# Lyrics as a selling narrative: original (targ) vs Taras vs ours

Copy-chief comparison of three versions of the same advertorial, read as a SELLING NARRATIVE, with a
syllable/time budget audit. Measurements are separated from judgments throughout.

## 0. Sources and method

| Version | Text used | Timing used | Runtime |
| --- | --- | --- | --- |
| ORIGINAL (targ, English, Seora melasma cream) | 65 owner-approved caption cards, `runs/video-dna-v1/targ/analysis.json` → `captions` | caption start/end (native sheet, ±0.2 s) | captions end 188.5 s; video 194.3 s |
| TARAS (owner-preferred, HANBIT Arbutin 7.0 + TXA 4.0, pigmentation) | ASR 43 segments, corrected against on-screen captions in `dna/taras/sheets/sheet01..07.jpg` | ASR segment + word times | 194.2 s |
| OURS (goPure neck cream, film v07 / song-v02) | The SUNG text = `line-anchors.json` `expected/heard` (= `song-v02/stress/lyrics-with-stress-v04.txt`) | `masterVocalTime` per line | 239.7 s |

Measurement notes (important):

- The inputs call `lyrics_uk-v07.txt` "our lyrics (song-v02)". That file is the earlier prose selling text
  (447 words / 843 vowel-syllables, only partly stress-marked). It is NOT what was sung. The sung master, per
  `line-anchors.json` `heard` and our film ASR, is **401 words / 766 syllables** (ASR: 398 / 762). All
  "ours" numbers below use the sung text. The "843" figure in the brief is the prose script, not the song.
- Ukrainian syllables = vowel-letter count (а е є и і ї о у ю я). Exact for Ukrainian.
- English syllables = heuristic vowel-group count on the caption text → 584. The measured rate in the brief
  (~199 syl/min over 188.5 s) implies ~625. Both numbers are used; the conclusion does not depend on which.
- Taras ASR corrections applied from the burned-in captions: "Лоб, шок і підборіття" → "Лоб, щоки,
  підборіддя"; "серетину" → "ретинол"; "півментації" → "пігментації"; "все кислоти, вітамін" → "все: кислоти,
  вітамін С, ретинол"; "На жаль" → "Не жаль"; "їй 32" is confirmed by the caption ("їй тридцять два") — Taras
  actually sings 32, not the original's 34.

## 1. Headline numbers (measured)

| | ORIGINAL | TARAS | OURS (sung) |
| --- | ---: | ---: | ---: |
| Runtime (sung span) | 188.5 s | 193.9 s | 239.4 s |
| Words | 467 (ASR 471) | 310 (ASR 308) | 401 (ASR 398) |
| Syllables | 584 heuristic / ~625 measured | 626 (ASR 620) | 766 (ASR 762) |
| Sung phrases / lines | 65 cards (ASR 76 segments) | 43 phrases | 65 lines (1:1 with caption IDs) |
| Syllables per phrase | 9.0 | 14.6 | 11.8 |
| Seconds per phrase | 2.9 | 4.5 | 3.7 |
| Syllables per second | 3.10 (3.32 measured) | 3.26 | 3.23 |
| Syllables per word | 1.25 | 2.02 | 1.91 |
| Instrumental gaps ≥ 1 s inside the song | 3 (70.5, 92.0, 116.1) | 5 (21.1, 58.3, 64.5, 148.9, 188.7) | 0 |

Reading (judgment): the three versions sing at the SAME density (3.1–3.3 syl/s). Length is therefore driven
purely by syllable count. Taras lands on the original's ~620-syllable budget by spending only 310 Ukrainian
words (66 % of the English word count, because Ukrainian words carry 2.0 syllables vs 1.25 in English).
Ours keeps 401 words → 766 syllables → +45 s. No tempo trick was used or is available: to be 194 s long, the
sung text must be ~620 syllables.

## 2. Function map with timestamps, % of runtime, syllables, words

Function = dramatic/selling function. "t" = start time of the function in each version, "dur" = seconds,
"%" = share of that version's sung runtime, "syl/w" = syllables/words.

| # | Function | ORIGINAL t · dur · % · syl/w | TARAS t · dur · % · syl/w | OURS t · dur · % · syl/w |
| --- | --- | --- | --- | --- |
| 1 | Hook: meet ex's new wife, her age, pity look | 0.0 · 8.8 · 4.7 % · 29/24 | 0.0 · 10.2 · 5.2 % · 32/16 | 0.0 · 8.8 · 3.7 % · 29/13 |
| 2 | Backstory: March, 16 yrs, office colleague, cliché, effortless pretty, time | 8.8 · 14.2 · 7.5 % · 45/39 | 10.2 · 18.0 · 9.3 % · 55/33 | 8.8 · 15.9 · 6.7 % · 51/29 |
| 3 | Aged/stress self-diagnosis (5 yrs in 6 months, crying, kids) | 23.0 · 10.0 · 5.3 % · 38/33 | 28.1 · 7.2 · 3.7 % · 25/14 | 25.0 · 15.9 · 6.6 % · 50/27 |
| 4 | Store scene + failed alternatives #1 | 33.0 · 9.1 · 4.8 % · 40/27 | 35.4 · 12.0 · 6.2 % · 35/18 | 41.7 · 14.3 · 6.0 % · 51/25 |
| 5 | Humiliation: voice, scan, sting, smile, "before version" | 42.1 · 20.9 · 11.1 % · 76/61 | 48.1 · 31.3 · 16.1 % · 93/46 | 56.0 · 24.7 · 10.3 % · 85/45 |
| 6 | Car + "angry at what I'd seen" | 63.0 · 6.2 · 3.3 % · 23/20 | 79.4 · 8.2 · 4.2 % · 30/18 | 80.7 · 11.8 · 4.9 % · 29/17 |
| 7 | VISIBLE PROBLEM named | 69.2 · 11.8 · 6.3 % · 25/20 | 87.6 · 4.7 · 2.4 % · 18/8 | 92.4 · 9.2 · 3.8 % · 29/15 |
| 8 | Mentor social proof (older woman, flawless skin, "what do you use") | 81.0 · 18.8 · 10.0 % · 51/42 | 92.3 · 16.0 · 8.2 % · 59/26 | 101.6 · 19.6 · 8.2 % · 65/34 |
| 9 | Failed alternatives #2 + mentor verdict | 99.8 · 9.0 · 4.8 % · 32/24 | 108.3 · 9.3 · 4.8 % · 39/19 | 121.2 · 15.5 · 6.5 % · 40/20 |
| 10 | Product handoff + ritual instructions | 108.8 · 10.4 · 5.5 % · 32/24 | 117.7 · 9.6 · 4.9 % · 33/13 | 136.7 · 14.1 · 5.9 % · 44/24 |
| 11 | First application (feel) | 119.2 · 3.7 · 2.0 % · 13/12 | 127.2 · 4.5 · 2.3 % · 14/7 | 150.8 · 4.7 · 2.0 % · 16/7 |
| 12 | Result 1 (3 weeks, mirror) | 122.9 · 5.3 · 2.8 % · 20/17 | 131.7 · 6.7 · 3.4 % · 25/12 | 155.5 · 7.2 · 3.0 % · 25/12 |
| 13 | Daughter proof | 128.2 · 6.2 · 3.3 % · 19/13 | 138.4 · 6.1 · 3.2 % · 14/7 | 162.9 · 6.7 · 2.8 % · 25/12 |
| 14 | Result 2 (week six, even skin, no foundation) | 134.4 · 12.0 · 6.4 % · 32/25 | 144.5 · 10.7 · 5.5 % · 33/16 | 169.6 · 18.2 · 7.6 % · 61/31 |
| 15 | Rematch/payoff (recital, compliment, "thank you") | 146.4 · 30.0 · 15.9 % · 78/64 | 155.2 · 22.7 · 11.7 % · 78/36 | 189.1 · 31.6 · 13.2 % · 103/55 |
| 16 | Self-worth reframe | 176.4 · 6.2 · 3.3 % · 17/10 | 178.0 · 10.8 · 5.6 % · 29/16 | 220.6 · 6.1 · 2.5 % · 23/12 |
| 17 | Product name + CTA | 182.6 · 5.9 · 3.1 % · 14/12 | 189.9 · 4.0 · 2.1 % · 14/5 | 226.7 · 12.7 · 5.3 % · 40/23 |

All 17 functions are present in all three versions. No version drops a function; the differences are in
weight and wording.

### 2a. Syllable deltas per function vs the original

| Function | TARAS Δsyl | OURS Δsyl |
| --- | ---: | ---: |
| 1 Hook | +3 | 0 |
| 2 Backstory | +10 | +6 |
| 3 Aged/stress | **−13** (cuts "holding your kids together") | +12 |
| 4 Store | −5 (cuts "leggings") | +11 |
| 5 Humiliation | **+17** (spends here) | +9 |
| 6 Car | +7 | +6 (keeps "п'ятнадцять хвилин" = 5 syl for "fifteen") |
| 7 Visible problem | **−7** (cuts forehead + lip) | +4 |
| 8 Mentor | +8 | +14 |
| 9 Failed alt #2 | +7 | +8 |
| 10 Handoff/ritual | +1 | +12 |
| 11 First application | +1 | +3 |
| 12 Result 1 | +5 | +5 |
| 13 Daughter | −5 (cuts "good different") | +6 |
| 14 Result 2 | +1 | **+29** |
| 15 Rematch | 0 (cuts "a few seats", "by the way", "I meant it too") | **+25** |
| 16 Self-worth | **+12** (spends here) | +6 |
| 17 Name + CTA | 0 | **+26** |
| **Total** | **+42** (584 → 626) | **+182** (584 → 766) |

Measured: Taras's per-function deltas net out to +42 syllables — he PAYS for the humiliation and self-worth
expansions with cuts in "aged/stress", "visible problem", "store", "daughter", "rematch". Ours has no negative
cell: nothing is ever cut, and three functions (Result 2, Rematch, CTA) alone account for +80.

### 2b. Timing drift vs the original (start time of each function)

Taras: within +6 s through the store scene, then +16 s to +18 s late through the humiliation/car/problem
block (he spends 31.3 s on humiliation vs 20.9 s), then claws back: mentor +11, handoff +9, results +9 to
+10, rematch +9, self-worth **+1.6 s**, CTA +7. Ends at 193.9 vs 188.5 (+5.4 s). Judgment: Taras treats
the original as a timing grid with ±10 s tolerance per beat and re-lands the emotional peak (self-worth)
almost exactly on the original's clock.

Ours: +2 s at backstory, +8.7 at store, +14 at humiliation, +23 at visible problem, +21 at mentor, +28 at
handoff, +33 at result 1, +35 at result 2, +43 at rematch, +44 at self-worth, +44 at CTA. Monotonic drift;
nothing ever claws back.

## 3. Problem section vs solution section

Definitions: Problem = hook through "visible problem named" (functions 1–7). Solution = mentor through
Result 2 (8–14). Payoff = rematch through CTA (15–17).

| | ORIGINAL | TARAS | OURS |
| --- | --- | --- | --- |
| Problem | 0–81.0 s = 81.0 s = **43.0 %** | 0–92.3 s = 92.3 s = **47.6 %** | 0–101.6 s = 101.6 s = **42.4 %** |
| Solution | 81.0–146.4 = 65.4 s = 34.7 % | 92.3–155.2 = 62.9 s = 32.4 % | 101.6–189.1 = 87.5 s = 36.5 % |
| Payoff + CTA | 146.4–188.5 = 42.1 s = 22.3 % | 155.2–193.9 = 38.7 s = 20.0 % | 189.1–239.4 = 50.3 s = 21.0 % |

Measured: proportionally, ours has the original's SHAPE (43/35/22 vs 42/37/21). Taras skews 5 points more
toward the problem (he lets the humiliation breathe with two instrumental gaps at 58.3–60.1 and 64.5–66.4
and a 7-s "before version of the woman" phrase). Judgment: ours is not mis-proportioned; it is uniformly
~27 % too long in every act. That is a syllable-budget failure, not a structure failure.

## 4. Where the product is first shown and first named

| | ORIGINAL | TARAS | OURS |
| --- | --- | --- | --- |
| Pack first on screen | pink tube at 1:49 (109.0 s, 57.8 % of runtime), "Next day she put a pink tube"; SEORA lettering readable ~1:56–1:58 (`targ/sheets/sheet04`) | pink HANBIT tube at 2:06 (126 s, 64.9 %), after "рожевий тюбик" is sung at 117.7 s; product-hand close-ups 2:06–2:10 (`taras/sheets/sheet05`) | blue goPure jar at 2:17–2:18 (137–138 s, 57.4 %), on "Назавтра дала мені баночку"; lettering readable 2:28–2:29 (`film/sheets/sheet05`) |
| Name first SUNG | "The cream is called Seora" 182.6 s = **96.9 %** (function 17) | "Крем називається Ханбіт" 189.9 s = **97.8 %** (function 17) | "go pure. Я стала читати" 139.4 s = **58.2 %** (function 10); again "go pure — це крем для шиї" 226.7 s = 94.6 % |
| Between showing and naming | 73.6 s of unnamed "pink tube / Korean writing" | 63.9 s of unnamed "рожевий тюбик / корейські літери" | 1.5 s (jar shown, brand sung on the next line) |
| End card | none (link below) | 3:08–3:13 HANBIT card: "ЧИСТА ШКІРА. ВПЕВНЕНА ТИ." + 4 bullets (Вирівнює тон / Повертає сяйво / Працює м'яко / Корейська формула) + "ПОСИЛАННЯ НИЖЧЕ" | goPure end card (per film plan) |

Judgment: the original and Taras both use the same device — the pack is a mystery object ("pink tube,
Korean writing, something I'd never tried") for a full minute and the NAME is the last thing you hear,
right after the self-worth line. Naming is the payoff of the story, not part of the mechanism. Ours names the
brand at the handoff and reads the label aloud ("Крем для шиї. На чисту шкіру — тонкий шар, щодня…"). That
converts the mystery-object beat into a product-facts beat. Cost in syllables: 0 for the brand word itself
(it replaced "Korean writing"), but the label-reading that follows is +9 syl (functions 10: 44 vs 32).
Selling judgment: hearing the brand twice is a plus; losing the "mystery tube" suspense is a minus; the
net is neutral. It is not where our length went.

## 5. How each version phrases the PROBLEM and whether the camera can show it

| | Sung problem line(s) | Syl | Zone | What the camera shows |
| --- | --- | ---: | --- | --- |
| ORIGINAL | "Melasma on my cheeks / Patches on my forehead and lip / Like the light had gone out from inside" (69.2–81.0) | 25 | face: cheeks, forehead, lip (3 zones) | face close-up; pigment patches rendered on the heroine in every close-up before 122.9 s |
| TARAS | "Пігментні плями на щоках, наче світло згасло зсередини" (87.6–92.3) | 18 | face: cheeks (1 zone) | brown patches on cheeks in every heroine close-up from 0:28 to 2:12 (`sheet01–05`); mirror macro at 2:12–2:18 shows them faded; clean face from 2:27 on |
| OURS | "На шиї — дрібні заломи, / Лінії впоперек. Шкіра на вигляд в'яла. / Ніби світло в мені згасло." (92.4–101.6) | 29 | neck (1 zone) | neck crop with transverse lines at 1:32–1:33; hands at neck 1:35–1:39 (`film/sheets/sheet04`) |

Additional device in the humiliation scan (function 5), which pre-names the zones the diagnosis will use:

- ORIGINAL 47.2–50.8: "Forehead, cheeks, jaw / Two second scan"
- TARAS 60.1–64.5: "Лоб, щоки, підборіддя, дві секунди і все зчитано" (verbatim)
- OURS 62.7–67.0: "Її погляд ковзає вниз — Від обличчя до шиї" (rewritten to move the eye to the neck)

Problem noun → result noun echo (measured):

- ORIGINAL: "patches / melasma" → "The patches on my cheeks were lighter" → "The melasma was almost gone"
- TARAS: "Пігментні плями" → "Плями стали світлішими" → "пігментації майже не було"
- OURS: "дрібні заломи, лінії впоперек" → "Заломи вже не так помітні" → "лінії впоперек уже не так помітні"

Judgment on visibility:

1. All three name a problem the camera can show; ours passes the test in principle — "заломи" and "лінії
   впоперек" are drawable and are drawn at 1:32.
2. The original and Taras put the defect on the FACE, which is in every shot of a 9:16 talking-head film. Every
   heroine close-up before the mirror beat is a "before" and every one after is an "after" at zero editorial
   cost. Ours puts the defect on the NECK, which is at the bottom edge or cropped out of a face-framed
   vertical shot; the film has to cut to a neck crop to prove it. The scan line "від обличчя до шиї" is a good
   fix for the lyric, but the picture still has to work harder.
3. Ours adds a hedge inside the diagnosis — "Шкіра **на вигляд** в'яла" — which is not a showable property
   (it is a judgment word). Taras has no hedge in the diagnosis. Cost: 3 syllables and a softer line.
4. Ours spends 29 syllables on the diagnosis vs Taras's 18 (Taras cut "forehead and lip"; ours effectively
   names two sub-features, creases AND transverse lines AND laxity). One noun is enough for the echo.

## 6. Claim style

| Beat | ORIGINAL | TARAS | OURS |
| --- | --- | --- | --- |
| Mentor verdict | "She said stop, all of that is making it worse" | "Вона сказала: зупинись, усе це робить тільки гірше" | "«А шия? Їй щодня потрібен власний догляд»" (new category argument, no "stop") |
| Ritual | "Use this every night / Throw the rest aside" (12 syl) | "Наноси щовечора / Решту викинь" (11 syl) | "«Крем для шиї. На чисту шкіру — тонкий шар, щодня. Легенько, знизу вгору, без обіцянок чудес»" (28 syl) |
| First feel | "it felt right No peeling, no sting" | "не пекло, не лущило" | "легко ковзає, зволожує шкіру" |
| Result 1 time | "Three weeks later" | "Через три тижні" | "Минуло кілька тижнів" |
| Result 1 | "The patches on my cheeks were lighter" | "Плями стали світлішими" | "Заломи вже не так помітні" |
| Result 2 time | "By week six I was in shock" | "На шостому тижні" | "Минув ще час. Мій догляд тривав / І знову шию порівняла" |
| Result 2 | "The melasma was almost gone / My skin more even than it had in a decade / No foundation on" (32 syl) | "пігментації майже не було / Шкіра рівна, як не була десять років, без тонального" (33 syl) | "Вона помітно гладша, / На вигляд — пружніша. / І лінії впоперек / Уже не так помітні. / Я лишилась собою. / І шию більше не ховаю." (61 syl) |
| Disclaimer | none | none (bullets on end card only) | "Без обіцянок чудес" (7 syl) + "Результат у кожної свій" (8 syl) |
| CTA | "The cream is called Seora / The link's below / If you need it" (14 syl) | "Крем називається Ханбіт, посилання нижче" (14 syl) | "go pure — це крем для шиї. Результат у кожної свій. / Посилання внизу. У формі — ім'я й телефон лиши, / Якщо такий догляд потрібен тобі." (40 syl) |

Measured: the original's claim grammar has five parts — (a) a NUMBER of weeks, (b) the SAME NOUN as the
problem, (c) an ABSOLUTE change verb ("lighter", "almost gone"), (d) a COMPARISON ANCHOR ("than in a
decade"), (e) a BEHAVIORAL PROOF ("no foundation on"). Taras reproduces all five verbatim. Ours keeps (b)
and (e), drops (a) and (d), replaces (c) with the comparative "не так помітні" — which it then uses TWICE
(160.3 s and 182.5 s), and adds two hedges ("на вигляд" ×2) plus two disclaimers.

Judgment: the hedged claim is both weaker and longer (Result 2: 61 syl / 18.2 s vs 32 syl / 12 s). The
adaptation map records the reason ("Three-week guarantee replaced with fictional passage of several weeks",
"no medical cure", "no depigmentation claim") — a compliance decision, not a writing accident. If numbers
really are off-limits for goPure, the copy-chief fix is: hedge ONCE, with one word, and never repeat the
same hedged phrase; keep a comparison anchor that is not a medical claim ("вперше за десять років шию не
ховаю" — behavioral, not clinical).

## 7. What OURS adds that the original does not have, and what it costs

| Addition (sung) | Line(s) / time | Syl added vs original beat | Function | Clarity/selling effect (judgment) |
| --- | --- | ---: | --- | --- |
| Category argument: "«А шия? Їй щодня потрібен власний догляд»" | 34 · 130.6–136.7 | +5 (13 vs 8) | replaces "stop, making it worse" | The ONLY genuinely new selling idea in ours (the neck needs its own care). Good for a neck product. Loses the "stop" drama; costs 6.2 s vs 2.6 s. |
| Label reading / ritual: "На чисту шкіру — тонкий шар, щодня. Легенько, знизу вгору" | 37–38 · 141.6–150.8 | +9 (28 vs 18 incl. disclaimer) | replaces "every night / throw the rest" | Usable instruction, but it reads like a leaflet. Original keeps ritual to two imperatives. |
| Disclaimer 1: "Без обіцянок чудес" | 38 · ~148–150.8 | +7 | none in original | Undercuts the handoff before the first application. |
| Disclaimer 2: "Результат у кожної свій" | 63 · 226.7–232.1 | +8 | none in original | Undercuts the CTA at the exact moment of the ask. |
| Hedges "на вигляд" ×2, "мов" | 25, 47, 10 | +7 | none | Each one dilutes a picture word. |
| Adherence beat: "Мій догляд тривав" | 45 · 169.6–177.4 | +5 | none | Filler; Taras uses a number instead ("на шостому тижні"). |
| Doubled result: "гладша… пружніша… лінії… не так помітні… лишилась собою… не ховаю" | 46–48 · 177.4–187.8 | +19 | expands "almost gone / even / no foundation" | Says the result three ways; the original says it once and proves it once. |
| Form CTA: "У формі — ім'я й телефон лиши" | 64 · 232.1–236.2 | +12 | expands "link's below" | A second action in the CTA; in Taras the action is one phrase and the details live on the end card. |
| Brand at handoff: "go pure. Я стала читати" | 36 · 139.4–141.6 | 0 (replaces "Korean writing") | relocates naming | Neutral: brand heard twice, mystery-tube beat lost. |
| Rhetorical store question: "Та чи допоможе і цей?" | 14 · 50.8–56.0 | +3 | softens "that wouldn't work" | Compliance softening; slight loss of bite. |
| Durations kept verbatim: "П'ятнадцять хвилин сиділа", "Ще не бачу — знаю, вона. Вона мене впізнала" | 22, 16 | +3, +3 | Taras cut these | Texture, not function. |

Sum of identifiable additions ≈ 80 syllables ≈ 25 s at 3.2 syl/s. The remaining ~100 syllables of overrun
come from never cutting texture (see §8) and from Ukrainian morphology on lines that were mapped 1:1 to
English cards (line 13: 23 syl for 14; line 21: 21 for 15; line 51: 18 for 9).

Does the addition cost clarity? Judgment: the category argument (line 34) and the ritual (37) are clear and
sellable. The two disclaimers and the doubled hedges cost clarity: they are placed at the two most
persuasive moments (handoff, CTA) and the listener hears the brand and its caveat in one breath.

## 8. What TARAS cuts (and does not cut)

Cut (texture words, no function lost): "ex-" (sings "нову дружину чоловіка"), "Holding your kids together",
"Leggings", "that Saturday" → "у суботу", "one aisle over" → "через ряд", "fifteen (minutes)" → "просто
сиділа", "Patches on my forehead and lip", "Mid fifties" → "за п'ятдесят", "I had to say", "She laughed …
Said honey", "Tri-Luma", "in my hand", "something I'd never tried", "it felt right", "I was in shock",
"good different", "a few seats from me" → "поруч", "she walked up to me" → "підійшла", "by the way", "And I
meant it too", "If you need it".

Kept verbatim: every number that sells (16 years, 5 years in 6 months, 2 seconds, 3 weeks, week 6, 10
years), every zone word (лоб, щоки, підборіддя, щоки), every claim verb (світлішими, майже не було, рівна),
the whole self-worth couplet, the recital, the compliment, the name-then-link ending.

Errors/deviations in Taras (measured, for the record): "тридцять два" for 34; "чоловіка" for
ex-husband; "дівчини" for "someone from his office". Judgment: none of these touch a selling function.

## 9. Phrase boundaries: 43 vs 65

Measured: Taras sings 43 phrases averaging 14.6 syllables / 4.5 s, with 5 instrumental breaths of 1.1–2.0 s.
Ours sings 65 lines averaging 11.8 syllables / 3.7 s, keyed 1:1 to the 65 caption card IDs
(`line-anchors.json` → `sourceCaptionId`), with 0 gaps ≥ 1 s. The original's 65 cards are caption cards,
not phrases (its ASR has 76 segments).

Judgment: AGENTS.md rule 18 says "source transcript rows do not dictate new sung-line boundaries". Taras
complies (2 cards → 1 phrase on average); ours is locked to the card grid. This does not by itself cost
syllables, but it prevents the two things that keep a song at budget: merging two thoughts into one
phrase (which forces dropping connective words) and leaving instrumental air.

## 10. Reconciling the owner's permanent rule with the budget evidence

The rule (`references/owner-generation-rules.md`, 2026-09-11): "Песня — это Advertorial который переводит в
песню… Она должна быть милозвучна… именно песней… Звучать, вовлекать." AGENTS.md 18: "A song is an
advertorial composed into a real song… Source transcript rows do not dictate new sung-line boundaries."
`references/adapt-phase-contract.md`: "Переписывать перегруженные формулировки **прежде, чем удлинять песню
или ускорять вокал**."

The evidence: the owner's preferred adaptation (Taras) is 194 s, ~620 syllables, 310 words, 43 phrases,
every function present, ~30 % of the texture words gone, claim grammar copied verbatim, one sung CTA
phrase with product facts moved to the end card.

Reconciliation (judgment): the two are not in conflict. "Advertorial rewritten into a song" defines the
MEANING source: write the full selling text so nothing is lost by accident, map every function. The
original song defines the BUDGET and CLOCK: its syllable count and beat timing are the container the
advertorial must be composed INTO. The contract already ranks rewriting above lengthening. Taras is what
"composed into a real song" looks like when the container is respected: full coverage of FUNCTIONS,
not of WORDS. Our version treated coverage as 65/65 thoughts ("allThoughtsMapped: true") plus additions,
and let the container grow by 27 % to fit — the exact ordering the contract forbids.

## 11. Top 10 rules for turning an advertorial into a song WITHOUT exceeding the original's syllable budget

1. **The budget is the original SONG's syllable count, not the advertorial's word count.** Target = original
   ±5 % (here 620 ± 30), measured on the SUNG text by vowel count, not on the prose script (ours: prose 843,
   sung 766, original ~620). Density is fixed at ~3.2 syl/s for all three versions, so every 32 syllables
   = 10 s of runtime; there is no tempo lever.
2. **Plan the Ukrainian word count at two-thirds of the English.** Ukrainian runs 2.0 syl/word vs 1.25;
   Taras spends 310 words for 467. Write the lyric to a word ceiling (≈ 0.66 × source words) before
   touching melody.
3. **Cover every FUNCTION, not every THOUGHT.** All 17 functions survive in Taras with 43 phrases; the
   words that go are durations, adverbs, second zones, "by the way / I meant it too / if you need it".
   Cut texture before cutting any function; never cut a number that sells.
4. **Pay for every expansion inside the same act, and re-land on the original's clock.** Taras spends +17
   syl on humiliation and +12 on self-worth, pays with −13 (kids), −7 (forehead/lip), −5 (leggings), −5
   (good different), and lands the self-worth line within 1.6 s of the original. Per-beat drift tolerance
   ≈ ±10 s; ending tolerance ≈ +6 s. A section that is >15 % over must name its offsetting cut.
5. **One zone, one noun, and the same noun in problem, mirror and result.** "плями → плями світлішими →
   пігментації не було". Ours already has the echo (заломи → заломи → лінії); spend 18 syllables on the
   diagnosis, not 29, and put the defect where the camera already is (face beats neck on a phone; if
   neck, keep the "погляд ковзає до шиї" scan line so the lyric moves the eye).
6. **Keep the five-part claim grammar: number + noun + absolute verb + comparison anchor + behavioral
   proof.** If a number is not allowed, hedge ONCE with one word and never repeat the hedged phrase ("не
   так помітні" ×2, "на вигляд" ×2 in ours). A behavioral anchor ("вперше за десять років не ховаю шию")
   is not a clinical claim.
7. **Disclaimers do not get sung.** Zero sung disclaimers in the original and in Taras; Taras's caveats
   are end-card bullets. Ours sings two (15 syl) at the handoff and at the CTA — the two most persuasive
   seconds. Move them to on-screen text.
8. **CTA = one sung phrase, ≤ 15 syllables, name + "посилання нижче".** Original 14, Taras 14, ours 40.
   Form fields, "результат у кожної свій", and "якщо потрібен" belong on the end card, not in the melody.
9. **Compose in musical phrases (≈ 40–45 phrases of 13–15 syllables / 4–5 s), not caption cards, and leave
   3–5 instrumental breaths of 1–2 s.** Merging two cards into one phrase is where connective words fall
   away for free; breaths cost zero syllables and are what makes it "a song" (Taras 5 gaps, ours 0).
10. **Reveal the pack at the handoff (~57–65 % of runtime) and sing the NAME as the last thought after the
    self-worth line.** Original 96.9 %, Taras 97.8 %. Naming it earlier is allowed only as a 1:1
    replacement of a mystery-object word (ours did this at zero cost); reading the label aloud is not.

## 12. Limitations

- English syllable count is heuristic (584) and lower than the rate-based estimate (~625); both give the
  same conclusion (ours +22 % to +31 %, Taras within ±1 %).
- Taras ASR was corrected from burned-in captions where the two disagreed; captions are the authority and
  were read from 1-fps contact sheets, so phrase boundaries carry ±1 s.
- Function boundaries are the author's cut; a different grouping shifts a few seconds between neighbouring
  rows but not the totals.
- "Where first shown" for the original and Taras is read from 1-fps contact sheets (±1 s).