Review gate

Review Gate

Overview

Match the review mechanism to the task profile. Verify checkable work directly against predeclared criteria; reserve hypothesis-first friction for consequential judgment where seeing an agent view first could anchor the human.

When to use

• Before a controlled output takes effect or any required release occurs.
• Before accepting, merging, sending, publishing, or deploying consequential work.
• At a consequential assistive or human-final judgment node.

When NOT to use: brainstorming; rough low-consequence material the human will independently rework; or routine autonomous output covered by the agreed spot-check schedule and not currently selected for review.

Friction modes

Mode	Behavior
full (default)	Run the applicable branch below and wait for explicit approve, revise, or reject. Required reviews always use full.
lite	Run a concise criterion/evidence check without waiting. Allowed only for confirmed autonomous work with blast radius 0–2 and no required review.
off	Skip optional review presentation only. Profile-, governance-, regulation-, permission-, security-, and host-required checks remain active.

Process

1. Read the profile and gate source. Record final mode, criterion at effect, blast radius, and why review is required or optional. Escalate an ineligible lite or off request to full.
2. Choose the review branch:
   • Verifiable branch (autonomous / controlled): show the result, named criterion, and evidence together. Compare directly against the criterion; do not manufacture a hypothesis-first ritual where an exact check exists.
   • Judgment branch (assistive / human-final): only when the node is consequential, show scope and neutral evidence first, then ask the human to record their own expectation, causal hypothesis, or weighting before showing mode-permitted agent analysis. In assistive, reveal the non-binding interpretation and compare divergence. In human-final, show only evidence and consequence simulations, without an agent-selected answer.
3. Preserve the mode boundary. In assistive, label any agent recommendation non-binding and require independent human weighing. In human-final, the human frames, weights, selects, approves, and owns the decision; the agent may supply information and consequence simulations but may not decide or execute.
4. Confirm load-bearing assumptions. Ask about the assumption that can invalidate the result, using a precise question rather than "looks good?".
5. Evidence every done-claim. Show test output, diff, query/source, or run log. Mark unsupported claims UNVERIFIED. Label claimed safety, competence, control, or effectiveness improvements without empirical evidence UNVALIDATED DESIGN HYPOTHESIS; do not turn software checks into impact proof.
6. Record assets/review-checklist.md. Include checks run and skipped, divergence where applicable, assumption status, and the outcome. In full, wait for approve, revise, or reject; silence is not approval.

Rationalizations

Excuse	Rebuttal
"Hypothesis-first is always more mindful."	On verifiable work it adds ceremony; compare the evidence with the criterion.
"The tests prove this design improves safety."	Tests prove software behavior, not real-world effectiveness.
"Off means there is no review."	Off cannot remove a required release or host check.
"A human click makes an unverifiable output controlled."	Controlled requires a criterion at the gate; a click does not create one.

Red flags

• Hypothesis-first imposed on a deterministic or exactly verifiable check.
• A conclusion shown first at a consequential judgment node.
• controlled approval without a named criterion and evidence.
• A software fixture presented as evidence of human or organizational impact.
• Required review bypassed through lite, off, or silence.

Verification

The gate worked when the review branch matched the mode; criterion-based work was checked against evidence; consequential judgment captured the human view before the agent interpretation; unsupported claims were labeled honestly; the mode boundary was preserved; and every required full review ended in an explicit human approve, revise, or reject before effect.