AI Governance

How AI assistants and agents participate in architectural evolution — their responsibilities, authority boundaries, and the human approvals they require. This is the governance counterpart to the AI Collaboration Model: that page describes how AI collaborates; this page sets the limits. Implemented behaviour is separated from future vision; the governing rule is AI proposes and drafts; humans decide and approve (Operating Principles O1–O3).

Authority model

graph TD
    subgraph MayDo["AI MAY (without approval)"]
        M1[Draft ADRs / digests / capability pages]
        M2[Generate diagrams & analysis]
        M3[Invoke skills; run deterministic tools]
        M4[Explain findings from structured facts]
        M5[Propose placements via the Decision Matrix]
    end
    subgraph NeedsHuman["AI MUST get human approval"]
        H1[Accept an ADR / change architecture]
        H2[Create / retire / archive a repo]
        H3[Publish a pack version]
        H4[Send anything customer-facing]
        H5[Promote a knowledge asset to canonical]
        H6[Resolve ambiguous capability ownership]
    end
    subgraph Never["AI MUST NOT"]
        N1[Auto-send external communications]
        N2[Mutate a consumed/immutable artefact]
        N3[Fabricate facts to fill gaps]
        N4[Overwrite HUMAN-marked regions]
    end

Per-actor governance

Actor	Responsibilities	Authority boundary	Required human approval	Status
Claude Code	Build/document repos & portal; draft ADRs; run skills & checks	May draft & generate; may not accept ADRs or change architecture unilaterally	ADR acceptance; repo create/retire; anything external	✅ Implemented
ChatGPT	Draft content that drops into templates; advise	Same as Claude Code; chat-window outputs are proposals	Same	✅ Implemented (working rules in chatgpt.handoff)
Shared Skills	Provide deterministic tools + method the assistants invoke	Skills execute within an assistant’s session; no standing authority	Inherit the invoking assistant’s approvals	✅ Implemented
In-product AI (assessment narrative)	Explain rule-based findings/recs from structured facts	Explain only — never decides; stricter for regulated verticals	Admin approves the report before sharing	✅ Implemented
OpenClaw	(Future) governed agent consuming services	Must consume via service contracts, under one claims/tone guardrail	Human approval for any external action; scoped tool access	⏳ Planned
Hermes	(Future) learning loop re-weighting intelligence	One-way: publishes a new pack version; never mutates	Human review before a re-weighted pack is adopted	⏳ Planned
portfolio-portal-orchestrator	(Future) regenerate derived docs	Writes only `<!-- GENERATED -->` regions; preserves `<!-- HUMAN -->`	Human review of structural diffs; can’t accept ADRs	⏳ Spec only

Approval gates AI cannot bypass (evidenced)

Nothing auto-sent — external communications require human approval (O2; outreachagent “golden rule”).
Human owns canonical intelligence — no asset becomes canonical without founder approval (FIP “human approval mandatory”).
Deterministic decides, AI explains — customer-facing recommendations are rule-based, not AI opinion (Principle 11).
Evidence-first — AI records uncertainty; it does not invent to fill gaps (Principle 8).
Immutable regions & artefacts — AI preserves  regions and never mutates an immutable pack/run.

As autonomy increases

The human approval boundary is preserved as automation grows: future agents (OpenClaw, Hermes, orchestrator) expand within gates — consuming services, drafting, regenerating — but capture, review, approval, and architectural governance remain human. New autonomous behaviour is introduced only via an ADR that defines its authority boundary explicitly.

Governance checklist for introducing a new AI capability

[ ] Passes the four-part test
[ ] Consumes service contracts, not venture repos (ADR 0005, proposed)
[ ] Its authority boundary is written down (may / must-approve / must-not)
[ ] Required human approvals are defined and enforced
[ ] Outputs are evidence-grounded; uncertainty is surfaced
[ ] An ADR records the decision