Rulebook

Evolve 07 (Iteration Playbook) + Zak scaling, with Fish's bars (locked 2026-10-12). Numbers live in config.json.

Data

• 7-day-click purchases and value only (Evolve + Zak: never scale on view-through). View-only purchases shown in brackets.
• Click share of purchases is the Zak gate: 70%+ scale aggressively, 50-70% cautiously, 30-50% don't scale, under 30% rebuild.

The four outcomes

• Breakthrough first 7 full days (or any later rolling 7 days): ≥15% of total account spend, so it competes with the proven SCALE ads, AND click ROAS ≥ the brand's breakthrough bar, AND the account's average daily spend the next 7 days is higher than the 7 days before.
• DawnBands: breakeven 1.31 (hub P&L 30d (ad-batch-hub build_matrix.compute_breakeven)), breakthrough ROAS 2.0
• Curvelle: breakeven 1.5 (PLACEHOLDER (Fish 10-12) until a Curvelle P&L exists), breakthrough ROAS 2.3
• Spend Winner ≥15% of account spend but misses the breakthrough bar. Under 0.5 ROAS it's flagged misleading (Evolve #12).
• Kpi Winner click ROAS ≥ the higher of breakeven and the account's own click ROAS over the same 7 days, with ≥2 click purchases.
• Loser everything else judged, including starved ads (under 1,000 impressions after 7 days: Meta's verdict).
• Pending under 7 full days old. Legacy launched before the data window; judged on health only.

Labels are sticky and only upgrade (Evolve 07). Copies that reuse one post are grouped; the best-labelled copy leads.

What each label gets

• Breakthrough: study it (9 components + hypothesis), never turn off over a bad week, iterate all types: 9 script iterations, 15 format iterations, watch lifespan risks #2 / #13 / #14.
• Spend winner: diagnose the ROAS input that lags the account (usually CVR). Fix belief #7, urgency #9, stakes #10, funnel #11 first, in the same format. Misleading #12 = reduce (Fish: never pause).
• KPI winner: leave running; iterate only on an execution issue (menu #1-#8).
• Loser: issues in order #1 hook → #8 offer. Ideation retries 2-3× only with new evidence; no iterations on losing imitations.

Ten-step learning (Evolve 07.02)

1. Batch number
2. Outcome + ad spend + campaign spend
3. Ad type: ideation / imitation / iteration
4. Hooks verbatim (hook/hold rate only past $1K)
5. Verify the 9 components vs the plan (awareness, valence+intensity, desire, avatar, angle, mechanism, belief, positioning, urgency)
6. Hypothesis + iteration ideas, cross-checked against other winners
7. Link the files
8. Add to the playbook
9. Pick iterations from the outcome's menu
10. Name it ITER#[N]_BATCH#[ORIG]

Fill components, type, hypothesis and parent in overrides/<brand>_annotations.json (keyed by ad id or post id). Atlas avatar/angle labels import automatically.

Zak overlays

• PDA: an iteration that keeps persona × desire × awareness is a variation, not new diversity.
• Kill vs extract: don't revive fatigued ads; extract the angle into new concepts with new hooks, formats, presenters.
• Never turn off main campaigns; reduce. Build creative diversity while scaling, not after the drop.
• LP EPC (CVR × AOV) multiplies every creative result.