Lessons 73.1 through 73.3 built a fast factory: a managed board, an evidenced angle, six disciplined variants. Speed is now the risk. A fluent model can quietly overstate a product, erase a price condition, invent a testimonial voice, or point a polished ad at a broken page — and paid distribution will amplify the error at cost within hours. This lesson installs the gate between production and spend: a practical preflight, a two-person rule, and an audit trail.

Why AI creative needs a gate

Model-generated copy fails differently from human copy. A tired human makes typos; a fluent model makes confident overstatements. It smooths away the awkward but load-bearing condition ("same-week, excluding formalwear"), upgrades "one sample filing" to "every filing ever," and writes testimonials in voices no customer ever used. Each reads beautifully and each spends money on a promise the business cannot keep.

The preflight therefore checks meaning before polish. Run every asset through these gates in order:

• Message accuracy: does the hook and body say what the brief approved, word for word on load-bearing claims? Any new adjective ("instant," "guaranteed," "all") must trace to the angle card's evidence.
• Product, price, and date truth: are offer, price band, availability, deadlines, and conditions exactly as the business can fulfill today? Check numbers against the booking sheet, coverage list, or terms — not against the draft.
• Evidence: is every proof asset real — actual sample, actual photo, actual customer words with permission? Fabricated before-and-afters, synthetic reviews, and stock photos presented as work fail here.
• Image, video, and on-screen copy: do visuals show the real product or work, is overlaid text legible and accurate, and does nothing in frame contradict the claim (wrong garment, wrong date, competitor branding)?
• Destination: does the linked page continue the promise per Lesson 72.2's message match, with one action working end to end? Click it as a stranger on a phone.
• Platform fit: does format, length, and safe-area framing suit the placement? Guidance observed in the TikTok Creative Center and comparable live executions visible in the Meta Ad Library inform this check — the claims remain yours alone.
• Accessibility: is contrast sufficient, are captions present and correct on video, does no meaning depend on color alone, and is essential copy outside tiny in-image text?
• Tracking: does each variant carry its distinct identifier, correct link or code, and counted conversion event from the metric sheet in Lesson 72.4?
• Approval owner: is a named human accountable for this asset's launch, with date and signature?

Two people, two lenses

Where possible, review with two people wearing two different hats. The maker checks craft: cut, copy rhythm, visual consistency, format correctness. A second person checks claim and customer interpretation: would a skeptical stranger read this promise, believe exactly what is meant, and find it true on arrival? The second reviewer has explicit permission to kill beloved lines. Solo builders can simulate the split by reviewing on separate days with the two checklists — craft first, claims second, never both in one skim.

Decisions are ternary, never vibes: pass (ships as is), change (ships after a specified edit, re-checked), or reject (does not ship; reason recorded so the error is not regenerated next week). Record all three outcomes. A review log showing only passes is not a review; it is a rubber stamp with formatting.

Exercise: review five assets

Assemble five assets — realistically, the control plus four variants from Lesson 73.3 — and document the review in AD-REVIEW.md:

• Asset list: ID, format, hook, destination link, and tracking identifier for each.
• Preflight table: one row per asset, one column per gate above, each cell marked pass, change (with required edit), or reject (with reason).
• Decision summary: which assets form the approved launch pack, which return for edits with named owners and dates, and which are rejected with the lesson learned.
• Audit trail: reviewer names, review date, brief version referenced, and the commitment that no unlogged edit ships.

A miniature outcome: V1 control passes after a caption fix; V2 hook variant is changed ("guaranteed" softened to the actual pickup-date promise); V4 demonstration clip passes and becomes the fatigue rotation; V5 creator clip is rejected because the speaker never used the product — a fabricated peer voice, however warm, never ships; V6 CTA variant passes with its tracking code corrected. Three shippable assets, two lessons recorded, zero mysteries for the media buyer or for Class 74's test reader.

Finish line: an AD-REVIEW.md with pass/change/reject decisions on five assets, an approved pack with traceable IDs, and a named audit trail — not a verbal "looks good."

Verify quickly: pick any approved asset and reconstruct its chain — brief promise, evidence source, destination page, tracking ID, reviewer name — from the documents alone. If any link requires asking someone, the trail is incomplete.

Common failure mode: reviewing for taste under deadline — kerning debated, claims unexamined — then discovering post-launch that the beautiful ad promises coverage, capacity, or pricing the business cannot honor. Polish never compensates for an untrue sentence.

Check your understanding

1. Why does model-generated creative specifically require a truth gate, beyond normal proofreading? 2. What do the maker and the second reviewer each own, and what three decisions can the review produce? 3. What must an audit trail contain so Class 74 can trust the approved pack?

Next

Class 73 closes with a creative pack that is evidenced, varied, and reviewed. Class 74 takes that pack into structured testing: Lesson 74.1 frames each round as one question with a hypothesis, control, and decision rule, so spend produces learning instead of noise.