# Wonder creative regression: what changed in the tooling

2026-09-10 · Existing Wonder task · File-backed investigation

The regression came from replacing individually composed art direction with a batch template, then trying to restore variety through fonts and decorations. The reference preference was available in the approved brief. The production and review process failed to enforce it.

## What the source establishes

1. **Original W41 and W42 were individually authored.** In `../2026-09-10-v4/render.mjs`, the brush concept uses large, heavy plum Archivo and a yellow highlight behind “Dread day.” W42 uses a condensed red Oswald hierarchy: opening phrase, smaller “WATCHING IT,” and dominant “SNAP.” Supporting copy and the product sit in a deliberate relationship. The original implementation was HTML/CSS captured with Playwright; it was not Remotion, although the later batches were. Both tools can render these layouts.
2. **The first expansion standardized the hierarchy.** `../2026-09-10-v11/qa/rejected-type-v1/index.tsx` explicitly says: “Font sizes are locked to the longest sibling headline, avoiding a second test variable.” Its sizing table reduces the opening lines to 43–63 px and emphasized lines to 78–101 px. The shared component uses headline/support/CTA blocks and a regular headline weight. This diluted the scale and weight relationships of the originals. A controlled copy test was prioritized over preserving the accepted visual character.
3. **The photography was also divided into template zones.** `../2026-09-10-v11/prompts/A2A-45.txt` requests the upper 39% empty and the objects in the lower 58%. Other prompts similarly reserve substantial text space. These instructions help readability but also encourage a repeated “copy above / subject below” composition. They are a contributor to sameness, not evidence that every resulting photograph is bad.
4. **The next revision substituted variation for art direction.** In `../2026-09-10-v12/remotion/index.tsx`, the A2A concept changes type and placement by variation index. The `arc()` helper draws a fixed SVG loop, rather than fitting an annotation beneath the rendered word. On the brush execution it is placed at fixed coordinates around the large handwritten phrase. This explains the awkward circle: it was a programmed decoration, not a necessary part of the original hierarchy. Other variants vary alignment, fonts and labels without a measured reason tied to the photograph.
5. **The review checks were insufficient.** Text-box checks detect overflow and canvas clearance. They do not judge hierarchy, image/copy balance, or whether a decorative stroke competes with a word. `../2026-09-10-v12/qa/visual-audit.json` recorded a producer PASS after contact-sheet and targeted checks, but the process did not require a direct quality comparison against the retained favorites. The operator subsequently rejected the result; `../2026-09-10-v12/qa/operator-rejection-20260910.md` withdrew that creative acceptance.
6. **The displayed scores overstated the review's specificity.** `../2026-09-10-v12/build-board.py` contains a concept-level score map with a visual score of 8 for all ten concepts, copied to all three siblings and both formats. Those were not sixty separate visual assessments. V13 does not assign new numerical scores.

## What the investigation does not establish

There is no evidence that an engine update caused the regression: the v11 and v12 packages use the same Remotion 4.0.523 and React 19.3.0 versions. The older HTML renderer versus the newer Remotion renderer is an implementation difference, not a demonstrated cause. The layout instructions and source code are the concrete differences.

The competitor archive contains 166 unique Foreplay records across eight brands and 122 captured images. This revision reuses that archive. It does not claim a fresh sweep, access to competitor ROAS, or proof that long-running ads were profitable. The failure was not an absence of any reference material; it was failure to translate the reference hierarchy into the expanded batch and hold that batch to the original benchmark.

## V13 correction

- Restore the original W42 and W41 photography for the SNAP and brush concept families, with the original condensed-red and heavy-plum/yellow visual hierarchies as anchors.
- Compose each of the ten concepts for its photograph. Use restrained serif, condensed and bold treatments with highlights attached to the actual words. Remove the floating circles and ornamental arrows. The small CTA direction glyph is retained as part of the link treatment.
- Keep each three-hook family coherent; adapt line breaks and scale when copy length requires it. These are hook-direction comparisons, not a perfectly isolated copy experiment.
- Bring directions and CTAs back into readable groups. The first twenty-render review identified directions over a comb and support over the lavender product scene; those executions were corrected before completing the batch.
- Review all sixty outputs in format-specific sheets, with larger checks for the original benchmarks and flagged compositions. Record producer observations, not an independent audit or a numerical quality guarantee.
- Preserve the exact four W41/W42 reference masters and six fictional Hair Diary masters separately. Preserve old routes and feedback. No new testimonial or verified-buyer claim is invented.

The changes are an accountable revision for operator review, not proof of advertising performance or a declaration that the operator will prefer them.