| date | learning | source |
|---|---|---|
| 2026-10-04 | QA caught 'you baby him' labelled as the husband's verbatim line; 0 hits in every corpus. Pronoun swaps, 'Mam' -> 'Mom' and a reordered quote also sat inside quote marks. Verify every quote before drafting; trim, never paraphrase. | output/dawn-iterate-2026-10-04/copy/qa-report-v1.md |
| 2026-10-04 | Lyric red-team v3: 4 BLOCKING, 3 of them created by the previous round's fixes (hook contradicting the body, root-cause line undercut). Every fix gets re-checked in a 'previously fixed items' table. | output/dawn-iterate-2026-10-04/copy/lyric-redteam-v3.md |
| 2026-10-06 | voc_check.py loads 54 files but misses the survey (glob empty), Okendo (path doesn't exist), voc.jsonl, bottleneck comment JSON and curly apostrophes. Stake-bank CPS line scores 0 in voc_check and 1 in 199olsy.json. Use the extended verifier; glob fix needs Fish's go. | scripts/statics/voc_check.py test run 2026-10-06 |
| 2026-10-06 | Okendo store reviews are not customer voice: 2,026 of 2,142 predate the store's first order and templated phrases repeat across reviewers. | output/dawn-bottleneck-2026-10-03/E-shopify-nonbuyer.md §6a |
| 2026-10-06 | Arctic Shift title search answered 'Timeout. Maybe slow down a bit' twice, then returned data after a 12 s wait. Reddit .json / old.reddit / WebFetch remain blocked. | curl test 2026-10-06; memory reference_scrapling_reddit_access.md |
| 2026-10-06 | Turn-stats snippet on the 10-01 attendance-office dialogue pack: CLERK ran 6 lines in a row with 18-word turns. Dialogue doctrine caps a speaker at 2 consecutive lines per scene. | output/concept-packs/dawnbands-s2-attendance-office-h1-2026-10-01.md |
| 2026-10-06 | Audit: voc_search.py --brand dawnbands only reads the voc-*.md compatibility pointers (no voc-tagged.jsonl), so its ranking is meaningless for Dawn; use grep + the extended verifier. Humanizer bans every dash but mode-B interruptions end on a line-final dash: that one is the exception. ad_dissector also spends EL Scribe quota unless --no-scribe. | scripts/voc_search.py:41-57,248-254; ~/.claude/skills/humanizer/SKILL.md §14; scripts/ad_dissector.py --help |
| 2026-10-06 | starpop articles: brief fields without a source carry [assumed] and block drafting until resolved; TikTok comments via tikwm /api/comment/list are a free, dated Tier 2 source (saved as tt_<id>.txt; extended verifier now globs voc-pull-*/*.txt); one Table 1 persona per script; logic pass checks product first-mention vs the reveal beat and the word budget. | ~/research/davidaistar/starpop/txt/{ai-customer-research-ad-copy,how-to-build-a-claude-skill-for-ads,create-and-clone-competitor-ads-with-ai}.txt; tikwm test call 2026-10-06 |
| 2026-10-06 | Extended verifier folded into voc_check.py 10-06: explicit raw corpus (voc.jsonl, comment JSON, voc-pull-*/*.txt, survey CSVs), curly-quote normalisation, per-hit file, competitor/Okendo reported not counted, `--files`, pickle cache in ~/.cache/voc_check. CPS line 0 -> 1. | scripts/statics/gen_statics.py + voc_check.py (10-06 builder D, backups *.bak-2026-10-06x) |
workflow-copy-research
Research-to-script workflow for Dawn VIDEO formats (song, spoken drama, podcast, courtroom, whistleblower, street interview, timeline, rated solutions, mascot explainer, skit): an on-demand VOC pull mid-brief into two grep-verified tables (ICP/persona + offering, every row a verbatim quote + source + date), 3 sourced hooks per concept, then the script in one of four modes (narrator beat sheet, two-person dialogue/banter, song lyrics with the versioned red-team method, or the long-form free-value explainer skeleton), ending in logic pass, Opus red-team, fix and re-check.