Archigov · docs
  1. Docs
  2. Concepts

AI in Archigov

Eight optional AI features that propose and narrate while a named human decides — off by default, per-feature, any provider, with a deterministic fallback for every one.

#The Boundary

AI in Archigov proposes and narrates; it never decides. Every AI output is stored (when it is stored at all) with origin: ai and verification: unverified. It never records a position, never approves, and never adopts anything. Citations an AI produces must resolve to a real artefact, decision, resource or design section, or they are discarded before storage.

Every feature is off by default and has a deterministic fallback, so Archigov is complete with no model configured.

#The Features

FeatureWho uses itWhat it doesThe human step
Pre-reviewRuns on a reviewCited advisory findings on top of the deterministic checksBoard seats weigh them; it never approves
Decision ReviewAuthorsCritiques a draft decision's argument — unsupported claims, strawman options, circular reasoning, unstated risks, contradictions with adopted artefacts — and offers a tightened version; suggests assessments for unevidenced critical claimsThe author applies or ignores it; nothing is stored
Decision DraftingAuthorsDrafts a decision from a planned-change work itemThe author edits and submits
Rule DraftingAuthorsTurns an artefact's statement into a rule; the draft comes back validated and dry-run against the live modelThe author saves; a board adopts
Control ReferencesAuthorsSuggests framework references for a draft artefactThe author prunes and saves
Assessment DraftingAssessorsFills the assessment type's template headings from the review; gaps say To be confirmedThe assessor edits and records the result as their own
Session Notes DraftingSecretary, chairDrafts the narrative notes from the session's own recordsThe secretary edits; the minutes stay derived
Evidence MappingPlatform adminSuggests control ids for unmapped Defender and Policy keysThe admin saves to pin
The decision form with its AI Review panel
Where AI meets the author: the Review Argument panel beside the decision form. With the feature off, the button explains why.
AI proposes, a named human promotes — no AI output changes state by itselfLanes: Person, Archigov, AI provider. Steps: Ask for a draft or a review (Person); Feature enabled? (Archigov); Button explains why; deterministic path (Person); Model call (AI provider); Validate; discard citations that do not resolve (Archigov); Edit, apply or discard (Person); Save? (Person); Stored as the person's act (Archigov); Nothing stored (Person). Transitions: Ask for a draft or a review to Feature enabled?; Feature enabled? to Button explains why; deterministic path (no); Feature enabled? to Model call (yes); Model call to Validate; discard citations that do not resolve; Validate; discard citations that do not resolve to Edit, apply or discard; Edit, apply or discard to Save?; Save? to Stored as the person's act (yes); Save? to Nothing stored (no).PersonArchigovAI providerAsk for a draft or areviewFeatureenabled?Button explains why;deterministic pathModel callprovider and model chosenper featureValidate; discardcitations that do notresolveEdit, apply or discardSave?Stored as the person'sactAI parts stay unverifiedNothing storednoyesyesno
AI proposes, a named human promotes — no AI output changes state by itself

#Providers

The platform admin chooses a provider and model per feature under Settings → AI: OpenAI, Anthropic Claude, Mistral, Ollama (local, no key needed), or any OpenAI-compatible endpoint. Pre-review can run on a local Ollama while decision review uses a hosted model. See AI providers.

#Data Handling

  • Keys saved in Settings are encrypted at rest with AES-256-GCM. A key supplied through the environment beats a stored one and never touches the database. Keys are never returned by the API, never written to the audit log, never exported.
  • What is sent is the context the feature needs: the review, design or decision text, the relevant catalogue items, and — for pre-review — the application's resources. Nothing is sent for a feature that is switched off.
  • Where it goes is the provider you configure. With Ollama, nothing leaves your network.
  • Every AI pre-review records the engine, provider and model in its data and on the timeline.

#When a Feature Is Off

Draft endpoints answer 409 when their feature is disabled, and the UI shows the button greyed with the reason. GET /api/ai/features tells any page which features are on.