Compliance packs and the AI registry
Governance > Compliance helps you show auditors how your AI systems are governed. Its tabs are Overview, Packs, AI registry, Reviews, Retention, Evidence and Bundles:
- an AI registry: one entry per AI system, with its owner, purpose, risk class, oversight, its role in decisions about people and what those people are told
- compliance packs: the controls of a framework, each filled in from evidence the platform reads live
- decision reviews: contests, explanation and human-review requests, corrections and overrides, with the reviewer, outcome and reason
- retention per jurisdiction: how long logs and content are kept, checked against every jurisdiction you operate in
- evidence bundles: a JSON and PDF export whose digest is anchored in the audit log, kept on the server
Compliance is for platform admins, the same as the audit log it reads. Two parts reach further: an organization's admins manage their own organization's registry entries, and a public notice page needs no sign-in at all.
The AI registry
Register each AI system you run: an agent (console-made or code, by project and name, covering every version), a catalog model, or an external system Orchestrate doesn't host. An entry records:
| Field | What it holds |
|---|---|
| Owner and purpose | Who is accountable, and what the system is for. |
| Risk class | Following the EU AI Act tiers: unassessed, minimal, limited, high or prohibited. |
| Data categories | The kinds of data it handles. |
| Human oversight | In the loop, on the loop, in command, or none, with notes. |
| Jurisdictions and licence | Where it is used, and under what licence. |
| Impact assessment | Intended use, affected people, benefits, risks, mitigations, residual risk and review cadence. |
| Decides about people | Not recorded, no, supports a decision a person takes, or decides alone. This is about decisions with legal or similarly significant effects (UK GDPR Article 22, Colorado's consequential decisions). |
| Notice to affected people | Three texts: what people are told before or when it is used, how a decision is explained to the person it concerns, and how to contest a decision, correct their data and reach a human reviewer. Required when the system decides about people or is high risk; the notice alone at limited risk. |
| Organization | Whose admins manage the entry. Install-wide entries are for platform admins only. |
| Evidence links | Console pages or web addresses. |
| Status | Active or retired. Retiring keeps the entry for the record. |
The registry lists the agents and models the platform has that nobody registered yet. Each entry also shows what the platform knows about that system, read live: its eval suites and recent runs, its live version and who set it, its guardrail profile, or the model's serving and access policy.
The registry checks a few rules: a high-risk system can't have no human oversight, an active entry classed prohibited fails the EU AI Act's prohibited-practices control, and a high-risk entry needs a complete impact assessment. Missing notice texts are shown as gaps on the entry. An entry that has decision reviews can only be retired, never deleted.
Packs
| Pack | Kind |
|---|---|
| EU AI Act (Regulation 2024/1689) | AI |
| NIST AI RMF 1.0 | AI |
| ISO/IEC 42001:2023 Annex A | AI |
| Colorado AI Act (SB 24-205) | AI |
| UK GDPR Articles 22 and 22A to 22D | AI |
| Korea AI Basic Act | AI |
| Singapore Model AI Governance Framework (2020) and its generative AI framework (2024) | AI |
| SOC 2 Trust Services Criteria | Security evidence |
| ISO/IEC 27001:2022 Annex A | Security evidence |
Choose which packs the install follows on the Overview tab. By default it follows every available pack.
Control references cite sections, and control titles are paraphrased. Controls no platform can show, such as a lawful basis, a domestic representative or reporting to a regulator, are left for attestation.
Each control names the evidence that shows it, and gets an automatic status:
| Status | Meaning |
|---|---|
| Met | All its evidence is present. |
| Partial | Some is. |
| Not met | None is, or some evidence is failing. |
| Unknown | A source couldn't be read. |
| Needs attestation | No platform evidence can show it, for example that you have an AI policy. |
A person can attest a control with a status, a note and links. The attestation is shown beside the automatic status and becomes the control's effective status. An attestation of "met" where the evidence says not met is kept, and flagged.
Evidence the platform reads
The registry and its risk classes, prohibited practices, impact assessments and oversight; approvals (tool-call approvals, eval-gate events and projects with a gate); the audit chain's verification; guardrails; evals (suites, passing runs in the last 90 days, gates and online evals); a live self-test of stored-content redaction; retention of run content; licences (the platform's, and models' in the registry); access control (the MFA policy and group-to-role bindings); change management (live versions and their audit events); monitoring (tracing, metrics and online evals); and model access (external providers and restricted models).
Also:
- Notices: the active entries that must tell affected people something, and how many do.
- Decision reviews: decision roles recorded, deciding systems with a contest route, and the reviews opened in the window, decided, open and past due.
- AI disclosure: every enabled phone number's first sentence still says the caller is speaking with an AI. A custom greeting can replace the default disclosure, so this is checked, in English, French and Arabic.
- Labelling of generated audio: the audio labeller's live self-test, and the newest stored call recording's label. A stored call recording is labelled as AI-generated speech inside the WAV file and in its stored metadata, and its download, like the gateway's text-to-speech answers, carries an
X-Vois-AI-Generatedheader. A live call is labelled to the ear by its AI disclosure. Recordings stored before labelling existed aren't relabelled. - Retention per jurisdiction: see below.
Decision reviews
On the Reviews tab, choose Record a review when a person asks about a decision an AI system took or supported:
| Request | What it records |
|---|---|
| Contest | The person disputes the decision. |
| Explanation request | The person asks how the decision was reached. |
| Human review request | The person asks for a human to look at it. |
| Data correction | The person corrects data the decision used. |
| Override | You replace the system's decision yourself. It is recorded as decided, with you as the reviewer. |
A review names the AI system, a decision reference (a run id, call id or case number) and a requester reference (a case or customer reference; no name is needed). A request is due in 30 days by default. Choose Decide to record the outcome (upheld, overturned, explained or withdrawn), with the reviewer's reason; an overturned decision names the decision that replaces it. Filter by open and past due; reviews past due are flagged.
Every step is an audit event. Reviews are never deleted. Through the API: POST /compliance/reviews and POST /compliance/reviews/{id}/decision.
Retention per jurisdiction
On the Retention tab, set per jurisdiction the minimum days the audit log must be kept and the maximum days run content may be kept. The defaults follow a stated basis, such as 183 days of logs in the EU (AI Act Articles 19 and 26(6)) and three years in Colorado for impact-assessment records; the others have no fixed period and say so. Change them to your counsel's reading.
The tab then checks the jurisdictions your registry names against what the install actually does: the audit log's retention, and how long each store of run content, and call recordings, is kept. A shorter audit retention than a jurisdiction's minimum, or content kept past its maximum, fails. The audit log is one hash chain, so its retention is one setting for the install. See Audit log.
Evidence bundles
Export a bundle on the Bundles tab, or with GET /compliance/export?packs=eu_ai_act,soc2&format=json. A JSON bundle holds the chosen packs with every control's automatic, attested and effective status, the evidence, the registry with decision roles and notices, the decision reviews of the window and those still open, the retention check, and a SHA-256 digest over the rest. The export is itself an audit event that carries the digest, so the hash-chained audit log anchors which bundle was handed out.
The PDF is rendered from a JSON bundle exported earlier, and only if its digest checks out, so the PDF and the JSON are the same bundle. It opens with a one-line summary per pack.
Every export is kept on the server, exactly as it was handed out. The Bundles tab lists them newest first, and a download is the same bytes again. Under Check a bundle, upload a bundle someone sends back: Orchestrate says whether nobody edited it, and whether this install issued it, by whom and when.
For a SOC 2 Type II observation window, export bundles on a schedule with a platform admin's token.
Your organization's AI registry
An organization's admins manage their own organization's entries under Governance > Your organization's AI registry, for the organization the switcher is on:
- They list, add, change and remove only their organization's entries: its projects' agents and code agents, models it owns, and the external AI systems it uses. Agents of its projects with no entry yet are listed for them to register.
- Anyone else's entry, or an install-wide one, is not visible to them.
- The rules are the install's, and every change is an audit event in the organization's own audit log.
Packs, evidence, reviews, retention and bundles stay with the platform admins. A member who isn't an admin is told who can manage the registry.
Public notice page
An entry's notice can be published for the people the system affects. On the entry, tick Publish the notice on a public page, then Open the public page or Copy link. It is off by default.
The page needs no sign-in and shows no console navigation. It follows the reader's language (English, Arabic or French), carries the organization's branding and asks search engines not to index it. It shows the system's name and purpose, the organization, its risk class, its role in decisions about people, human oversight, the data it uses, where it is used and the three notice texts. It never shows the owner, the impact assessment, evidence links, the licence, or which project or agent it is.
Publishing needs the notice text and every notice field the entry's risk class and decision role require. A retired or unpublished entry's page answers as not available. Publishing and withdrawing are recorded in the audit log. The page's data comes from GET /public/ai-notices/{id}, which is rate limited per client.
Not available yet
Organization admins managing decision reviews. Reviews stay with the platform admins for now.