Overview & workflow

K-AI Audit is the module that drives the document-audit workflow: verifying indexation, surfacing duplicates, answering mandatory questions, resolving conflicts, identifying missing subjects. The Document Steward animates the workflow day-to-day; Document Producers contribute through targeted reviews; LLM agents drive the same flow via the K-AI MCP audit-tools surface.

The Audit MCP exposes a 20-tool curated subset of the 47-operation REST API, organised conflict-first. Admin operations (instance setup, user management, stats) remain on the REST API.

The Clean discipline

K-AI Audit is the operational expression of the Clean discipline (the second of the three DKP capabilities — Govern · Clean · Activate). It implements Principle 2 (Clean before consume) and Principle 5 (Continuous observability): the platform never silently swallows inconsistencies, and every conflict, duplicate, or missing subject is materialised as a piece of work that a Steward or agent can pick up and resolve.

Conflict-first workflow

The recommended path through K-AI Audit is conflict-first. A Steward (or an agent) walks through five steps:

1. Land on the dashboard (audit_dashboard_get) to see open work across all audit instances.

2. Pick an instance with open conflicts; list them (audit_question_conflict_list).

3. Inspect a conflict (audit_question_conflict_get) — read the conflicting excerpts and the source documents.

4. Record the expert answer (audit_question_conflict_set_answer); the platform generates per-document modification recommendations inline.

5. Confirm modifications were applied in the source system; close the conflict (audit_question_conflict_close).

The same five-step shape applies to the broader-scope "full-audit" review (whose operations live under audit_full_audit_conflict_*), to mandatory-question follow-ups, and to missing-subject clusters.

Who uses K-AI Audit

• Document Steward — daily, via the K-AI Audit web app or MCP. The Steward animates the workflow: triages the dashboard, dispatches mandatory questions to Producers, resolves conflicts, marks duplicates managed, and decides on missing subjects.

• Document Producer — occasional, via mandatory-question prompts surfaced by the Steward. Producers are the SMEs whose authoritative answers feed the Document Products.

• AI agent — programmatic, via the Audit MCP. Agents typically run batch triage, propose draft expert answers, or generate companion documents for missing subjects, leaving final validation to a human Steward.

See Roles model for the full DKP role definitions.

Surfaces

• K-AI Audit web app (browser) — the full UX for daily Steward work, at https://audit.kai-studio.ai. See K-AI Audit web app.

• Audit REST API — 47 operations across 8 domains: instances, documents & indexation, duplicates, mandatory questions, questions & answers, conflicts, missing subjects, stats. Base URL https://api-audit.kai-studio.ai. See Instances and the following API reference pages.

• Audit MCP — 20 curated tools. A conflict-first subset of the REST surface, exposed at https://api-audit.kai-studio.ai/mcp. See the MCP tool catalog and the schema stability policy.

Base URL & conventions

• Base URL: https://api-audit.kai-studio.ai

• Auth: Bearer JWT (MCP and direct API) or kai_auth HttpOnly cookie (K-AI Audit web app). See OAuth 2.1 and Cookies.

• Methods: all POST, JSON request body.

• Response shape: every successful response is wrapped as {"response": <payload>}.

• operation_id convention: audit_<domain>_<action> (e.g. audit_instance_get, audit_question_conflict_set_answer). Used as the MCP tool name when exposed.

• Errors: standard HTTP status codes; body {"detail": "<message>"}.