---
name: ask-the-twins
description: Use two independent clean-context reviewers as an adversarial plan or verification gate.
version: 1.0.0
updated: 2026-08-14
---

# Ask the Twins

Use two capable, independent max-effort reviewers with fresh context. Anything flagged for review
must go to a clean-context twin. Give each the primary request, governing plan, relevant repository
artifacts, rubric, and evidence; do not give only the orchestrator's summary or substitute the
orchestrator's own reading.

Each reviewer returns `PASS` or `FAIL` plus numbered findings covering omissions, contradictions, safety, user stories, evidence gaps, and honesty risks. Reviewers work independently.

- Plan gate: both must pass the revised plan before a high-risk build begins.
- Reality gate: both judge screenshots, contact sheets, tests, and read-only data checks from the built result.
- The orchestrator inspects the evidence too; that is a third check, not a substitute.
- On any failure, integrate every finding, run a bounded fix wave scoped to those findings,
  regenerate evidence, and use fresh clean-context reviewers.
- Never silently skip a reviewer. Report model/tool unavailability and leave the gate incomplete.

Completion review begins only after the same build wave re-read the original human prompt,
phase-0 story, and final specification; ran its own E2E capture and regressions; validated distinct one-second frames;
self-scored every story and final-spec step against evidence; and wrote `docs/evidence/receipts/<item-id>.md` using
the canonical [evidence receipt schema](../tutorial-verified-done/references/evidence-receipt-schema.md).
A missing or invalid receipt is `FAIL`, not a request for reviewers to reconstruct evidence.

Technical convergence authorizes in-scope technical execution; spending, publishing, destructive changes, and human values decisions still require the user's authority.

## Exact-source convergence

Pin each review to the exact commit and evidence-byte hashes. Planning twins compare stable requirement IDs against the raw prompt, transcript, or microphone submission; any missing numbered requirement is `FAIL`. Completion twins compare all four one-second tutorial contact sheets and all four audio-derived narration transcripts against those same user-story IDs. A conditional finding identifies the exact required edit and is not a pass. Rebuild the evidence and obtain a fresh binary verdict on the new bytes. Convergence means both twins and the orchestrator inspected identical bytes; the implementer never owns green status.

Each completion twin writes its binary verdict into that receipt with `cleanContext: true` and the
complete evidence-path list it inspected. The two lists must exactly match the receipt's canonical
evidence list. Any twin `FAIL` starts a keeper rework wave limited to its numbered findings, followed
by fresh capture, regression checks, receipt validation, and fresh clean-context twins.