DawnBands — Creative Test Readout

Angle Scoreboard (30 days)

CTR/CPM excluded from verdicts on purpose — CPA is the only decision metric. (Long-form natives show 20%+ "CTR" because text-expansion clicks count; it means nothing.)

What's Working

1. ADHD Convenience/Relief — the account backbone, but concentrated

80% of all spend runs through one angle family, and mostly through three creatives: slop 4 ($4.6K, 136 purchases, $26–38 CPA), slop 9 ($2.2K, 66, $21–42), slop 7 ($2.3K, 64, $26–116 depending on ad set). It's earning its budget. Two flags:

• Fatigue drift: family CPA moved $34 → $37 in the 7d window. Not a kill signal, but the refresh batch should already be in flight.
• Duplicate-stacking variance: the same slop 7 creative runs $26 in one ad set and $116 in another. Stacking one creative across 3+ ad sets splits learning and buys the same audience against yourself.

2. Offer / risk-reversal — quietly the best CPA-per-dollar in the account

$33 CPA over 30 days, $15 CPA over the last 7 on the residual spend it gets. This angle has converted across every window since March and has never been given a real budget. Most underfed asset in the account.

3. Fable humanized witness natives (7-19 launch) — best early read of any July batch

Batch CPA $24 = 33% under target at day 3. Caveats before celebrating: it's day 3 of a 7-day read, and the batch's result is concentrated in one variant — Meta chose v10 and fed it. That's still meaningful (the algo picking a winner fast IS signal under Andromeda), but v10 is the finding, not "long-form natives" as a class. The humanizer + witness-mode + voice-calibration stack shipped this batch; it's the first batch where the full feedback loop (prior-batch data → new brief → humanize → launch) ran end-to-end, and it produced the account's best sub-$25 CPA cluster.

4. Sniper P1-B blame-removal — thin positive

sniper-b2-kara-composite: $41, 2 purchases, $20 CPA at day 2. Watch, don't conclude.

What's Not Working

Older-adult angle — weak prior, not falsified

12 executions, $280 total, 2 purchases, $140 CPA. Direction is consistently bad across multiple creatives, but $280 across 12 ads doesn't meet the falsification bar (that would need ~$750+ of consistent feeding). Verdict: weak prior — stop feeding it passively; it only earns a retest with genuinely upgraded execution and a protected budget, or drop it.

Independence statics — same tier

$137 / 1 purchase. Thin and negative. Note the portfolio already knows independence converts as a listicle/lander angle — the static-image expression of it is what's failing. Angle ≠ format: retest independence with the fable-native format before writing the angle off.

Legacy creative stack — fatigued, extract don't revive

27 pre-May ads re-stacked across ad sets: $528 / 9 purchases / $59 CPA. Doctrine call: don't revive fatigued creatives — extract the angles that made them winners (wakeforcollege = future-independence, notdiscipline = blame-removal) and rebuild them in the currently-winning formats.

The structural problem: most "tests" were never tests

How To Structure Future Tests (feedback-loop framework)

Shaun's core principle — every batch is built with intent from the data of the previous batch — only functions if every batch actually gets read. Structure to guarantee reads:

1. One batch = one question. Angle first, then hook, then format. Never mix ("is fake-UI a format we can use?" and "does blame-removal beat convenience?" are separate batches — the current Sniper launch did this correctly, keep that).
2. Minimum viable read per concept: ~$110 (3× target CPA) before it has earned ANY verdict. Budget the batch backwards from that: $750/day test budget ÷ $110 = max ~6-7 new concepts per weekly cycle actually readable. Launching 27 at once = 27 lottery tickets, not 27 tests.
3. Cap concurrent test ads. 10-15 live test ads max in the test CBO. New concepts queue instead of diluting.
4. Fixed read cadence: no verdict before day 4, no kill/scale decision before day 7. Then 3-day rolling: above target + strong attribution → graduate; below → reduce 30-50%; zero spend for 7-14 days → kill slot, re-queue next concept.
5. Graduation bar: ≥$110 spend, CPA ≤ $36, ≥3 purchases → duplicate into the winners stack. "Graduated-" prefix rules still apply.
6. Kill = extract. Every killed creative logs its angle + hook + why it died. That log is the input to the next batch brief. A winning angle should spawn 5-10 iterations before it's considered mined out.
7. One creative, one ad set. Stop running the same creative in 3+ ad sets ($26 vs $116 CPA on identical creative = self-competition). Consolidate per Andromeda: winners stacked in ONE ad set.
8. Never pause — reduce. Standing rule, restated because fatigue management is about to matter on the slop family.

Recommended Next Steps (in order)

1. Read the two ACTIVE batches (fable natives, Sniper P1-B) at day 4-7 — 7-24 through 7-27. Nothing gets killed or scaled before then. Note the rest of the July tests (P1-A fake-UI, both slop tests, fable natives 2, B2) are currently PAUSED mid-read — frozen UNREAD forever unless re-queued. Don't resume them all at once: feed them back through the restructured test slots (#5) a few at a time so each earns its $110 read. P2-C is a deliberate hold to 7-26 per its ad set name — leave it.
2. Brief the fable-v10 iteration batch now. The outcome-proof / college-retrospective angle at $24 CPA is the best new signal in the account. 5-8 new witness natives iterating that angle (different narrators, same retrospective-proof spine), full stack: witness mode → humanizer → voice calibration. Ready to launch the day v10 confirms at day 7.
3. Brief the convenience-angle refresh batch in parallel. The backbone is drifting ($34→$37) and the account has no hedge if it fatigues. Same angle + mechanism, new hooks and new visual formats — including a long-form native expression of the convenience angle (the account's best format × the account's best angle has never been tested together).
4. Feed the offer angle. It's at $15 CPA on 7d and has been at-or-under target for months on scraps. Move it into a properly funded ad set and brief 3-4 fresh offer-forward variants (guarantee-first, price-anchor, risk-reversal openers).
5. Restructure the test CBO per the framework above — cap live tests, min $110/concept, queue the rest. Re-queue the best 5 B2 concepts (chosen by concept quality, not their meaningless $2 reads) into the first proper cycle.
6. Consolidate duplicate creative stacking — reduce (not pause) the redundant ad-set copies of slop 4/7/9, keep each creative's best-performing placement.
7. Retire the older-adult angle from passive spend. Reduce to floor. It comes back only as an upgraded-execution retest with protected budget if there's a strategic reason to want that audience.
8. Log this readout as the feedback-loop input for both new briefs (#2 and #3): winning angles (outcome-proof, convenience, blame-removal, offer), dead expressions (older-adult statics, independence statics, fake-UI pending day-7), format learnings (humanized witness natives = best new-batch CPA; statics = proven scale vehicle).

Round 2 — Winning-Combination Isolation Test (approved direction 7-22)

Day-3 fable signal is concentrated in fable-v10-zero-missed-8ams ($260 / 11 purchases / $24 CPA). v10 differs from the batch's losing variants on four axes simultaneously, so the win is real but the driver is unidentified:

Format (long-form witness native + humanizer + voice calibration) is held constant across all arms — it's the proven container, not the variable under test.

Test arms (12 new variants + marksman parallel arm — one variable isolated per arm, volume added WITHIN arms so Andromeda picks the best expression of each dimension)

1. Arm A — Angle robustness (4 variants). Retrospective-proof structure ported to NEW avatars/desires: (A1) mom of 16yo son — first summer job, zero missed shifts; (A2) partner/spouse buyer — "I sleep through the night again," told from the after; (A3) dad narrator — son's first semester of trade school, never late; (A4) mom of younger teen — full semester attendance streak. If Arm A hits → the retrospective-proof angle is the driver and it's avatar-agnostic.
2. Arm B — Desire isolation (3 variants). College/independence desire WITHOUT the retrospective structure: (B1) forward-looking college countdown (kitchen-table brochure moment, "who wakes her up there?"); (B2) first-week-at-college wobble→save, told mid-story; (B3) senior-year "two years left to fix this" forward frame. If B ≈ A → the independence desire is the driver. If B flops while A hits → the retrospective structure itself is the driver.
3. Arm C — Register recut (3 variants). The v01, v07, and v03 stories — all thin performers — recut to open at the victory and flash back. Same narrators, same beats, one flipped variable. Cleanest single-variable test in the batch: hope-forward vs pain-forward with story held constant.
4. Arm D — Compression probe (2 variants). 150-250w short-form cuts of v10's angle. Tests whether the angle survives outside the long-form container.
5. Arm E — Marksman self-purchase (ALREADY LIVE, launched 7-22). Self-voice college/young-adult natives (n1-n3) in ad set "Marksman Self-Purchase Natives 7-22-2026" → new /pages/adhd-college-alarm LP. Reads as the avatar-dimension arm (self-buyer vs mom-narrator) on the same day-4-7 window. Round 2's Arm A deliberately excludes the adult self-buyer avatar to avoid double-testing this lane.

Volume rule: extra variants go INSIDE arms (all share the tested dimension — Andromeda concentrates on the best one, and the arm verdict strengthens). Never as extra thin ad sets — that's the pattern that produced 172 unread ads. In testing, the binding constraint is dollars-per-read, not creative count.

Launch structure

• New ad sets inside the existing Copy Test CBO (never a new campaign), one ad set per arm, labeled R2-A/B/C/D Isolation 7-2X-2026.
• Budget: ~$250/arm/week floor across A-D + feeding marksman ≈ $1.2-1.5K committed before any verdict; 5-7 day read window. No kill/scale calls before day 4. Arm-level budget is what buys the read — variant count within an arm is free.
• Read logic at day 7: which arm(s) produce ≤$36 CPA identifies the driver → Round 3 pours iterations into that dimension only. v10 itself keeps running untouched in its current ad set as the control.
• All copy through the full stack (witness doctrine → humanizer → per-speaker voice calibration → qa-gate) and operator review before anything is uploaded to Meta — standing rule.
• Images: Pinterest-reference lane via Higgsfield, unchanged. Real amateur phone photos sourced from Pinterest (logged in references/phone-shot/sources.json per batch) passed via --image-references to GPT Image 2 on Higgsfield, quality medium, 2K, 1:1. Product-bearing scenes: phone-photo ref first, product reference second. No Soul/Flux/Kontext/Nano-Banana chains for this lane (per 7-18 pilot baseline decision).

Sources: Meta Graph API insights (level=ad, last_30d/last_14d/last_7d, pulled 2026-07-22, system token, read-only) · campaign-state.md · media-buyer skill (Zakariah + Andromeda) · Shaun Eng feedback-loop philosophy per workspace Discord archive. Verdict tiers per the 2026-07-20 account-autopsy correction: FALSIFIED requires real spend + consistent direction; starved tests are UNREAD, not failed.