The pipeline from Lesson 62.2 can generate a finished-looking video with a false claim, an unreadable caption, and a broken link — all rendered beautifully. AI generation is fast; AI judgment is not. This lesson installs the human gates that make a complete piece trustworthy.

The four passes

Do four separate passes, not one big watch-through. Each pass has one question and one tool — eyes, ears, or links. Separating them is the point: when you check everything at once, you check nothing.

1. Message + facts. Does it say what the brief promised, and is every claim supported? Watch with the brief and script beside you. Pause on every number, name, and causal claim ("because," "causes," "proves"). Each must trace to the brief's evidence lines. Anything new gets cut or sourced — never shipped on a generator's authority. 2. Visual + brand. Do the shots match the system and the story? Check palette, fonts, character consistency across cuts, and logo placement against VISUAL-SYSTEM.md. Then check meaning: does the visual show what the narration says? A line about "state between steps" over a stock clip of a rocket launch is a continuity failure, however pretty. 3. Audio + captions + accessibility. Listen on headphones and on a phone speaker; read the captions with the sound off. Narration must be intelligible, music must sit under the voice (not over it), and captions must match the spoken words exactly. Then run the accessibility rules below — they are part of this pass, not an optional extra. 4. Publishing + delivery. Click every link, read every title and description, confirm files and formats. This pass uses the PUBLISHING-CHECKLIST from Lesson 62.2 and previews Lesson 62.4: correct aspect ratio, correct length, CTA verbatim, UTM present, thumbnail legible at small size.

Log each pass with a verdict — PASS, FIX, or BLOCK — and the timestamp. A FIX returns to one pipeline stage. A BLOCK stops the release: false claim, inaccessible captions, or broken CTA link.

Accessibility is a release gate

Accessibility rules are not polish. For many viewers they are the product: captions are how a Deaf viewer watches, contrast is how a low-vision viewer reads, alt text is how a screen reader describes the thumbnail. Apply these minimums to every piece:

• Captions: every spoken word captioned, synchronized, with speaker labels where needed. Auto-captions get a human correction pass — proper nouns ("Sonariq"), numbers, and the CTA must be letter-perfect.
• Intelligible audio: narration peaks clearly above music; no line is understandable only with captions compensating for a muddy mix.
• Readable text: on-screen text large enough to read on a phone, held long enough to read aloud twice.
• Contrast: body text meets a strong contrast ratio against its background (light text on the dark navy from our system passes; thin gray-on-gray fails — darken the box or thicken the type).
• Alt text: every published visual gets one to two sentences describing content and function, not "image of graph." Example: "Four-step loop diagram: think, act, observe, decide, with an arrow labeled state returning to think."
• No visual-only essentials: anything the viewer must know — the CTA, a URL, a key number — is also spoken or captioned. Nothing essential lives only in pixels that captions and screen readers never reach.

If a piece fails any of these, it fails review. Fix it in the edit or captions stage, not with an apology in the comments.

Claim review: strength, sources, qualifiers

Generators strengthen claims by stripping qualifiers. "May help in some setups" becomes "will transform." Your job in pass one is to restore the exact strength the evidence supports:

• Evidence strength: rank each claim — direct source you opened, reputable report, or generator assertion. Generator assertions ship only after you verify them elsewhere or cut them.
• Sources: every number and proper-noun claim needs one. Keep the source next to the claim in MEDIA-REVIEW.md so a later repurposing (Class 66) inherits the citation instead of inheriting the sentence without it.
• Qualifiers: carry the brief's qualifiers into the final ("for short scripted tasks," "when state is stored"). If a qualifier was cut for time, the claim must be softened to match, not left strong.

A useful test from Class 40 (unverified autonomous decisions): for each claim, ask "what would I show someone who doubts this?" If the answer is "the AI said so," the claim is not ready.

The matrix: score, then revise the weakest two

After four passes, score six dimensions from 1 to 5:

| Dimension | 1 (blocked) | 3 (shippable with fixes) | 5 (strong) |

|---|---|---|---|

| Message | Off-brief, CTA missing | On-brief, CTA present but soft | One message, verbatim CTA lands |

| Truth | Unsourced or false claim | Sourced but qualifier dropped | Every claim sourced + qualified |

| Continuity | Shots contradict narration | Minor mismatch, fixable in edit | Every shot serves its line |

| Readability | Text unreadable on phone | Readable but cramped | Large, held, high-contrast |

| Audio | Voice buried or harsh | Clear on headphones only | Clear on phone + headphones |

| Captions + CTA | Auto-caps uncorrected / link dead | Caps fixed, link works | Caps exact, link + UTM verified |

Rule: revise the two lowest scores before shipping, even if the average looks fine. Averages hide failures — a 5 on visuals plus a 2 on truth is a beautiful lie. Write the scores, name the two weakest, state the fix and the pipeline stage it returns to.

Exercise: write MEDIA-REVIEW.md

Review your 30-second explainer draft (script + shots + any generated sample) and create MEDIA-REVIEW.md:

# MEDIA-REVIEW.md — explainer-30s

- Pass 1 message/facts: FIX — shot 3 narration adds "never forgets"; brief allows only "carries state when stored." Fix: soften line, return to SCRIPT.
- Pass 2 visual/brand: PASS — palette matches; shot 2 diagram legible.
- Pass 3 audio/captions/a11y: FIX — CTA only on screen, not spoken. Fix: add spoken CTA, return to VOICE.
- Pass 4 publishing/delivery: PASS — links open, UTM present.
- Matrix: message 3 / truth 2 / continuity 4 / readability 4 / audio 3 / captions 3.
- Weakest two: truth (2), message (3) — revise script lines 3–4, re-record, re-score before export.
- Verdict: HOLD — no export until truth re-scores ≥4.

Finish line: a MEDIA-REVIEW.md with four pass verdicts, a six-number matrix, the two weakest areas named with fixes, and a HOLD or SHIP verdict.

Verify fast: re-score after the fix and confirm both weak areas rose. If you cannot point to the changed line, file, and timestamp, the revision did not happen. Common failure: scoring everything 4 to avoid decisions. A matrix of all 4s means the passes were skimmed — rewatch with the brief in hand.

Check your understanding

1. Why four separate passes instead of one careful watch? 2. Name three accessibility requirements that gate release. 3. A generator upgraded "can help" to "will guarantee." Which pass catches it, and what is the fix? 4. Your matrix reads 5/5/2/5/5/4. What do you revise?

Next

Reviewed media still needs to survive the real world of placements. Next, Lesson 62.4 packs the delivery — masters, variants, thumbnails, titles, captions, credits, and archives — so one finished piece ships correctly to blog, Shorts, email, and beyond.