do you let someone correct wrong data after an adverse automated decision? · How do you write principal reasons for an adverse decision that a person can actually use? · All ReviewDesk guides · ReviewDesk home · Try the free desk · Free letter generator · Free notice builder · Pricing

ReviewDesk is built by Ingenious AI. Templates and workflow tools — not legal advice.