Skip to content

Human Review and Approval — Policy Update (GitHub Issue #2)

Objective

Revise docs/ai/human-review-and-approval.md so that an explicit, session-specific instruction from a human to execute a specific action is recognized as that human's own authority being exercised through the AI as a tool — while making clear the AI must never infer or extend that authority beyond the specific instruction, and must never treat a standing/autonomous instruction as substituting for contemporaneous human approval on high-risk work.

Scope

  • Edit docs/ai/human-review-and-approval.md to add/clarify a principle distinguishing:
  • Explicit, contemporaneous instruction ("do X now") → counts as the human's own act; the AI is a tool executing it, not an independent approver.
  • AI-inferred permission (extrapolating from a prior approval, a general "proceed" instruction, or standing autonomy settings) → does NOT count as approval for high-risk categories.
  • Check docs/governance/decision-rights.md for consistency; touch it only if the new language in human-review-and-approval.md would otherwise contradict it (currently they appear consistent — decision-rights.md already says "a prior approval of a different document, or a standing instruction to work autonomously, does not carry over").
  • Close/reference GitHub issue #2 in the work log once the edit is proposed.

Out of scope

  • Issue #1 (ADR-0007 / repository boundary / venture repo scaffolding tooling) — separate task.
  • Issue #3 (Shelfery food-safety claims ADR) — separate task, requires domain decision content this plan does not attempt to write.
  • Any change to status fields (this document stays draft; only a human can approve it).
  • Building new tooling or automation around approval recording (out of scope; this is a text clarification only).

Applicable authority

Files expected to change

  • docs/ai/human-review-and-approval.md

Assumptions

  • The issue's intent is a clarification of how approval is granted (explicit instruction = valid authority) not a loosening of which categories require approval. The issue text itself confirms this: "the ai should not autonomously make such changes. It needs human approval." Treating this as narrowing ambiguity, not removing a guardrail.
  • No change to the high-risk category list in decision-rights.md is implied or needed.

Risks

  • Classified as a policy change (per the issue's own "Type of change" field) — this is moderate/high-risk work per task-classification-and-routing.md and requires human sign-off before the document's status can move past draft.
  • Risk of the new language being read as broader than intended (i.e., as license for the AI to treat vague or past instructions as current authorization). Mitigate by being explicit that the instruction must be specific and contemporaneous to the action being taken.

Validation plan

  • python scripts/check.py
  • Manual review: confirm the edited document is internally consistent and consistent with docs/governance/decision-rights.md.
  • Human review/approval of the proposed text before status changes.

Progress checklist

  • [x] Read current docs/ai/human-review-and-approval.md and decision-rights.md.
  • [x] Write this work plan.
  • [x] Draft the edit to docs/ai/human-review-and-approval.md.
  • [x] Run python scripts/check.py (all checks passed) and python -m unittest discover tests (38 tests, OK).
  • [x] Present diff to user for review.
  • [x] User approved explicitly, naming this document, in-session (2026-07-12). Set status: approved, approved_by: Stuart M. Davis (Founder), approved_date: 2026-07-12, effective_date: 2026-07-12 per ADR-0011 / document-metadata-standard.md. Re-ran python scripts/check.py — all checks passed.
  • [x] Closed GitHub issue #2.
  • [x] Archived this plan.

Decisions required

  • Human (Founder) must review and approve the proposed text change before docs/ai/human-review-and-approval.md status can move to approved.
  • Whether to also close GitHub issue #2 or leave it open until formally approved — human call.

Handoff notes

Plan created and drafting proceeding in the same session. If picked up later: the proposed text edit has not yet been applied as of plan creation — check git diff / this file's progress checklist for current state.

Report an issue about this page