Meta ad pipeline

Who decides what

Blue marks the steps where a model decides. The other steps are a rule or a person. Three of the line's steps are blue.

The review pass is a model, so it is blue, and its verdict is still subject to the three-attempt limit. Segmentation is a model, drawn as a rule: it returns a box and a rule reads it.

Every choice is in the spec file

Each rendered ad carries a JSON spec. Nothing on the image is unexplained. These values are copied from specs/poc_7968200130623.json and the run manifest.

1. 1

   Headline ← the product page

   "500mg nitrate / in a 60ml shot." The model may only claim what it can quote. The quote has to appear verbatim on thefeed.com.

   copy_provenance.claim_source = "500mg of pure dietary nitrate in a 60ml shot"

   model gemini-3.6-flash · prompt c133edcb · attempts 1 · reviewed true
2. 2

   110px ← the house ad type scale

   One size for every ad in the batch. Copy has to fit it; nothing auto-shrinks. The 9:16 headline is 10.2% of the short edge, which reads at about 40px on a phone.

   type.headline.size_px_at_1080 = 110 · max_lines 2 · line_height 0.98 · tracking −0.03em

   measured: "500MG NITRATE" 748px · "IN A 60ML SHOT" 748px · limit 920px
3. 3

   Support line on ← the subhead gate

   The vendor name goes on only when the label is too small to read on a phone and there is room above the product. Here the product is 39.3% of the frame height.

   support_decision.shown = true

   support_decision.reason = "label too small to read on a phone, and there is room"

   product_share 0.393 · product_top 714
4. 4

   Product box ← segmentation, 60px of air

   The same background-removal model already cached on the Mac finds the product. The type block ends at y560; the product starts at y714; the rule needs 60.

   product_bbox = [406, 714, 692, 1468] · layout.block_top 260 · block_max_width 920
5. 5

   Navy wordmark ← sampled brightness

   The pixels under the wordmark are averaged. Below 128 the mark is white, above it navy. No lime, by my call on 22 Oct.

   logo.region_luminance = 160.4 → color #09023d · width 300 · bottom 290 · lime_uses 0

What failed

Two sources, labelled. The first set is a proof of concept from 22 Oct where an assistant wrote the headlines and layout by hand. Those renders are superseded. The second set is what the finished process caught on its own during the real run, straight from copy_log.jsonl.

First pass, before the checks existed

Fuel and hydration, one bottle

pack-word check · review pass

It is a sachet. "bottle" is now only allowed when it matches the form in describers.csv.

Recovery, strapped on

double-meaning rule

Flagged in review. The lint regex now lists the phrase by name.

Chest strap accuracy

concreteness check

A category claim with nothing to quote. The real run wrote "Real-time heart rate tracking" from the product page.

A gel / you can drink

line-balance rule

The shorter line must be at least 55% the width of the longer, measured in Futura.

Carbs and caffeine, / one gel

line-balance rule

Same rule. Also the headline was set at a different size from its neighbours.

Tart cherry / for recovery

segmentation + 60px clearance

Type on the pouch. Position now comes from u2net, with 60px of air required.

Cherry shot, / post-session

segmentation + 60px clearance

Same failure. The edge heuristic missed low-contrast packs; segmentation does not.

Just carbs, / nothing else

product-too-high routing

Type over the product. The real run refused this one entirely; see below.

Caught by the process during the real run

Increases blood / flow by 400%

Speeds recovery / while you rest

review pass · double meaning

Attempt 1 was quoted correctly from the product page and still failed. The reviewer flagged a second reading. Attempt 2 passed.

Cut inflammation

Less soreness

fit at 110px · 922px against 920

Two pixels over the line limit, as the log recorded it. The rule does not round.

Enhance strength

1.5g per gummy

fit at 110px · 922px against 920

Same width, same rule. The rewrite moved from a benefit claim to a spec claim with a quote behind it. All three rejections are in the log at the end of the page.

product top y426

25% rule · y480

Not rendered. Routed back.

Routed back, automatically

Carbs Fuel Original Mini. Segmentation put the product top at y426. At 110px, not even one headline line fits above it with 60px of air. The gradient pipeline's own placement rule says the product top stays below 25% of the frame, y480, so this is the gradient's error, and the fix belongs there.

The rule is that products that sit too high are not squeezed. The product went to copy_needs_review.json with a note to regenerate the gradient lower. 19 of 20 rendered.

The nineteen ads

Manifest order. Every headline below was written by the process, from a quote, and checked the same way. Type size is identical on all nineteen.

Counts, cost and limits

Cost

About $0.15 per gradient image, plus a few cents of copy per product. Copy ran one attempt for 16 products and two for 3. Carbs Fuel never reached the copy step.

What stays human, and why

• Product clipping. Automated cutouts were tested on 22 Oct and rejected. The edges were not good enough. Clipping stays in Photoshop or with a vendor.
• Health claims sign-off. The quote check proves the claim exists on the product page. It does not prove the claim should run.
• Contact-sheet approval. Every run ends on a contact sheet and a person looks at it.

Known copy weaknesses from this run

• It ran on gemini-3.6-flash because Pro was not available on the key.
• Several headlines lift the vendor's phrasing almost whole. "Lightweight at just 47g" is the product page's "Lightweight: Just a mere 47g."
• Some line breaks end on "of" or "to": "38 hours of / battery life", "120 hours of / active tracking", "Absorb up to / 90g of carbs".

Fixes for all three are specced in the prompt and lint rules and have not yet been run. The manifest records model, prompt version and attempts for every ad, so the next run can be compared to this one line by line.

Not yet wired

• 4:5 and 1:1 cuts. The gradient pipeline outputs 9:16 only.
• A real sales ranking. top20.txt is the first 20 top-100 SKUs with a QA-passed gradient, in manifest order.

One command