verify git:20261004.cbb0989A

Immutable. This exact content never changes and is served at /api/v1/blob/12cdf628bc27803a.

---
name: verify
description: Check an explainer before it ships — renderings against model.yaml, the video review sheet, the page in a browser, and a blind understanding test run by a fresh agent. Use after building an explainer, before publishing one, or when the user asks to check, review, or test an explainer (/verify <slug>).
---

# Verify — is the explanation right, and does it teach?

Run the checks in order. Stop and fix at the first failure. Record what you ran in `explainers/<slug>/review/understanding.md`.

**CLI.** `explainer <command>` with the plugin; `uv run explainer <command>` inside the Explainer repo.

## 1. Consistency with the model

```bash
explainer check <slug>
```

It fails when a rendering shows a number the model does not explain, omits a value that `require` lists, embeds an out-of-date model (`explainer sync <slug>` fixes that), leaves a claim incomplete or uncovered, or when `model.md` uses a function that no `why` claim justifies. A number that is right but missing from the model goes into `values` or `allow`. Never silence a real disagreement.

## 2. Each medium

- **Video.** `explainer review <slug>`, then read `video/review.png`: a frame at each line, bookmark, and predict pause, with the spoken text under it. Every layout issue it lists must be fixed. Check that each frame shows the evidence for its line.
- **Page.** Open `index.html` in a browser (Playwright when available). Use every control once. Check: no console errors, no horizontal scroll at 390 px, both light and dark themes, the predict gate hides its result until a prediction is committed.
- **Prose and diagrams.** Read them against the STE-80 rules in `../explain/references/writing.md`. Render each Mermaid block once.

## 3. Blind understanding test

Use this for anything that will be published. A fresh agent sees only one rendering.

1. Print the prompt: `explainer quiz <slug> --rendering <prose|diagram|html|video>`. For video, run `explainer review <slug>` first: the reviewer reads the captions and the review sheet.
2. Start a fresh subagent with the prompt text exactly as printed (the Agent tool, general-purpose). Do not add context. It must not read `model.md` or `model.yaml`.
3. Score its answers with `explainer quiz <slug> --rubric`: each general item 0–2, each quiz item 0–2. Count an answer that states a listed misconception as 0.
4. Pass: every quiz item scores 2 and no general item scores 0.
5. Write the scores and the reviewer's "gaps" to `review/understanding.md`. Fix real gaps in the rendering, then run the test again on the changed rendering.

A blind test that a deliberately broken copy also passes proves nothing. When you change the quiz, run it once on a broken copy (one key claim reversed) to confirm that the test can fail.