skills/static-localization/references/performance-feedback.md

# Performance feedback loop

Production quality is the prerequisite; profitable acquisition is the objective.

## Preserve attribution

- Keep stable final filenames and ad-set numbers from production through upload — the
  filename grammar (`juicylucy` skill, `naming.json`) is the attribution key.
- Store the campaign manifest with the upload record.
- Never rename a live winner without preserving the original identifier in the result
  export or experiment log.
- Record language, hook, visual concept, batch, audience, placement, and launch date.

## Collect comparable outcomes

From the business's trusted reporting source, where available: spend and impressions;
clicks, CTR, CPC; landing-page views; leads, trials, purchases, or the campaign's real
conversion event; conversion rate, CPA/CAC, revenue, ROAS when attribution is
reliable. Do not declare a winner from tiny samples, and do not compare languages with
materially different audiences, budgets, placements, or optimization events as one
test.

## Learn at the creative level

Group results by stable filename fields and annotated concepts. Look for repeatable
patterns in: hook framing; benefit axis (as named in the brand's `copy-patterns.md`);
human versus object/landscape visual; copy density and CTA; language-specific phrasing
and script; source creative family. Separate production defects from market feedback —
a low performer with unreadable copy is not a clean test of its angle.

## Feed winners into the next run

1. Identify winners only after the agreed evidence threshold.
2. Preserve the winning promise and hierarchy.
3. Create controlled variations in one major dimension at a time.
4. Update the brand's `copy-patterns.md` or `history/` with supported learnings, not
   guesses — dated evidence goes to `history/`, standing patterns to `copy-patterns.md`.
5. Keep compliance and product-truth constraints even when a risky claim appears to
   attract clicks.

Optimize for profitable customers and durable learning — not asset count, CTR alone,
or short-lived policy-risk tactics.