| date | learning | source |
|---|---|---|
| 2026-10-06 | 10-01 S3 rate-them mirror passed assemble technically (freezes 0, dark 0, 86.0 s, 9 batches) but gemini_qa final failed 4/10: character inconsistency, 'poor lip-sync', confusing hallway scene (family on the floor), inconsistent product design. | output/cartoon-h3/dawnbands-s3-rate-them-mirror-2026-10-01/gemini-qa-final.txt + assemble-receipt.json |
| 2026-10-06 | S3 was never read: $0.40 / 10 impressions through 10-02 in the shared SCALE ad set 120249989091270232 ($0.18 lifetime in the 10-05 hub pull); moved 10-03 to Testing ad set 120250033290300232 (ad 120250033291530232). A shared ad set starves new concepts; give each its own ad set with a floor. | dawn-bottleneck-2026-10-03/audit/ad_level_30d.json; kalda-ads-2026-10-05-v2 active_ads.json; testing-restructure-2026-10-03/state.json |
| 2026-10-06 | S3 pack product_truth said 'never a screen' (obsolete: the display exists and is dark when idle) while intents 24/26 asked for an LED glow and three LED dots; the band audit flagged batches 7/9 for screen/glow and all four product batches (6-9) were regenerated. Use the product-truth.md sentence and never write a glow into an intent. | concept pack + gen/_failed-qa + project_launch_2026-10-01_beauty_format_cartoons |
| 2026-10-06 | The 1:1 Smooche mirror kept zig-zag scores (6/5/7/6/10), had no root-cause line and no stake beat, kept similes and sass ('pays rent', 'The audacity'), and an influencer joke line dragged at 2.31 wps (vo_check WARN); it also said 'Sixty day guarantee' instead of the 60-night storefront promise. | SCRIPTS-v4-mirror.md S3 + vo-check.md + product-truth.md |
| 2026-10-06 | Rung source = post-purchase survey 'already tried' (n=134): waking them myself 90%, phone alarms full volume 79%, loud alarm clock 58%, routines/bedtimes/sleep apps 48%, smartwatch 15%, bed shaker 12% (research-only), medication 3%. Tactile rungs and medication never appear in copy. | knowledge/brands/dawnbands/survey-intel-post-purchase-2026-09-04.md + feedback_dr_copy_anchoring |
| 2026-10-06 | Lane limits for this format: canvas prompts ban numerals and check_canvas fails readable text, so scores live in VO + captions only; title.py overlays one PNG from t=0 (no per-line score stamp); assemble has no per-beat SFX; gen_statics.py bans 'phone' and gates every static on ADHD + guilt wound + dawn band/vibrat/wrist/startle + calm payoff; no carousel builder and stage-session's review gate only runs for --type static. | scripts/cartoon_h3/canvas.py, title.py, assemble.py; scripts/statics/gen_statics.py; ~/ad-batch-hub/stage-session.py |
| 2026-10-06 | S3 used one combined ref sheet for 'Dad, the little sister and the dog' and framed Mom looking straight into the lens on her own lines; Gemini read the first as identity drift and the second as missing lip-sync. Two characters, one sheet each; presenter at the board, not to the lens. | output/dawn-beauty-format-2026-10-01/run_s3_asm.sh + storyboard.md beats 10/24 |
| 2026-10-06 | 10-06 audit: video now packargs + shot tiers + draft→final; the static ladder card is a NEW layout → gen_statics --new-layout --improvise --candidates (dry-run first), floor like every new static format; gen_statics adds its own 'Wake up guaranteed · 60 nights' pill unless the lines carry the exact '60 nights' (a '60-night' bullet renders two guarantees). | scripts/statics/gen_statics.py text_parts(); skills/dawn-statics/SKILL.md 10-06 upgrades |
| 2026-10-06 | starpop ai-static-ads-that-convert: a reasons/ladder static is expected to show lower CTR with better-qualified clicks; read the static ladder card on CPA and click→purchase, never cut it on CTR. Reasons-frame headline (brand voice) is a later one-variable test; first-person "I ditched" is a declared mom-first variant only. | ~/research/davidaistar/starpop/txt/ai-static-ads-that-convert.txt |
| 2026-10-06 | Ladder statics: declare a per-spec/per-static `chain` or waive `rule9` with a reason in `waivers`; gen_statics applies both on every path since 10-06, so the ladder card can render via --new-layout/--improvise/--recycle. | scripts/statics/gen_statics.py + voc_check.py (10-06 builder D, backups *.bak-2026-10-06x) |
| 2026-10-06 | Built 10-06: tiers auto-tagged (tiers.md) + run.py --tier-models; --draft 480P stage; h3.py FAL_PRICE fixed to the fal page (768P $0.08/s), so 9-12 batches ≈ $10-17 video. | scripts/cartoon_h3/run.py, h3.py, tiers.py (Builder B) |
| 2026-10-06 | ScoreCard Remotion comp BUILT 10-06 (name + bar/X/circle mark, numerals off by default, card at 16 %): `python3 scripts/remotion_stitch.py comp ScoreCard <props.json> <out.mov> --prores` = alpha overlay; SFX cue sheet layer built too (sfx.py). Not wired into run.py. | apps/orange-doc/src/FormatComps.tsx, scripts/remotion_stitch.py render_comp (Builder C) |
format-rated-solutions
The "rate everything I tried" ad. A presenter scores each failed wake-up method out of ten with an honest credit and a one-line, mechanism-linked reason it fails, climbing to the band at ten out of ten. Video is an animated mom at the fridge whiteboard rating her own attempts (cartoon-h3 narrator). The static is a brand-voice ladder card (dawn-statics).