The child becomes the rescuer

What this reference actually does

The Resilia ad sells a family future before it sells an aged-garlic supplement. Its strongest structural choice is to give the quest to the seven-year-old daughter while leaving the purchasing decision with her father. The father initially offers reassurance he cannot substantiate. The child acts on the fear the adults try to contain. A neighbor supplies an explanation and a product; the father must then decide whether to act. The closing image returns the family to ordinary life together.

That is the transferable narrative DNA: a loved one notices a threat → reassurance fails → the loved one searches → a trusted helper makes the problem understandable → the buyer tests the explanation → a small ritual becomes possible → an earlier relationship wound is repaired → an offer enables the viewer's next step. The supplement's medical claims are not part of the reusable truth.

The owner described the reference as having “1M+” in ad spend. Currency, period, account data, profitability and attribution were not supplied. This report analyzes observable craft and proposes hypotheses; it cannot establish why money was spent or what caused conversion.

Evidence and coverage

• Original: resilia-aged-garlic-family-cholesterol-story.mp4; SHA-256 and complete FFprobe output are in source.json.
• Measured duration: 325.682 s, picture 325.567 s, 720 × 1280, 30 fps, 9:16, H.264 picture and stereo AAC audio. The 115 ms audio tail is a container/track difference, not a separate story beat.
• Gemini through the authorized neighboring Kie account: two full-video native analyses plus six overlapping-edge, approximately 55-second segment analyses. The native route received actual inline video/audio; the measured proxy is 360 × 640 at 6 fps and retains the full spoken track.
• Source annotations: 76 speech entries and 114 shot/state entries. Entries include reaction shots, inserts, fragments and segment-boundary repetitions. Neither total is a verified count of unique sentences or hard cuts.
• Edit measurements: FFmpeg detected 89 candidate cuts at threshold 0.22. Candidate interval median 2.517 s, mean 3.617 s. Soft transitions can be missed and flashes can be counted; these are algorithmic candidates.
• Contact sheet samples every 10 seconds. Full moving media and segment evidence underpin dialogue/motion observations; the sheet alone cannot establish acting or audio.
• The English transcript is model-assisted, with approximate times. The aggregate contains 758 space-delimited words including repeated boundary fragments. It is not a certified transcript or exact words-per-minute measurement.

Corrections to the machine analysis

The first OpenAI-compatible Gemini route did not ingest the media reliably: one response said no attachment and the other returned code 524. Both are retained as failed attempts. The native Kie/Gemini route then reported approximately 30,000 input tokens for the complete video and described verifiable beginning, middle and ending events.

The broad native pass still summarized dialogue while calling it verbatim, invented a “40%” retention-drop estimate, called the footage live action without production evidence, and assigned arbitrary DNA scores. Those claims are rejected here. The segment pass gives substantially fuller dialogue. Segment 3's shot timestamps were returned as decimal clock notation (0.13 for 13 seconds) although its transcript uses seconds; the compiled shot ledger normalizes those values and records the correction. One final phone-call answer was labeled “Doctor” by the model even though the scene follows a call thanking the helper; speaker identity is flagged for review rather than used to prove a medical endorsement beyond the actual clinic sequence.

A hook that develops for thirty seconds

The first spoken number is not the whole hook. The doctor delivers the event, the father supplies vulnerability, and the daughter's question gives the event its emotional meaning. The scene also sets up a return to the same clinical situation near the end. The viewer is invited to compare two states of the same family.

The ad delays brand disclosure. A pouch becomes identifiable around 118–120 s; the brand is spoken around 130 s, roughly 40% into the runtime. These are approximate observations, not frame-certified first-visibility measurements. Early attention must therefore be earned by the story itself.

Causal story, beat by beat

Screen time is 5:26; story time includes one night, the following morning and an eight-week progression. The night-to-morning reunion creates a small plot climax well before the product-result climax. That gives the middle more than one reason to keep watching.

Character passports and relationships

Father — buyer avatar, recipient of rescue, eventual testimonial voice

Presented as a middle-aged man with short dark hair, some grey, light stubble and dark casual layers. He is initially seated, listening and looking between doctor and family. His dialogue starts with short questions. He later inspects the pouch and asks substantive objections. His eventual ease is shown through participation with his daughter and family, not merely an isolated smile.

His external want is to improve the situation described in the consultation. His deeper need, as the story frames it, is to remain available to his family. His defensive behavior is reassurance before certainty. The decision beat matters because the child cannot purchase or adopt the routine for him. He must move from recipient of concern to acting adult.

Lily — emotional focal point and plot-driving rescuer

Named in dialogue and described later as seven. Long brown hair, pink knitwear/jacket and a tablet establish continuity. The important progression is not “sad child → happy child.” It is listening → checking the adult's promise → investigating → committing → asking for help → testing an explanation → recruiting the parent → receiving reassurance embodied in family life.

Her facial cues include raised inner brows, sustained upward looks toward adults, a small tightened mouth during worry, and more open expressions during later activity. The cup held with both hands and blanket around her body keep her vulnerability visible even while she drives the plot. Her unusually adult rescue mission is also a plausibility risk; it may feel moving to some viewers and manipulative to others.

Mother — protective realism and the objection voice

Brown shoulder-length hair, soft cardigan/neutral clothing; she drives the car, monitors the father, touches his hand and reacts strongly to Lily's return. She holds the family together while the plot temporarily assigns rescue to the child. Her “what if it doesn't work?” prevents the product decision from feeling unanimous too quickly. She is not an antagonist.

Male neighbor — helper and explanatory authority

Short dark/grey hair, dark jacket over a light shirt, calm posture. His stated connection to Japan supplies the cultural authority cue; appearance alone does not establish nationality or expertise. He invites Lily into warmth, explains, retrieves the pouch, escorts her home and speaks directly to the father. The handoff from child to buyer is dramatized rather than assumed.

He combines hospitality, personal experience, a memorable mechanism, product discrimination and confident answers. None independently proves medical expertise. His claim to be the father's age and to have a particular reading is dialogue evidence, not independently established identity or health data.

Female neighbor — the threshold of safety

Light knitwear and an open doorway contrast with the cold exterior. She recognizes the child, asks whether her mother knows and admits her. This brief role makes the mentor encounter feel like a social relationship rather than a product spokesperson appearing from nowhere.

Doctor — bookend authority

White coat, chart, clinical environment, precise measurements and initially contained seriousness. Later surprise changes his behavior. The same authority device that creates the problem is made to validate the solution. This is a powerful rhetorical closure; it remains a staged claim within the ad.

Pain, emotion and moral permission

The visible condition is a number and a narrated scan result. The emotional pain is prospective absence: the child might lose her father, and he might fail to stay for her. The ad makes the cost of inaction interpersonal. It reframes self-care as an obligation to people one loves.

The emotional movement is approximately alarm → uneasy reassurance → private fear → purposeful curiosity → suspense → warmth → understanding → cautious hope → family alarm at the reunion → skeptical inquiry → commitment → relief → gratitude → protective resolve. This is an analyst interpretation supported by events and performances, not a measured audience response curve.

Three controlled changes in emotional intensity are particularly useful to study:

1. The warm kitchen slows the fear after the night walk. It makes space for exposition while retaining the child's purpose.
2. The morning reunion briefly intensifies emotion again. It reminds the father, and the viewer, why a technical product discussion matters.
3. The father's objections restore adult agency after the child's plea. That gives the commercial explanation a more credible conversational form.

The ending transfers motivation from “I want better numbers” to “I want to remain here for them.” It also risks guilt-based pressure. In a cosmetic adaptation, a woman's worth and place in her family must not be conditional on a product changing her appearance.

Open-loop and objection ledger

The product explanation repeats, but its audience changes: child first, purchasing adult second, viewer third. This is repetition with a new dramatic job, not merely repeated features. Each listener raises a different question. That is a useful generator constraint for future narratives.

Persuasion architecture and its limits

The presumed audience knows the health concern and may be familiar with treatment and supplement alternatives. Solution fatigue and skepticism are reasonable hypotheses from the dialogue, not demographic facts about delivered impressions.

The ad's mechanism stack is: arteries as pipes → buildup as a cause → ordinary approaches described as incomplete → garlic transformed by time → S-allyl cysteine as active compound → visible softgel color as authenticity cue → personal use and marketplace popularity → claimed outcome. It makes a complicated account feel inspectable through a kitchen conversation and a pouch in the buyer's hand.

Several parts of that stack are hazardous to treat as fact. Claims that a supplement scrapes off calcium, melts cholesterol, replaces the role of prescribed management, or changes how a doctor treats patients are asserted in the ad, not validated by this analysis. The same applies to the result attributed to the supplement, “real” reviews, marketplace rank and competitor accusations. The montage has no control group and does not disclose concurrent care. The final number is not proof of cure.

The selling sequence also answers practical questions: what to buy, how to recognize it, how often to take it, how long to try it, what to expect, why to trust it and how to reduce purchase risk. These questions can be retained in a new category while all category-specific answers are replaced and substantiated.

The final scarcity argument uses the earlier aging mechanism again: long production time supposedly constrains supply. That makes urgency narratively connected to the explanation. It still needs inventory evidence. In the Ukrainian goPure adaptation, the birthday supplies a story deadline; no invented stock deadline or copied discount is used.

Visual, performance and edit grammar

The footage has a photoreal dramatic style. Whether it was shot with actors, generated, composited or mixed cannot be established from the supplied file alone. No production provenance was supplied.

Composition: faces dominate vertical close-ups and medium shots; the family relationship is established through seated groups, hand contact and shared eye lines. Product inserts arrive after the viewer has a reason to inspect them. The father reading a pouch makes the label part of a decision, not only a beauty shot.

Camera: restrained push-ins, stable conversational framing, shot/reverse-shot and reaction coverage support emotional legibility. Exterior/car transitions establish movement between places. The tablet inserts externalize thought; the child does not need a long explanatory monologue to communicate her search. Later montage shots vary activity and distance to suggest time and recovered participation.

Color and light: neutral/cool clinical scenes, a dark blue exterior and screen-lit bedroom give way to warm doorway/kitchen practical light. Later daylight and greenery expand the frame's emotional world. The movement from constriction to warmth is story-motivated. A simple “sad = blue, happy = orange” LUT recipe would miss the role of actual locations and light sources.

Props: chart = consequential information; tablet = active search; pink clothing = child continuity; door = access to help; blanket and mug = being cared for; pouch = portable explanation; softgels = discriminating detail; water glass = repeatable ritual; family meal and dog = ordinary life worth preserving.

Acting: worry lives in listening and held looks, not only spoken lines. The father checks other people's reactions; Lily looks up to adults while holding her mug; the helper softens his manner toward her and becomes more explanatory with the father. The mother protects and questions. A generation prompt saying only “expressive family” would erase those distinct tasks.

Edit: brisk early alternation among doctor, father and child gives one event several emotional readings. The search/walk uses non-dialogue action. The long kitchen middle obtains variation through questions, reactions, inserts and audience changes. The montage compresses weeks. The clinical return provides visual rhyme; lifestyle scenes then carry the offer's voiceover. The final graphic is allowed to feel like an ad because the story has already closed.

Captions: short all-capital white phrases with a dark edge/shadow sit around the lower-middle picture. They track speech in small chunks. They aid silent comprehension but can compete with hands or pouch labels. The end card changes register to a conspicuous red/white promotional headline, guarantee seal and button. Exact font identification is unverified.

Audio-first DNA

The source is primarily multi-character spoken drama, later moving into a first-person commercial narration. It is not a song. Its speakers provide different rhythms: clipped doctor statements, brief parental reassurance, a child's emotionally direct questions, a helper's longer teaching phrases, skeptical buyer interruptions and a more fluent closing testimonial.

The absence of speech during parts of the search and movement matters: picture can carry intention while the viewer waits. The child's small hesitations and the parents' pauses carry vulnerability. Product terms and numbers create deliberate diction demands. Background underscore and ambience are audible in the source, but their exact instrumentation, stems and mix levels have not been independently isolated here; the broad model's “piano” label is not treated as a measured instrumentation fact.

For adaptation, the audio is generated and reviewed before shot timing. One stable Ukrainian daughter-narrator preserves the child's/relative's point of view in an adult voice and makes quoted dialogue intelligible across generated silent scenes. This deliberately changes the source's multi-voice dramatic mode. It avoids pretending silent generated mouths are synchronized dialogue; the picture uses listening, reactions, hands, movement and over-the-shoulder angles. The tradeoff is less immediate conversational realism. No music is generated or added.

The goPure adaptation: what stays, what changes

The story's concern is loss of comfort and self-consciousness, not diminished worth. Natural lines remain on the mother's neck throughout. No de-aging, body reshaping, clinical scan, before/after split screen, “minus ten years,” competitor injury claim or guaranteed four-week reversal is included. The source's “foreign secret” is replaceable surface rhetoric, not essential DNA.

The source business brief contains claims that heavy creams physically pull skin down and that the neck has no sebaceous glands or supporting bone. These are not adopted. The adaptation uses the live page for product identity and seller terms and a restrained hydration explanation consistent with the American Academy of Dermatology's guidance. The AAD distinguishes modest moisturizer effects from lifting sagging skin.

The Ukrainian seller page was fetched on 2026-09-09. Its snapshot lists the 50 ml jar, 4% cupuaçu, olive lipid complex, hyaluronic acid and caffeine, upward application, phone orders, payment on receipt and 60-day return terms. These are seller assertions, not an independent ingredient assay or fulfillment audit. The global goPure site has different product variants/label details; its studies, sizes and percentages are not merged into this SKU. Seller authorization and product authenticity were not independently established.

A better analysis system, inspired by this ad

The reusable passport format expands video analysis into eighteen linked sections. Five changes matter most for future use:

1. Store causal turns and belief changes beside shots. A scene becomes useful for adaptation when its job is known. “Kitchen conversation” is not enough; “buyer challenges category equivalence and receives a discriminating explanation” is.
2. Track who knows what. The daughter, father and viewer receive information at different times. This controls suspense and the purpose of repetition.
3. Separate identity, concern and performance. A stable face does not guarantee stable neck texture or emotionally correct acting. Each needs its own reference and review.
4. Make open loops addressable. Each opening question needs a specific answer and a timestamp. A generator can then preserve the loop while changing product and culture.
5. Make claims and provenance first-class fields. Observed speech, seller assertions, hypotheses and measurements must never collapse into a single “truth” field. Failed or overconfident model analysis should be auditable.

The machine-readable DNA is the populated handoff for this source and the adaptation. It connects timing, narrative function, evidence status, localized replacement and production constraints.

What to test after production

These are hypotheses, without invented lift estimates. Randomize comparable audiences/placements and keep product offer constant. Define sample size and decision rules before spending.

Spend is not success; attention is not conversion; a generated film is not an ad test. Actual effectiveness, Ukrainian audience response, pronunciation judged by a native listener, fulfillment and product outcomes remain separate validation tasks.

Delivered adaptation and production evidence

The complete audio-led goPure story is «Місце поруч», with Ukrainian script, caption timing, 37-shot production plan and production report. The interactive report connects the reference video, causal turns, source evidence, full analysis and adaptation.

The accepted narration lasts 349.440 seconds and was locked before any picture generation. It uses Microsoft Edge Ukrainian speech after two Kie ElevenLabs provider failures. The model-assisted listening audit and exact normalized word alignment are retained. Motion uses Kie Seedance 2 Mini at 480p; no music or native video audio is included. See the production report for measured export results, provider credits, corrected failures and remaining visual limitations.

Video DNA passport — evidence before interpretation

Version 2 production contract. Read the callable meta-function, the populated contract, the vocal performance passport and the LevchykUGS study. New generation goes through scripts/compile_narrative.py; a descriptive document alone cannot pass the factory gates.

Required passports: video craft; transferable persuasion function; each character; product identity/use; product-driven change; wardrobe/props; attention; vocal performance; Ukrainian pronunciation; regional visual references. Link them by stable IDs and exact asset hashes. Describe what the provider must make and what the reviewer must actually see/hear.

A reusable way to reverse-engineer an ad and hand its structure to production. The unit of analysis is a change: someone learns, wants, feels, believes, acts, or sees something different. A shot list alone misses the causal story; a marketing summary alone misses the craft.

1. Source and provenance

Record asset ID, original path/URL, SHA-256, duration, dimensions, frame rate, audio tracks, languages, caption method, date analyzed, model and request receipts. Record campaign claims separately: who supplied spend, currency, date range, placements, attribution window, revenue, and whether raw account data was inspected. A large spend is a reason to study the creative, not proof that every feature caused success.

2. Evidence vocabulary

Every substantive annotation carries start_s, end_s, evidence, status, confidence, and reviewer. Use these statuses:

Confidence measures certainty about an annotation, never predicted sales. High = directly legible/audible and corroborated; medium = clear pattern with approximate timing; low = ambiguity, inference, or unverified source assertion. A transcription system must not label summaries “verbatim.” Preserve rejected provider output with reasons. Test a whole-video model’s clock against distinctive start/middle/end anchors. If it assigns an event to the wrong time, reject those timecodes and adjudicate short excerpts with explicit media hashes and known offsets; the model must not become the authoritative edit clock.

3. Source fact sheet

Capture category, format, aspect, visual realism, dialogue/narration mode, perspective, main viewer promise, apparent audience, destination/placement if known, first product appearance, first spoken brand, first benefit, first CTA, final CTA, total words, speaking time, silence, and normalized reveal position. Separate screen time from story time. Count speaking roles independently of visible roles and extras.

4. Hook microscope

Inspect the first frame and 1, 3, 5, 10, and 30 seconds. At each checkpoint record: visible subject, spoken proposition, information withheld, immediate stakes, emotional invitation, motion or edit interrupt, legible caption, and reason the next second is needed. Identify overlapping hooks: visual incongruity, danger, relationship, knowledge gap, contradiction, identity, and specificity. Do not reduce the hook to its first sentence.

5. Story and belief timeline

Use one row per meaningful turn, not a forced universal scene count:

time → event → cause → character goal → obstacle → decision → viewer question → emotional change → belief changed → evidence → next beat

Identify inciting incident, commitment to the quest, failed approaches, mentor arrival, explanation, product discovery, objections, adoption, elapsed-time proof, relational payoff, and commercial close when present. Mark absent stages rather than inventing them. A beat may perform several functions. Explain why the next beat follows from this one.

6. Character passport

For each recurring person: stable identity markers; screen and story role; relationship graph; what they know at each stage; external want; internal need; fear; misconception; agency; turning point; dialogue style; recurring props; wardrobe; body posture; gaze; brow; eyelids; mouth tension; gesture; spatial position; entrance and exit. Record the narrator separately from the protagonist. A child may be the plot's agent while the parent is the buyer avatar.

7. Product-concern continuity

Identity is independent of the concern. Bind affected person, exact body/object region, separate observable markers, baseline, actual use, story days, intermediate states and resolved endpoint. Keep age, face shape, body size, hair, lens, pose and lighting stable. The cream example requires the problem clearly present in week one, visibly lower in week two, lower again in week three, and visually absent at week four. “Less severe” cannot pass as “resolved.” First use remains at baseline; it is not instant erasure.

Use four distinct matched references, hold each state long enough to inspect, and use a short matched dissolve at labeled elapsed-time boundaries. The generator must not morph a face mid-shot, change clothes or relight the neck to imply improvement. Review actual pixels at state, board, clip and final-film levels. A smile, caption, removed scarf or another actor's neck cannot stand in for the result. These are illustrative visual targets, not clinical measurements or guaranteed results. Mentor and family members are explicitly excluded from the target concern. Other categories use their own observable and timescale: a cleared error queue or a controlled alert test, not cosmetic weeks.

Wardrobe and prop changes have passports too. Each garment has an owner, ID, neckline, color, sleeve, material, fit and reference. Every shot assigns a wardrobe state to every visible person. Preserve states across reverse angles and adjacent scenes. A change needs an explicit event and visible action, such as mother putting on the high collar or changing into the birthday blouse. The daughter cannot inherit the mother's concealment garment. Bind narrated actor → action → object → expected visible result to the audio words; inspect the actual frame interval and both sides of every cut. Reject a mother already wearing the supposedly newly offered blouse and wrong-sleeve product inserts.

8. Emotional architecture

Map surface discomfort → functional loss → social consequence → identity wound → threatened relationship → desired future. For each beat record the emotion invited, valence, arousal, perceived control, intimacy, and uncertainty. Optional 0–4 curves are analyst ratings with written anchors, never audience measurement. Include recovery intervals, dignity/absolution, and whether persuasion relies on fear, shame, belonging, hope, anger, curiosity, relief, or agency.

9. Retention ledger

Give every open loop an ID, opening time, exact question, reminders/escalations, answer time, delay, and whether the answer creates another loop. Separate plot suspense from product curiosity. Track microhooks, specificity, character reversals, location changes, interruption, reveals, callbacks, pattern breaks, progressive repetition, surprise, payoff and residual uncertainty. Measure viewer drop-off only when actual platform data exists.

10. Persuasion ledger

Describe awareness and market sophistication as hypotheses. Store problem definition; mechanism of problem; mechanism of solution; comparison/failed alternative; authority source; discovery story; demonstration; proof claim; objection; answer; commitment; offer; risk reversal; urgency; CTA verb; destination; and transaction friction. For each mechanism ask whether it is literal, metaphorical, asserted, or independently supported. Distinguish a credential cue from a verified credential, a character result from clinical evidence, and a seller review from an audited sample.

11. Visual grammar

Annotate shot scale, camera angle, lens impression, movement, horizon, subject placement, eyeline, negative space, foreground/midground/background, occlusion, scene geography, focus/depth, color temperature, contrast, light direction, motivated practical light, wardrobe, production design, weather/time, props and motifs. Note visual continuity and probable artifacts without declaring the production method from appearance alone. Use contact sheets for identity and color arcs; moving excerpts for acting and continuity.

12. Editing grammar

Create a candidate cut list from scene detection, then adjudicate boundaries. Record hard cut, dissolve, fade, insert, reaction shot, shot/reverse shot, montage, match action, flashback, temporal ellipsis, repeat and hold. Measure shot durations and cadence by act, first 10-second cut count, close-up proportion, visual novelty interval and time spent on reaction vs explanation. A scene detector's threshold output is not a ground-truth shot count.

13. Audio and captions

Audio is the clock. Preserve the approved spoken script, voice IDs and language. Every line requires context, speaker, intention, emotion, pitch contour, energy/timbre, pace, pause, emphatic words, quoted-role switch, source delivery reference, negative direction and listening criterion. Analyze actual reference speech, not just the words or face. Questioning, hesitant reassurance, patient explanation, doubt, joyful release and the CTA cannot share one generic warm reading. The actual provider request must carry the directions its model supports. Listen for each intended turn; reject flat or overacted delivery independently of pronunciation. Measure loudness, peak, silence and duration as separate technical checks. For this format do not create or decompose music.

Ukrainian stress is an occurrence-level contract: contextual evidence for every inflected word, U+0301 after the stressed vowel of every multisyllabic word in actual submitted text, explicit foreign-brand pronunciation, and exact script/request/score hashes. Listen to every occurrence in the real recording and record intended vs heard stress, consonants, time and pass/fail/uncertain. ASR spelling and written accents do not prove stress. Uncertainty blocks acceptance; repair in a versioned attempt, realign and recheck the assembled voice. Strip only U+0301 for plain captions, preserving Ukrainian letters.

Captions carry the heard words, real alignment source, cue start/end, line breaks, position, face/product occlusion, contrast and phone legibility. Distinguish measured provider word alignment from manually verified timing and from estimates. Record where picture anticipates, illustrates, reacts to, or contradicts a spoken phrase. Reject orphaned words and cuts through meaning units.

14. Localization and product mapping

For each source beat, specify function kept → local event → product truth → visual → audio → reason for change. Translate idiom, social roles, settings, familiar objects, authority, currency, transaction model, calendar and CTA. Localization is not flag decoration. Keep dignity and recognizable daily life. Do not transplant a country's supposed biological superiority, disease claims, numerical outcomes, discounts, guarantees, or competitor allegations into a new product.

Preserve image attraction: saturated complementary accents, subject/background contrast, layered space, tactile materials, appealing silhouettes and memorable environments. Attach actual local visual references with source/rights, hashes, palette, local cue and a precise usage role. Put these images into generation refs; prose alone is not a reference bank. The original Ukrainian moodboard demonstrates teal/coral street contrast, emerald/amber tea, cobalt/orange care and a plum/green courtyard. It conditions setting/color, not cast identity or product application. Do not reproduce its grid.

For attention, annotate every retained second with a stimulus, meaningful change, small payoff and next question. Keep a ledger of open loops and handoffs. Flag unexplained holds and visual repetition; let a deliberate suspense hold pass with an explicit acting reason. Do not confuse frequent edits with meaning or promise zero drop-off without audience data.

15. Generation handoff

Lock source analysis → claim ledger → adapted script → rendered and reviewed audio → word/phrase timeline → character and concern references → shot plan → video generation → edit → captions → validation. Every generated item stores prompt, model, tier, references/hashes, task ID, receipt, output path, retained range, review and dependency. A changed line invalidates its alignment and dependent shots, not the entire paid run. Poll existing task IDs before resubmitting.

16. Quality gates and score anchors

Use pass / revise / unknown with concrete evidence: story causality, preserved hook function, sustained questions, clear mechanism, claim traceability, role distinction, Ukrainian diction, identity continuity, concern continuity, picture/speech relationship, mobile legibility, complete CTA, no music, correct video model/tier, and actual playable output. Subjective 1–5 craft scores require anchors: 1 absent/broken; 2 inconsistent; 3 intelligible; 4 clear and deliberate; 5 unusually coherent with multiple observed receipts. Do not create a single “DNA score” pretending to predict performance.

17. Experiment plan and limitations

Specify one-variable ablations, hypotheses, primary metrics, guardrails, audience and placement matching, random assignment, spend/sample allocation and decision rules before a test. Useful comparisons include caregiver quest vs direct explanation; early vs delayed reveal; dialogue vs single narrator; technical vs simple mechanism; relational vs appearance-only payoff. Track attention, qualified traffic, conversion and adverse feedback separately. Never invent an expected percentage gain from one reference.

18. Report packaging

Provide an executive reading, watchable reference, synchronized timeline, expandable evidence, character/concern passports, emotional and belief maps, source-to-adaptation table, claim ledger, full script, audio master, generated film, editable project, provider receipts, test plan and explicit unknowns. Machine-readable JSON should describe the same evidence as the readable report. Link timestamps to the actual local asset; keep temporary provider URLs distinct from durable deliverables.

Evidence that can be checked

114 model-assisted shot/state annotations, 76 speech entries, and 89 independently detected candidate cuts. Counts describe different units. Timing, speakers and claims carry explicit limitations in the report.

Character identity reference

Search the spoken evidence

Provider and verification receipts

Production decisions, costs and final quality review · Measured export verification

• Final film QA — accepted and rejected model findings
• Full-video forensic pass
• Full-video persuasion pass — raw, with rejected claims discussed in report
• Actual audio listening audit
• Audio lock and measured loudness
• Reproducible cut-detection measurements
• Ukrainian product page snapshot text

Two Kie ElevenLabs speech attempts failed with provider 500 errors. The accepted speech uses uk-UA-PolinaNeural through Microsoft Edge speech; the owner was informed before the change. Gemini analysis and Seedance video remain on Kie. No music was generated.

Visible friction → earned proof → restored possibility

RESTORE_POSSIBILITY(AUDIENCE, PROBLEM, STAKES, ACTOR, WITNESS, TRIGGER, FAILED_MODEL, HELPER, REFRAME, PRODUCT, RITUAL, PROOF, PAYOFF, OFFER, REGION, LANGUAGE, CLOCK) → a narrative in which the viewer understands a problem, tests an alternative, sees the declared outcome and knows the next action.

This function organizes a sales argument. It is callable across categories and situations; it cannot guarantee demand, attention, truth, or sales for every product. Missing proof must become an honest demonstration or a different argument, not fabricated evidence.

The viewer-state transformation is:

That affects someone like me → the easy explanation fails → I need an answer → this alternative makes sense → I can inspect what changes → I can imagine using it → I know how to act.

The source's daughter/father/doctor/garlic plot is one concrete implementation. Invariants are causal relationships; casting and calendars are slots. The machine-readable contract contains the full constraints, source windows and complete calls for cream, bookkeeping and a water-leak sensor.

Every stage needs a local mini-loop: a concrete stimulus, withheld answer, meaningful change, small payoff and next question. Let information arrive in increments. Alternate faces, action, objects and environments according to meaning. A long hold can work when a decision is visibly forming; another angle of the same explanation may still be dead air.

Swap test. Bookkeeping: an unreconciled ledger threatens an owner's family evening; an accountant exposes duplicate entries; a sandbox import and reconciliation clear them; the owner closes on time; a demo is the next step. Water sensor: a missed puddle threatens a finished room; an installer demonstrates placement and a controlled wet test; the alert arrives before water spreads; the homeowner leaves reassured; select a kit. Neither requires a daughter, a doctor, a cosmetic transformation or four weeks.

Proof grammar. Bind baseline observable → actual use → appropriate elapsed interval → intermediate checks → final observable → return to the initial test. For cream the illustrative target concern can fade week by week until absent at week four. For software the observable may be an error count; for a sensor, detection latency in a controlled test. Art-direction marker ratings are not efficacy percentages. Generated skin is not customer evidence. Packaging, mechanism, outcome and purchase promises each have separate sources.

Do not lose: opening image attraction, legible stakes, earned explanation, active buyer choice, inspectable transformation, callback and actionable close. Do not copy: disease assertions, a child alone at night, an ethnicity-as-authority cue, source cadence as a universal law, or an unverified guarantee/discount.

Call the same function with another product

The product-change passport

Week one must visibly contain the declared problem. Intermediate weeks must be distinct. Week four resolves the concern while keeping the person. First use stays at baseline.

Illustrative references, not clinical results. These use the upstream Olena example, not Natalia’s identity in the historical film. The generated week-two candidate needs refinement; its crop/texture do not yet pass the matched-view gate. The old endpoint is useful as an intermediate state, but failed as a final state.

Transition: inspect each complete state, then a 0.65 s matched dissolve at the labeled elapsed-week boundary. Preserve face, pose, lighting and garments. This is the production specification; these stills are not a newly rendered film.

People, clothing and actions stay connected

A mother’s high-collar sweater stays on the mother. New clothing requires a declared event. Each narrated action names its subject, garment/object and required visible result; the actual clip must satisfy it.

Local life with visual appeal

Teal and coral streets, emerald and amber tea, cobalt and orange care, plum and green celebration. Used as setting and palette conditioning in compiled prompts. The board does not define cast identity or prove cream application.

Keep the next second worth watching

Every retained interval needs a stimulus, a meaningful change, a small payoff and a next question. The compiler flags missing coverage, unmotivated holds and static acting. These are authored craft checks; audience attention remains unmeasured.

What transfers from LevchykUGS

Studied the active tree and recent history through d3b23e6fe0199533224bfec45c4fec7243dffdc6 (2026-09-09). Paths and content hashes are frozen in sources.json. The older September 2 worktree alone missed the most relevant September 7–9 fixes.

The useful failure is visible below. These are original upstream illustrative reference images, copied byte-for-byte. They establish a reusable target-state example, not the identity of Natalia in the historical film and not evidence of cream efficacy.

Baseline: concern clearly present	Rejected endpoint: concern persists	Resolved endpoint: concern absent

We adopt the visual endpoint rule and selective invalidation, but do not import song generation, melody/tempo quotas, or upstream provider requirements for songs. We also do not transplant unverified clinical claims, the source's character de-aging, or its claims that a particular mechanism caused audience dopamine or sales.

A voice that performs the story

The owner rejects the historical master as emotionless. It is not accepted under this new gate. A prior model saying “emotional and deliberate” is insufficient: it did not bind phrase-level expectations to audible evidence.

We listened through five bounded native-media audits: four reference excerpts plus the historical opening. Full receipts and local clocks are in vocal-reference. The model's adjectives remain interpretations, not acoustic measurements. The reference payoff audit misidentifies the remote helper as a doctor; do not copy that role label.

The complete vocal score assigns all 21 original paragraphs a speaker, context, intention, emotion, contour, pace, energy, pause, emphasis, negative direction, source reference and listening criterion. Quoted speech changes the narrator's stance; it must not become a cartoon imitation of age or gender. Emotion means fit to the moment, not shouting or crying throughout.

The delivery ladder is curiosity → realization → unease → vulnerable listening → determination → skepticism → trust → cautious hope → informed choice → relief → joy → gratitude → helpful invitation. Reserve the largest release for the photo callback. Quiet factual language still has an intention and rhythmic contrast.

New speech uses provider-compatible direction in the actual request, not a separate document the provider never sees. The expressive runner supports the documented Kie Eleven v3 dialogue contract. Tags are suggestions that must be auditioned; punctuation and tags can shape delivery. ElevenLabs prompting guide.

For Ukrainian, every multisyllabic occurrence in the submitted script receives contextual stress adjudication and U+0301. The ledger binds the exact script, direction score and request bytes. Preserve plain spelling for captions by removing only U+0301. Do not use generic Unicode accent stripping, which can corrupt Ukrainian letters.

Acceptance requires listening, not metadata. Each occurrence must have a local time, heard stress, intact consonants and pass/fail/uncertain. Each vocal line also needs heard delivery, context fit, no flat reading and no overacting. A correct ASR transcript, loudness, pitch range, tag or written accent cannot approve it. Check the opening, repetitions, quotes, product name and last CTA word. Repair a failed phrase in a new attempt, re-align, then listen to the complete assembled voice because transitions can expose seams.

The short directed audition is an experiment, not a replacement for the full 5:49 voice. Failed provider requests and listening verdicts remain visible in the review. New picture cannot use it as a complete narration master.

Listen to the reference delivery

Directed Ukrainian audition

Three-line expressive audition; independent machine listening accepted all 16 words and three distinct deliveries. This is not the full-film replacement. Actual listening verdict · Pronunciation ledger

All 21 paragraphs have a delivery

Check what the film actually shows.

Historical film: not accepted under v2. All 349.44 seconds were submitted across seven overlapping media windows. Direct frame inspection confirms the collar ownership error. Automated findings still require adjudication; submission coverage is not a claim of flawless human review.

Full-film revalidation evidence

• 0–54.00 s clip · Raw review with local times
• 50–104.00 s clip · Raw review with local times
• 100–154.00 s clip · Raw review with local times
• 150–204.00 s clip · Raw review with local times
• 200–254.00 s clip · Raw review with local times
• 250–304.00 s clip · Raw review with local times
• 300–349.44 s clip · Raw review with local times

Reject the raw model suggestion to keep the daughter in the erroneous turtleneck; preserve the intended owner instead. Retrospective narration does not require lip sync. The final window’s text response ends early, so its submission is not complete line-by-line proof.