# Memoryless Review Protocol ## Purpose Use memoryless subagents to verify completion without inherited assumptions. ## Input Packet Provide only: - Final plan - Acceptance criteria - Deliverable references - Minimal repro instructions Do not include: - Internal thoughts - Previous draft answers - Suspected fixes - Intended conclusions ## Multimodal Parity **Critical rule: verifier capability must match task modality.** - If the task involves images, screenshots, diagrams, UI layouts, rendered output, PDFs with visual layout, or any non-text artifact, the verifier MUST use a multimodal/vision-capable model. - Before spawning a verifier, check: 1. Does the task produce or consume visual artifacts? 2. Does the parent session use a multimodal model? 3. Does the candidate verifier model support vision? - If the answer to (1) or (2) is yes, and (3) is no, select a different model for the verifier. - Never strip images or visual context from the verifier input packet when the task requires visual reasoning. - When verifying image generation, UI rendering, layout correctness, chart output, or screenshot-based tests, the verifier must see the actual visual output, not a text description of it. ## Verifier Contract Each verifier must: 1. Restate the goal in one sentence. 2. List observable artifacts only. 3. Check each criterion independently. 4. Return `PASS` or `FAIL: `. 5. If `FAIL`, include: - exact unmet criterion - missing evidence - safest remediation hint ## Independence Rules - Verifiers must not share memory. - Verifiers must not read each other output. - Verifiers must not assume earlier steps passed. - Verifiers must treat missing evidence as failure. ## Escalation Rules - If two verifiers disagree, spawn a third verifier. - If conflict persists, treat as blocking and revise the plan. - Stop only when all verifiers independently return PASS. ## Exit Criteria A task is complete only when: - Every acceptance criterion is satisfied - Evidence exists for every satisfied criterion - All verifier reports are PASS