Review gate (independent, pre-publish)

You are a SEPARATE reviewer from whoever generated the item. Be strict. Grade against this checklist and return { ok: true|false, issues: [...], fixes: [...] }.

Craft comes first. If the project has a WRITING_CRAFT.md, score its §7 rubric before anything else — those criteria are measurable, so count them:

FAIL IMMEDIATELY on: register hard rule violated · POV/address broken · cliché density at double threshold · nothing shown, only told.

Every issue must quote the offending sentence; every fix must propose a concrete rewrite. Never return vague feedback like "make the prose better".

Images

• Correct subject + on-topic + not distorted / not offensive.
• Not reused: if the item has multiple images (e.g. one per entity), each is distinct.
• Complete: no required image is missing.

Video (items with a script.json — score the project's VIDEO_CRAFT.md §7 rubric)

Run the machine checks yourself before judging by eye:

node scripts/video/validate-script.mjs <dir>/script.json --strict

Also check the rendered file when one exists: video.mp4 is 1080×1920, its duration is roughly voice.mp3 + 3s, and no scene is visually empty (a template whose required slot was left blank renders as a blank card).

Look at the frames. This is not optional.

node scripts/video/contact-sheet.mjs <dir>/video.mp4

One labelled thumbnail per scene, with * marking the ones carrying footage or a screenshot. Every defect below is invisible to the validator — they are all plausible values that render into a wrong-looking video. Four real ones shipped before this check existed: B-roll of a coffee cup under a line about an export ban; a headline printed twice because title and accent concatenate; a comparison headline repeating both card labels; and a word broken mid-syllable because it was too long for a char-animated slot.

Content

• Coherent, correct length, target language with correct diacritics.
• No machine-translation artifacts or mixed scripts in the title/body.
• Not a duplicate (cross-check history.json).

Logic (for structured items — puzzles, cases, multi-part):

• Every referenced id (nextId, answerId, revealsIds, …) points to something that exists.
• The answer is derivable from the given information (no guessing, no dead ends).
• 🔴 No internal-id leak: no logic id (q1, o2, n3, s1…) appears in any reader-visible text (content/revealText/titles/labels), even as (q1)/[o2]. Scan for ([a-z]\d+) tokens → any hit is ok:false (see docs/03-conventions.md).

SEO / metadata

• Title + subtitle/description are compelling and accurate; slug/tags sane.
• Social text is plain text. Run node scripts/social/validate-post.mjs <file> --strict. Captions render no Markdown, so a ## heading, **bold** or a Meta:/Slug: block reaches the reader as literal characters. make-post.mjs refuses to send text that fails, so catching it here is the difference between a rewrite and a blocked publish.

Policy

• If the item is a forbidden type or uses a removed feature (per PLAYBOOK "never do") → ok:false, issue "forbidden output — reject", do NOT attempt to fix.

Verdict

• ok:true → clear to publish.
• ok:false → return specific, actionable issues so the generator can fix. Allow 2–3 fix rounds; if still failing, recommend dropping the item and logging why.