STORYBOARD — V3 In 2003, Thirteen Million Sick Women Became Healthy — Documentary VO × No Presenter (D10 cinematic film) [KD_C0086]

The locked script is an Echo-chain deliverable and does not carry v2 Fields 19/20/21/25. Mapping used:

The density table (review check-10 note, boarded rather than footnoted)

The review recorded that the script's declared per-beat seconds sum to 233s while the uniform-wpm total is 234.5s, and that the divergence is deliberate: the donor's own per-beat speech density is wildly non-uniform (donor beat 5 = 120 wpm, beat 10 = 269 wpm) and the script tracks that curve. It recommended the board carry both numbers. It does:

What the editor does with this. The boarded M:SS spine is the CUT plan. Where a beat's word-derived seconds exceed its boarded seconds (3, 6, 10, 12, 13), the narrator is genuinely reading faster than 221.8 there and the picture must keep up: those five beats get the tightest cutting on the board. Where the word-derived seconds fall short (5, 8, 9, 11), the surplus is picture time, not pause time — the narrator does not slow down and does not breathe; the film holds the shot longer or takes an extra cut. Beat 5 is the only large gap (16s boarded against 9.7s of copy) and it is the board's one deliberate air pocket, spent entirely on the clinical-white mechanism insert. Nowhere on this board does a gap become silence: the donor's grammar is "no room for the viewer to think," and the bed plus the picture carry every surplus second.

Review-flag dispositions (ship order, gates delegated 2026-07-28)

Copy is LOCKED — this storyboard changes zero words. Verdict was SHIP AS-IS with 0 HARD, and option 1 (ship v1 as-is) is what the board executes. The five SOFTs are dispositioned as follows:

① F1 (zero narrator contractions; the house CTA is de-contracted to "Go see if it is in stock."). Not a storyboard defect and no fix was taken. Boarded as a VO direction constraint: the read is uncontracted by design and every take must stay uncontracted — a narrator who drifts back into "it's" mid-session will produce an audibly mixed register across a 3:53 documentary. The line is also never burned on screen, so the de-contracted form is heard once and never read. ② F2 (Hooks A and C are near-duplicate doors; C's cited H07 lineage drifts to H14 on the catalog's own tie-break). Rebuild not taken; the board records what the copy IS and boards it accordingly (the S1 precedent). The board's real archetype spread is H14 / H04 / H14, not the cited H14 / H04 / H07. The three doors are given three genuinely different visual strategies (period room / hands and volume / bare number over a printout wall) precisely because the copy-level differentiation between A and C is thin, and the human should know the spread is two archetypes across three hooks before the test reads. ③ F3 (Hook B's line one qualifies the speaker, not the viewer). Copy untouched, mitigated storyboard-side and this is the strongest visual disposition on the board: Hook B's t=0.4s card reads "THE RANGE THAT DECIDES YOUR THYROID IS NORMAL" — the viewer's door is open on screen while the VO is still delivering the credential, so the muted viewer is qualified in line one even though the spoken line qualifies the narrator. ④ F4 ("most plans" vs the file's "many plans"). Copy untouched. Boarded as a hard staging constraint in the opposite direction: no payer name, no insurer logo and no real plan document appears anywhere in beat 7. The claim-denial artifact is a generic, unbranded explanation-of-benefits line item. The board will not put an institution's name on screen behind a quantifier the copy softened. ⑤ F5 (product-entry depth reported on two bases). Recorded honestly in the board's own numbers: product enters at beat 12 [3:10] = 79.9% of boarded runtime, which on the like-for-like basis is 78.9% word-depth against the donor's 80.5% (1.6pp shallower) and on the brand-name basis is 82.0%. No compensating visual delay was added — the review explicitly ruled against structural change here, and the Kindled bottle simply does not exist on this board before 3:10.

Also carried, not "fixed" (review C1 to C5): the systemic-villain framing (C1), beat 13's "Not the number. The ENTRIES." (C2), the staccato noun-phrase fragments (C3), Hook B's single first-person credential (C4), and the female re-cast (C5). A downstream pass must not simplify any of them. C2 in particular is load-bearing and the board stakes a visual on it (see beat 13).

Hook-variant handling (Echo 1×3 contract)

Beat 1 exists as THREE storyboard variants (1A / 1B / 1C) — interchangeable opening sections, all boarded at 12s, all 24 words. Beats 2 to 13 are the shared locked-body storyboard and do not shift by a single frame between variants (this is the first board in the chain where that is true; on S1 and S2 the hooks differed in length). All three variants end on the IDENTICAL seam state: a paper macro, institutional beige grade, the caption field cleared, the camera already drifting — so any variant cuts straight onto beat 2's first line ("Nothing about a single body changed that year.") with no re-establishment. There are zero variant-conditional body elements on this board (S2's finding-3 trap does not fire here — no hook plants a visual receipt that the body pays off).

Delivery markup

The script carries exactly one bracketed token: [pause] ×1, in beat 12 on the mid-narrative product hinge. It is stripped from all on-screen text and boarded as a single tight hold (see beat 12). CAPS emphasis in the copy (MILLION · FINE · GOOD · CONTROL GROUP · ORDER · ARRIVED · REJECTED · HEALTH · PAPERWORK · LIFE · STARTED · GAP · IN · CONVERSION · RATIO · ENTRIES) is voice only and is never burned as text unless a card is separately specified below.

Pause budget, stated honestly. 1 delivery token × 0.8s = +0.8s on every variant (the hook variants carry no token, so A, B and C are identical). Against the band ceiling of 251s (4:11):

All four bases are inside the band on all three variants, and the headroom is the widest any board in the Echo chain has carried at this beat count. It is not spare runway. The donor's defining property is that it gives the viewer no room to leave; the surplus belongs to picture (extra cuts, longer holds on the four under-dense beats), never to silence, never to a second pause. Editors must not add a second bracketed pause and must not pad the hook past 12s.

Thirteen-beat discipline (donor budget preserved)

13/13 donor beat order reproduced, no beat added, dropped or re-ordered. Boarded shares track the donor's duration shares closely: heaviest beat is the Offer at 25s / 10.7% (donor 23s / 10.6%), joint-second are the Protagonist and the Reframe at 23s / 9.9% each (donor 21s / 9.6% each), lightest is the Single Variable Reveal at 9s / 3.9% (donor 8s / 3.7%). The donor's defining shape is the EVEN curve — nothing above 10.6%, no swell anywhere — and the board holds it: no beat on this storyboard gets a "big moment" treatment. There is exactly one tempo change in 3:53 and it is the packaging beat.

This account has never shipped a D10 cinematic documentary. There is no in-house precedent, and this concept carries a specific hazard the other four in the batch do not: the entire angle rests on a real, dated, citable institutional anomaly — the 2003 TSH reference-range change — and the ad names hard numbers (2.5, 4.5, thirteen million, two percent a month, permanent after two years) that are historical facts about a range, not outcome claims. Every one of those numbers is file-sourced (winners.md §2.5 Voltage Card, class AS-IS). The board's single largest job is to make them read as documentary evidence rather than as an invented statistic card, because a blocky number floating on a coloured field is exactly what a fabricated stat looks like, and this ad cannot afford to look fabricated for even one cut.

The two OST classes (borrowed from the S1 finding, which recommends exactly this split whenever a donor's grammar carries on-screen numeric evidence)

The two classes never share a frame for more than one second.

The binding rule (this is the whole contract)

> Every NUMBER SLAB must be within ONE CUT of the DOCUMENT that sources it.

A slab may precede its document or follow it, but a slab may never be more than one cut away from a filmed piece of paper carrying that number. There is no such thing on this board as a number that lives only in typography. This is the mechanism by which "2.5 → 4.5" reads as a reference range someone printed, and "13,000,000" reads as the size of a stack of results, rather than as two statistics an advertiser typed.

One declared exception, with its resolution: Hook C. The script mandates a cold open on the bare number with no context and roughly 2s of dwell. Boarded as follows: the slab is full-frame from frame one over a field that is already the printout wall at heavy defocus (legible as texture, not as paper), and the lens racks to focus by t=2.0s so the document resolves under the number inside the dwell. The shock of the bare number is preserved; the evidence arrives inside two seconds rather than after a cut. This is the only place on the board where a slab precedes its document, and it is declared rather than discovered.

The provenance ladder for the 2003 material specifically

The 2003 change is never asserted by a card. It is boarded as three physical artifacts in a fixed order, and the number 13,000,000 does not land until the third is on screen:

1. THE DECISION — a period conference room and a bound guideline volume. What a committee looks like after it has left.
2. THE CONSEQUENCE — the printed reference-range column on a lab requisition, upper bound reading 4.5. What the decision became on paper.
3. THE SCALE — a wall or a stack of identical result printouts. What the decision became in bodies.

Beat 8 (True Discovery) re-walks the same ladder in reverse (where the range came from), which is why it can re-use the 13,000,000 slab without it reading as a repeat: the second time, the viewer has seen the curve it came off.

Hard evidence rules (QC gates, not preferences)

• No invented document, ever. No fabricated journal citation, no fabricated study title, no fabricated abstract, no fabricated screenshot of any real organization's page.
• No real institution appears anywhere on this board. No association name, no payer name, no insurer logo, no hospital mark, no journal masthead, no letterhead belonging to anyone. Every artifact is a generic, unbranded period-correct document: an unmarked bound clinical volume, an unmarked lab requisition, an unmarked explanation-of-benefits line item, an unmarked results printout. The copy indicts a system, not a named body, and the picture must not name one the copy did not.
• No number appears on screen that the narrator does not speak. The full permitted on-screen numeric set is: 2003, 2.5, 4.5, 13,000,000, 3.6, 3.8, 3.4, 3.9, 2%, two years, eleven years, 2014, 2016, 2018, 2021, 2.8, 3.4, 3.9, 12 minutes, 30 days. No price, no discount, no stock count, no review count, no percentage, no p-value, no sample size.
• No lab-outcome result claim in any frame. brand_bible §9.2. Carol's numbers only ever appear as her IN-RANGE historical readings. No frame anywhere shows a number improving. Beat 13's payoff is explicitly the handwriting column, not the number column, and that distinction is load-bearing copy (review C2).
• The bell curve at beat 8 is hand-drawn on paper, by a hand, in the frame, with only 2.5 and 4.5 marked. It is not a motion-graphic chart with an axis. A charted curve with tick marks asserts a dataset; a drawn curve asserts an explanation, which is what the copy actually makes.
• Carol is never seen in full face. Hands, sleeves, the back of a head, a jaw cropped by the frame edge, a figure out of focus at a window. This protects the documentary register and it is the difference between a documentary subject and a testimonial actor.

The five worlds (the board's whole location vocabulary)

Colour arc, declared: the base grade is W1/W2/W4 institutional beige and fluorescent, and it owns beats 1, 2, 7, 8, 9 and 11. Warm domestic is an EXCURSION, not a phase — it is owned by beats 3, 4, 6 and 10 (Carol's world and the viewer's phone) and it always returns to institutional on the cut out. Clinical white is the terminal state — it appears once early as the mechanism insert (beat 5), then takes the film from beat 12 to black. This tri-partite arc is the script's own DELIVERY NOTES instruction, boarded as an ownership map so an editor cannot accidentally warm a system beat.

Cutting law

1 to 2s cuts, everywhere, with no held frames. Nothing on this board is allowed to breathe except the packaging beat, which stretches to roughly 3s cuts and only there. There is no dissolve anywhere in 3:53 — hard cuts exclusively, with two declared exceptions (a rack-focus in Hook C and a slow push in beat 1A/beat 10), both of which live inside a single shot rather than between two. A held frame anywhere else is a format failure, because the donor's measured mechanism is that the VO and the picture together give the viewer no exit point.

Generate long and wide, cut short and tight

At 1 to 2s the board needs roughly 150 cuts across 233 seconds and it has roughly 50 distinct assets. Those numbers reconcile in exactly one way and it is a production instruction, not an accident: every generated asset is made LONGER and WIDER than the cut needs (generate 5 to 8s at maximum resolution with headroom on all four edges), so post can pull two to four distinct cuts from a single generation by re-framing, pushing and re-timing without a re-generation. A board that generates 150 assets for 150 cuts will not be finished, and a board that generates 50 tight 2-second clips will run out of picture in the edit. Generate 50, cut 150.

Rotation: N/A this run — the donor is PINNED per the run order (SEED, binding; no rotation re-query executed, reference_usage_log.jsonl not read as an EXCLUDE source, per the task directive). Supporting anchors are the script's own CITED hook-formula exemplars plus the batch's only other no-presenter donor, all declared choices per §2f. Log append SKIPPED (echo run — handled at package).

Pool note: no fresh §2b/§2c query ran. This follows the R3 finding-3 convention — a pinned SEED board inherits its anchor set from the donor plus the mini-brief's and script's citations, and re-queries only if a beat type is left uncovered. Nothing is uncovered: the donor is itself a D10 and supplies grammar for all thirteen beats, and the one thing it cannot supply (how a no-presenter VO film is geometrically built when the account has never made one) is covered by Anchor E under the S2 finding-4 convention.

Set composition note: all five anchors are video. No static demotion applies on this board — a first-of-mode film had no business drawing a single-frame anchor, and none surfaced in the cited lineage.

[Anchor A — CORNERSTONE] 27357369823856563 — lirae — winner_score 52.24 — video (THE DONOR, SEED — pinned CORNERSTONE) Hook pattern: H14 Forbidden Lore over a 12s open. Third-person dated anomaly lede with zero second-person address; second-by-second [W]: s1 extreme close-up of a hash brown being bitten (contextual contradiction — health ads show salads), s2 "HE ATE MCDONALD'S EVERYDAY" in high-contrast caps with a 3s dwell, s3 hard cut to a glucose meter reading 121, which is the commitment moment. Thumb-stop 9.5/10 [W] On-screen text style: Hormozi-blocky high-contrast caps, numbers rendered MASSIVE ("140", "92", "29%"); phrase-level, never a continuous caption track B-roll cuing: the film is nothing but B-roll — there is no presenter. Third-person documentary POV intercut with first-person product shots; extreme close-ups to force visceral response (biting, blood on a strip, powder scooping); raw-ingredient footage; clinical inserts (meters, doctors) Transition style: very fast, 1 to 2s cuts, no pauses. Slows only for the packaging shot Mute-test asset: massive numbers over visceral close-ups. The format IS the mute test — captions and objects, no talking face to depend on

Best-fit beat type in our script: ALL (SEED full-pattern borrow). Primary: the dated-anomaly hook (1A), the single-variable reveal (2), the hyper-specific protagonist (3), the accidental-control-group protocol (4), the one-sentence mechanism (5), the dated verification ladder (6), the suppression explanation (7), the sourced forbidden fact (8), the systemic indictment (9), the second-person reframe (10), the restated wager (11), the late packaging slow-down (12), the risk-free close (13)

Tooling stack inferred: documentary capture plus archival plus clinical inserts plus heavy post captions — for us, AI-generated cinematic story worlds plus practical product macros plus post text. No presenter in either stack, which is why this donor maps to our target with zero adaptation strain on the one axis that usually strains. ⚠️ Honest inheritance: ws 52.24 is below the 60 winner bar and this donor executes EUREKA's non-winner structural signature (13 beats with a full disqualification arc). It is an exemplar of SHAPE and DELIVERY, not a proven winner — recreate its D10 grammar and its 221.8 wpm with confidence; the structural caveat is recorded, not acted on

[Anchor B] 1269484834872935 — try-forge — winner_score 64.06 — video (H14 ARCHIVAL-CINEMATIC grammar; the script's own cited H14 cross-reference) Hook pattern: the textbook H14 Forbidden Lore lede — "In 480 BC, 300 Spartan soldiers…" Dated, third-person, institutional, zero second-person address, commitment number early On-screen text style: text-scroller — the film carries its argument in typography over cinematic imagery, with no spoken voiceover at all and a dramatic instrumental bed B-roll cuing: high-contrast atmospheric AI-generated period imagery (gritty warriors against a stone wall) — the archival/epic look is entirely generated and it works, which is the single most useful fact this anchor contributes to a board with no presenter and a 2003 period requirement Transition style: cinematic, driven by the instrumental Mute-test asset: the generated period image plus the typography, alone. It has no other option and it scores

Best-fit beat type in our script: Hook 1A (the dated period lede and its archival grammar) — PRIMARY. SECONDARY and load-bearing across the board: the "period authenticity is generatable" proof, which is what licenses beat 1A's 2003 conference room, beat 1B's bound volume and beat 8's institutional world to be AI-generated rather than sourced as real archive. FLAG (VS6 cross-adaptation logged): its ANCIENT framework, its 373.7 wpm text-scroller pacing and its wordless construction are all rejected wholesale — our film is VO-led at 221.8 wpm on an EUREKA spine. Borrowed at archival-look and typography-weight level only, never at spine or delivery level

Tooling stack inferred: AI-generated cinematic period imagery plus post typography plus a scored bed. Maps 1:1 to our generation layer

[Anchor C] 1336103951757183 — trysculptique — winner_score 65.63 — video (HOOK-B H04 FORMULA source) Hook pattern: H04 Credential Cold Open — "I'm a urologist. Here are 15 questions about prostate problems, and the answers your doctor won't give you." The catalog's strongest shape (40.2% winner share); trust transfers before any claim is made On-screen text style: question-numbered structure; the credential and the index carry the text load B-roll cuing: minimal in the open — an older credentialed-looking male actor in front of home window blinds carries it Transition style: sustained direct address, no early cutaway Mute-test asset: the face and the burned promise

Best-fit beat type in our script: Hook 1B ONLY (the credential-first door). FLAG (VS6 cross-adaptation logged, and this is the board's largest single adaptation): its entire staging is an on-camera credentialed persona, which D10 forbids and product_facts §5 separately forbids (Kindled is not a fake-doctor persona network). Borrowed at FORMULA level only and re-staged as the board's hands-and-volume open: the credential is shown as an object (a working hand turning the pages of the guideline volume it reads for a living) rather than worn as a costume. The narrator's single "I" is heard once and never embodied. Review C4 cleared this construction explicitly

Tooling stack inferred: real credentialed-actor talking head — rejected wholesale. Our restaging needs no talent and no face

[Anchor D] 1957809038129580 — elare — winner_score 81.52 — video (HOOK-C SHOCK-FACT formula source; the set's highest winner_score) Hook pattern: H07 Lecture Bait in its shock-statistic form — "42% of men who have their first heart attack die within a year." A bare, startling number in line one with no story, no credential and no offer; the gap between what the viewer believed and the stated fact IS the hook On-screen text style: the statistic itself carries the frame B-roll cuing: split-screen pairing a man clutching his chest with a freshly dug grave Transition style: fast, fear-register cutting Mute-test asset: the coffin visual plus the number

Best-fit beat type in our script: Hook 1C ONLY (the bare-statistic cold open, before any history). TWO FLAGS, both VS6 cross-adaptations logged: ① its mortality imagery is rejected wholesale and on doctrine — the terminal on this concept is a file-sourced, explicitly non-mortal permanence clock, and the standing rule is never invent mortality. No grave, no body, no medical emergency appears on this board. ② per review F2, our Hook C's actual copy ("Thirteen million women were declared healthy in 2003") is a dated historical institutional lede, which routes to H14 by the catalog's own tie-break, not to the cited H07. The board carries the copy as locked and boards it as what it IS — the visual treatment is built for a bare-number-then-evidence read, and the archetype mismatch is recorded rather than papered over

Tooling stack inferred: split-screen fear composite — rejected; ours is a slab over its own defocused evidence

[Anchor E] 26907486858952165 — resilia — winner_score 65.38 — video (NO-PRESENTER VO GEOMETRY reference — the S2 finding-4 slot) The batch's only other donor with no on-camera presenter (D04 Faceless Anatomy & Timeline Narrator, the catalog's #2 carrier at 32.2%, +5.5pp — "the highest-confidence overweight after D10"). Included under the convention S2 established: when a board is the account's FIRST in a delivery mode, reserve one anchor slot for a same-mode geometry reference regardless of niche or angle fit. D10 has never run on this account, and "how is a no-presenter VO film actually built, shot to shot" was the single thing this board most lacked Hook pattern: H03 Time-Lapse Trial — a 3D arterial cross-section carrying frame one with the VO over it On-screen text style: sparse; the imagery is the argument and the text supports it B-roll cuing: the transferable lesson — a faceless film sustains attention by alternating two distinct visual registers (clean explanatory graphics against tactile real-world footage) rather than by staying in one. Pacing is explicitly variable: fast cuts on the montages, deliberate slow-downs on the explanatory beats so the mechanism can be processed Transition style: variable by register, and the alternation is itself the grammar Mute-test asset: the graphic plus the sparse burn — no face anywhere, and it wins without one

Best-fit beat type in our script: global spine (how a presenter-free film is constructed) — PRIMARY; and specifically beat 5, the clinical-white mechanism insert, which is the one beat on our board that behaves like an explanatory graphic and is also the one beat with 6.3s of surplus picture time. Its "slow down for the mechanism" instruction is the direct source of how that surplus is spent. FLAG (VS6, at role level): product_facts §5 names resilia by name as a positioning lane Kindled does not play (fake-doctor persona network). The faceless-VO carrier is imported; the persona lane is not. Its counted scarcity ("642 pouches left") and its discount close are also explicitly not imported — ours is demand-driven scarcity with no invented count

Tooling stack inferred: 3D/AI explanatory animation plus UGC lifestyle plus product tactility plus VO — the same four-layer stack as ours minus the archival layer

Global visual spine (applies to every beat unless overridden): 9:16 vertical. THERE IS NO PRESENTER AND NO PERSON IS EVER ESTABLISHED ON CAMERA. The default frame is a documentary subject: a document, an object, a room, or a pair of hands. Camera behaves like an observational documentary camera — locked or on a slow deliberate drift, eye-level to objects, never handheld-jittery and never gimbal-glossy. Extreme close-ups are the board's dominant grammar (donor-exact: "extreme close-ups to force visceral response"). Two OST classes only — NUMBER SLABS (Hormozi-blocky, massive, full-frame) and DOCUMENT MARKS (ink-behaviour marks on filmed paper) — never sharing a frame for more than a second, and every slab within one cut of its document (THE EVIDENCE CONTRACT). Grade is institutional beige and fluorescent by default; warm domestic is an excursion owned by beats 3, 4, 6 and 10; clinical white is owned by beats 5, 12 and 13's final card. CUTTING: 1 to 2s hard cuts throughout, no dissolves, no held frames, the single declared slow-down at beat 12. AUDIO: a low, flat, unresolving bed from 0:00 with no swell (the donor's arc is built by accumulation, not volume); functional SFX only — a paper-handling sound on every document cut, a soft mechanical stamp on every NUMBER SLAB entry, a pen-scratch on every handwriting mark. Zero dead air anywhere. The bed thins by a half-step at the packaging beat and resolves out under the end frame. CHARACTER LAW: Carol exists as hands, a sleeve, a back, an out-of-focus figure — never a full face, never a piece-to-camera, never a smile. INSTITUTION LAW: no real organization's name, mark, letterhead or page appears in any frame.

PRIMARY_LAYER: higgsfield-generate (NO avatar tool, NO presenter, NO Soul ID — this format has no performer)
SHOT_1 (0:00-0:02): GPT Image 2 keyframe -> "photorealistic cinematic still, an emptied 2003-era
        institutional conference room, rows of chairs pushed back at odd angles, a ceiling projector
        throwing a blank warm rectangle onto a pull-down screen, one abandoned water glass, fluorescent
        overhead light with a faint green cast, institutional beige walls, documentary film grain,
        completely empty of people, 9:16 vertical" -> Seedance 2.0 i2v, SLOW PUSH IN, generate 6s
SHOT_2 (0:02-0:05): practical or generated macro -> a bound unbranded clinical guideline volume lying
        open, a printed numeric reference-range line on the page, copy-stand flat even light, extreme
        close-up, shallow depth. NO institution name, NO logo, NO masthead anywhere in frame
SHOT_3 (0:05-0:08): generated macro -> an unbranded lab requisition form, slow push down the page
        toward a printed reference-range column, cold fluorescent white, flat light
SHOT_4 (0:08-0:12): SHOT_3 re-frame (tighter crop on the column) hard-cut to a generated wall calendar
        with two years of pages turned back, institutional beige
ON_SCREEN_TEXT (post): SLAB "2003" t=0.3-1.9 | MARK "THE THYROID RANGE MOVED" t=0.9-1.9 |
        SLAB "2.5 -> 4.5" t=2.1-5.0 (3s dwell, donor-mandated) | SLAB "13,000,000" t=5.4-8.0 |
        MARK "TWO YEARS. THEN PERMANENT." t=9.4-beat end
SFX: paper handling on each document cut; a soft mechanical stamp on each slab entry; low flat bed from t=0
TRANSITION (editorial): hard cuts only, no dissolve; hard cut out to the shared seam state
VO DIRECTION: female investigative narrator, forensic, absolute facts, true-crime-podcast seriousness.
        FLAT AND FAST. No warmth, no aside, no company-"we". Receipts spoken bare: "2.5", "4.5",
        "thirteen MILLION" get NO softener and NO approximation. Uncontracted read (F1 — hold it).
        The hook runs 12s for 24 words = ~120 wpm: this is the ONE slow beat, and the slowness is
        CAPTION DWELL, not narrator hesitation. She still reads flat-fast; the picture holds.
DURATION: 12s (24 words; 6.5s of speech, 5.5s of caption dwell — donor parity, donor hook is 21w in 12s)

PRIMARY_LAYER: higgsfield-generate + hands-only practical pickup (page-turn is one of the board's
        three highest AI-artifact actions — REAL-SHOOT PREFERRED for SHOT_1, see cross-ref #4)
SHOT_1 (0:00-0:04): hands only, turning the pages of a bound unbranded clinical guideline volume,
        single warm desk lamp, institutional dark surround, night, adult unmanicured hands, plain watch,
        NO face, NO shoulders, nothing above the forearm ever enters frame. Locked, slightly overhead,
        shallow depth. Generate 7s / shoot 15s so post can pull two cuts
SHOT_2 (0:04-0:08): tighter macro on the open volume, a fingertip resting beside a printed range line.
        NO institution name, NO logo anywhere
SHOT_3 (0:08-0:12): generated blood-draw tray, extreme close-up on an unbranded tube label,
        cold fluorescent, flat light; second cut = tighter crop of the same generation
ON_SCREEN_TEXT (post): MARK "THE RANGE THAT DECIDES YOUR THYROID IS NORMAL" t=0.4-3.6 (THE F3 FIX) |
        SLAB "RECLASSIFIED" t=4.2-5.6 | SLAB "13,000,000" t=5.8-8.2 |
        MARK "TWO YEARS INSIDE IT" t=9.6-beat end
SFX: page turn (practical), paper handling, mechanical stamp per slab
VO DIRECTION: the "I" in "I read reference ranges for a living" is a CREDENTIAL, not an endorsement,
        and it is the only first-person moment in 867 words. Deliver it as a flat statement of
        occupation, never as a boast and never as a confidence. She introduces herself once and
        never refers to herself again — do not let a take add warmth here to "pay it off" (review C4).
DURATION: 12s (24 words; same caption-dwell structure as 1A)

PRIMARY_LAYER: higgsfield-generate
SHOT_1 (0:00-0:02): the printout wall from SHOT_2, generated at heavy defocus, then RACKING TO FOCUS
        across t=0.0-2.0s. Generate as one continuous rack, 4s, so post can time the rack precisely.
        THE SLAB SITS OVER IT FROM FRAME ONE — this is the declared evidence-rule exception
SHOT_2 (0:02-0:07): GPT Image 2 keyframe -> "a wall covered edge to edge in identical printed lab
        result sheets, pinned in a grid, cold fluorescent light, flat even illumination, institutional,
        no branding of any kind, documentary film grain, 9:16 vertical" -> Seedance 2.0 i2v,
        STEADY LATERAL TRACK, constant rate, generate 8s (post pulls 2 to 3 cuts)
SHOT_3 (0:07-0:12): extreme close-up macro, a fingertip resting on the printed word "NORMAL" in a
        results column, cold fluorescent, shallow depth. Hands-only, no body
ON_SCREEN_TEXT (post): SLAB "13,000,000" t=0.0-2.4 | MARK "TOLD YOUR LABS WERE NORMAL" t=2.6-6.6 |
        INK CIRCLE drawn around the word NORMAL, one stroke, t=8.2-beat end |
        MARK "TWO YEARS IN" t=10.0-beat end
SFX: a low fluorescent hum under shot 2; paper handling; ink scratch on the circle
VO DIRECTION: the bare statistic is delivered with NO build and NO emphasis — it is stated the way a
        report states a finding. The flatness is what makes it land as fact rather than as a claim.
DURATION: 12s (24 words)

PRIMARY_LAYER: higgsfield-generate + practical macro
SHOT_1 (0:12-0:15): macro, a printed reference-range line on a page; a hand slides a second later page
        over the first, same line, different upper number. ONE continuous move, shot practically if the
        hands pickup day happens (artifact risk), otherwise generate 5s and cut on the slide
SHOT_2 (0:15-0:18): a single unlabelled blood tube in a rack, cold fluorescent, LOCKED AND DEAD STILL,
        zero camera movement. Generate 5s
SHOT_3 (0:18-0:21): a woman seen from behind, out of focus, standing at a clinic window, waiting,
        long lens, shallow depth, the window slightly blown out. NEVER her face. Generate 5s
ON_SCREEN_TEXT (post): SLAB "ONE NUMBER MOVED" t=1.4-4.0, the word ONE at slab scale
SFX: paper slide (practical), fluorescent hum, bed continuous
VO DIRECTION: "the same blood, the same exhaustion, the same woman, read as FINE" is a list delivered
        without escalation. FINE lands in CAPS as a pitch-drop, never as volume (no beat swells).
DURATION: 9s (34 words @ 227 wpm — at pace)

PRIMARY_LAYER: higgsfield-generate + hands-only practical pickup (notebook writing = artifact risk #1)
CHARACTER SPINE (generate + lock BEFORE anything else on the board):
        THE CANONICAL CAROL KEYFRAME — mid-50s woman's hands, unmanicured, one plain wedding band,
        a specific soft cardigan sleeve in a muted colour, a specific spiral notebook with ruled
        two-column pages, a specific pen, a specific kitchen table with visible wood grain.
        NANO BANANA reference-chain EVERY later Carol clip from this single keyframe.
        NEVER her face. Hands, sleeve, back of head, out-of-focus figure only.
SHOT_1 (0:21-0:25): Carol keyframe, hands on the table, notebook open, pen resting, warm domestic,
        locked overhead. Generate 7s
SHOT_2 (0:25-0:29): generic clinic corridor, institutional beige, locked centred; second cut = tighter
        crop on a numbered door from the same generation
SHOT_3-6 (0:29-0:35): FOUR macro requisition forms, unbranded, each with one printed reading;
        the ink circle is a POST animation drawn on in one stroke, timed to the spoken number
SHOT_7 (0:35-0:39): extreme close-up, the notebook's two-column spread, dated numbers left,
        handwriting right. Carol-chained
SHOT_8 (0:39-0:42): macro, the handwritten words "Made of lead" being written, pen in frame.
        REAL-SHOOT PREFERRED (writing hands are the board's worst AI case)
SHOT_9 (0:42-0:44): a wall clock, then a calendar. Institutional beige
ON_SCREEN_TEXT (post): SLABS "3.6" / "3.8" / "3.4" / "3.9" t=8.5 / 9.9 / 11.3 / 12.7, each on the frame
        carrying its own circled printed number | MARK "FINE." small, corner, t=13.4 |
        HANDWRITING "Made of lead" written on t=17.8 | SLAB "2% A MONTH" t=20.0-21.4 |
        SLAB "AFTER TWO YEARS, GONE FOR GOOD" t=21.6-beat end
SFX: pen scratch on every handwriting mark; ink scratch on every circle; paper on every requisition cut;
     a single soft clock tick under the last two shots and NOWHERE else on the board
VO DIRECTION: THE ONE FEAR SPIKE. Deploy it hard and deploy it ONCE. The mechanism is the donor's:
        terrifying stakes delivered CALM read as reportage, not salesmanship. The permanence clock
        ("two percent of her metabolism, gone. After two years, gone for GOOD.") is spoken FLAT,
        bare, with no softener and no approximation. GOOD is a pitch-drop, not a volume lift.
        Nothing in the remaining 3:09 is allowed to be louder than this and this is not loud.
DURATION: 23s (95 words @ 248 wpm — the board's second-densest beat; cut tight)

PRIMARY_LAYER: higgsfield-generate (Carol-chained from the canonical keyframe)
SHOT_1-4 (0:44-0:49): THE SAME kitchen table, IDENTICAL locked framing, four different lights —
        winter morning / summer evening / grey afternoon / night. Generate as FOUR relights of ONE
        keyframe (Nano Banana relight chain), NOT as four separate scenes. Frame match is the beat.
SHOT_5-8 (0:49-0:54): 6am alarm clock / a work badge on a hook / the same front door / the same coffee
        mug. One per cut, locked macros, warm domestic, Carol-chained
SHOT_9 (0:54-0:58): a medicine shelf, one bottle, same position, dust undisturbed, SLOW PUSH.
        NOTE: no selenium bottle, no iodine bottle, no thyroid supplement of any kind on this shelf —
        their presence would fight the beat-12 reveal. QC gate, not a preference
SHOT_10 (0:58-1:01): the notebook closed, then opened to a full spread. Carol-chained
SHOT_11 (1:01-1:04): extreme close-up, the left column running down the page, year after year
ON_SCREEN_TEXT (post): MARK "SAME BODY. SAME JOB. SAME ALARM." t=6.0-10.0 |
        SLAB "THE CONTROL GROUP" t=15.0-18.0
SFX: paper, a single alarm-clock tick, a shelf-bottle set-down; bed continuous, no swell
VO DIRECTION: the list of negations ("Never changed her diet. Never changed her dose. Never changed
        doctors.") is delivered as an INVENTORY, flat and even, one item per breath, with no
        accumulation of emphasis. CONTROL GROUP is a pitch-drop. The beat is at pace (234 wpm) —
        do not slow down for the montage; the picture is already doing the slowing.
DURATION: 20s (78 words @ 234 wpm — at pace)

PRIMARY_LAYER: higgsfield-generate (graphic register — the one beat that behaves like an explainer)
SHOT_1 (1:04-1:08): clinical white field, a printed order slip placed into frame by a hand and slid
        away out of frame in one move. Locked, dead still, high key, shadowless
SHOT_2 (1:08-1:13): split composition — the order slip left, an EMPTY SPACE right. Locked, dead still.
        This is the board's longest hold; generate 8s
SHOT_3 (1:13-1:17): extreme close-up, a lab report line showing ONE measured value and nothing under it
SHOT_4 (1:17-1:20): the empty right side alone in frame, locked
ON_SCREEN_TEXT (post): SPLIT CARD "THE ORDER" (left) / "THE PACKAGE" (right) t=4.4-9.0
SFX: paper slide; the bed thins very slightly here and ONLY here before beat 12; no silence
VO DIRECTION: 135 wpm — the slowest read on the board, and it is a DENSITY fact, not a delivery
        instruction: there are simply fewer words here. She does not slow her articulation and she
        does not add pauses. ORDER and ARRIVED are pitch-drops. "It is not the hormone." keeps its
        uncontracted form (F1).
        THE SURPLUS IS PICTURE, NOT SILENCE — the bed and the held frames carry 6.3 seconds.
DURATION: 16s boarded (36 words = 9.7s of speech; 6.3s of deliberate picture time)

PRIMARY_LAYER: higgsfield-generate (Carol-chained) + hands-only pickup for the two handwriting inserts
SHOT_1 (1:20-1:23): extreme close-up, the notebook's right column at "2014"; then a desk at 2pm, flat
        light. Carol-chained
SHOT_2 (1:23-1:27): car interior, driver's seat reclined, a coat over the chest, lunch-break light.
        Shot from the passenger side, NEVER her face. Carol-chained
SHOT_3 (1:27-1:31): a July window, green outside, a heavy knit sleeve in the foreground; then a macro
        on the handwritten line "I don't recognize who I've become" (REAL-SHOOT PREFERRED)
SHOT_4 (1:31-1:34): a workplace desk, a head down and out of focus, a bag still on the chair.
        Shot from behind. Carol-chained
SHOT_5-8 (1:34-1:38): FOUR requisition macros, unbranded, each with an in-range printed number
ON_SCREEN_TEXT (post): HANDWRITING "2014" t=0.6 / "2016" t=3.4 / "2018" t=7.2 / "2021" t=11.4 |
        HANDWRITING "I don't recognize who I've become" t=9.0 |
        MARK "FINE." x4, hard stamps, t=15.0 / 15.9 / 16.8 / 17.6
SFX: pen scratch per handwriting mark; a hard stamp hit per "FINE."; paper per requisition cut
VO DIRECTION: FASTEST BEAT ON THE BOARD (283 wpm). The four "Fine. Fine. Fine. Fine." are four
        identical flat words — no rising anger, no irony, no colour. The flatness is the indictment.
        Do not let a take play the last one differently.
DURATION: 18s (85 words @ 283 wpm — cut tighter than anywhere else)

PRIMARY_LAYER: higgsfield-generate
SHOT_1 (1:38-1:42): extreme close-up, a GENERIC UNBRANDED explanation-of-benefits line item, a status
        column at the end of the row, slow lateral drift. NO payer name, NO insurer logo, NO plan
        document, NO real institution mark ANYWHERE (review F4 disposition + INSTITUTION LAW)
SHOT_2 (1:42-1:46): macro, an unbranded lab order form with a printed test box left UNTICKED, a pen
        resting past it
SHOT_3 (1:46-1:50): a wall clock reading twelve minutes past the hour, locked, institutional beige
SHOT_4 (1:50-1:53): a corridor of identical closed doors receding, long lens, locked, centred
ON_SCREEN_TEXT (post): INK UNDERLINE under the status field t=1.8 | SLAB "REJECTED" t=3.6-5.0 |
        MARK "NOT REIMBURSED" t=6.0-8.6 | SLAB "12 MINUTES" t=9.6-11.4
SFX: paper; a single hard stamp on REJECTED (the beat's one heavy hit); the clock is SILENT here
     (the only tick on the board is beat 3's — do not repeat it)
VO DIRECTION: cold indignation, and cold is the operative word. "Nobody suppressed Carol's notebook.
        Nobody had to." is indictment-by-denial and must be read as a flat finding, never as a sneer
        (review C1: a literal reader would false-fail this line; it is correct as written).
        REJECTED is a pitch-drop. Noun-phrase fragments ("Twelve minutes an appointment. One number
        on the sheet.") are donor-native and read clean at pace (review C3).
DURATION: 15s (54 words @ 216 wpm — at pace)

PRIMARY_LAYER: hands-only practical pickup STRONGLY PREFERRED (a hand drawing a curve is the third of
        the board's three artifact-risk actions) + higgsfield-generate for shots 3 and 4
SHOT_1 (1:53-1:58): blank paper, a hand, a pen. A bell curve DRAWN BY HAND in one continuous stroke.
        Then two marks added to the baseline. Locked overhead, flat copy-stand light.
        THIS IS NOT A MOTION-GRAPHIC CHART. No axis, no ticks, no gridlines, no data points.
        A charted curve asserts a dataset; a drawn curve asserts an explanation, which is what the
        copy actually makes. HARD RULE (EVIDENCE CONTRACT)
SHOT_2 (1:58-2:03): tighter macro, the hand DRAGS the right-hand mark across to a second position.
        One continuous move, in frame
SHOT_3 (2:03-2:07): the Hook-C printout wall generation, tracked the OPPOSITE way (re-use, re-time)
SHOT_4 (2:07-2:11): the Hook-A/B bound guideline volume, closing. Re-use the same generation
ON_SCREEN_TEXT (post): INK MARKS "2.5" t=3.2 and "4.5" t=4.0, in the same ink as the curve |
        SLAB "13,000,000" t=11.0-13.4 | MARK "THE PAPERWORK DID" t=15.4-beat end
SFX: pen on paper (practical); paper; a soft stamp on the slab; the volume closing
VO DIRECTION: "It was never a measurement of HEALTH." HEALTH is a pitch-drop. The line "It is the
        average of sick people and well people, and it is not where well people sit" is quoted from
        mechanism_library §6's own Variations line and keeps its uncontracted form (F1). Cold
        indignation, still flat, still no volume.
DURATION: 18s (59 words @ 197 wpm — slightly under pace; the surplus goes to the drawing, which is
        the board's most patient shot outside beat 5)

PRIMARY_LAYER: higgsfield-generate + practical macro (four generic bottles is a trivial practical set)
SHOT_1-4 (2:11-2:16): a plain surface. FOUR generic unbranded pill bottles, different shapes, set down
        ONE AT A TIME, one per cut, each landing on its own spoken prescription. Hard key, defined
        shadows. NO real drug name, NO real brand, NO pharmacy mark on any label
SHOT_5 (2:16-2:21): wider, the four bottles in a row plus A FIFTH EMPTY SPACE beside them. Locked
SHOT_6 (2:21-2:26): four generic copay receipts fanned, extreme close-up, slow drift across.
        NO amounts legible (no invented numbers on screen)
SHOT_7 (2:26-2:31): the fifth empty space alone in frame, SLOW PUSH IN
ON_SCREEN_TEXT (post): four small typed specialty labels under the bottles, t=1.4 / 2.4 / 3.4 / 4.4 |
        SLAB "FOUR PRESCRIPTIONS. FOUR SPECIALISTS. FOUR COPAYS." t=11.0-14.0
SFX: a distinct set-down per bottle (four different weights); paper on the receipts; a soft stamp on
     the slab; the bed stays flat — this beat does NOT swell
VO DIRECTION: "Nobody planned that. The incentive did not need anyone to plan it." — this is NOT
        exoneration (review C1, checked and explicitly cleared). It is a systemic attribution that
        makes the indictment worse, and the second sentence closes the escape hatch. Read both flat,
        as findings. Never as sarcasm, never as a shrug. A downstream pass must not "fix" this.
DURATION: 20s (67 words @ 201 wpm — slightly under pace; the surplus goes to the four set-downs and
        the final push, both of which are picture, not air)

PRIMARY_LAYER: higgsfield-generate (Carol-chained table) + post-composited screen content
SHOT_1 (2:31-2:35): a phone FACE DOWN on a kitchen table, warm domestic light, LOCKED, nothing happens
SHOT_2 (2:35-2:39): the phone turned over, a patient-portal login screen. LOCKED, SLIGHTLY OVERHEAD.
        SHOT AS DOCUMENTARY EVIDENCE, NOT AS UGC — no hand-held phone-in-hand framing, no selfie
        energy, no vertical-video-within-vertical-video. This is a camera looking at an object
SHOT_3 (2:39-2:44): extreme close-up, a results row on the screen, a thumb resting beside it
SHOT_4-6 (2:44-2:50): THREE requisition macros, one per spoken number, each with its printed number
SHOT_7 (2:50-2:54): back to the phone screen, SLOW PUSH into the results row
SCREEN CONTENT: composited in POST, never rendered in-generation (no model renders legible UI text
        reliably). Generic portal UI. NO real health-system name, NO real portal brand, NO logo
ON_SCREEN_TEXT (post): SLABS "2.8" / "3.4" / "3.9" t=14.0 / 15.4 / 16.8, each on its own printed source
        | SLAB "YOU ARE ALREADY IN IT" t=19.6-21.6 | MARK "GO LOOK" t=22.0-beat end
SFX: a phone turning over; paper; a soft stamp per slab; the bed holds flat
VO DIRECTION: THE REGISTER TURNS HERE and it is the only turn in the ad. Third-person reportage
        becomes direct address. Do NOT warm the voice to mark it — the turn is in the pronoun, and
        letting the read soften would cost the reportage credibility that bought the pivot.
        "It is sitting in a patient portal" / "If it is under... If it is over" keep their
        uncontracted forms (F1). IN is a pitch-drop. "Go look." is flat and final, not urgent.
DURATION: 23s (106 words @ 277 wpm — the board's densest beat; every cut under 2s)

PRIMARY_LAYER: re-use + higgsfield-generate
SHOT_1 (2:54-2:58): the beat-8 drawn curve, on the same paper, lying alone. Slow drift. RE-USE the
        beat-8 generation/shoot, re-framed and re-timed
SHOT_2-4 (2:58-3:02): a calendar page turning, three times, three cuts, slightly accelerating
SHOT_5 (3:02-3:06): a BLANK unbranded requisition form waiting on a desk. Locked
SHOT_6 (3:06-3:10): the beat-1A conference room, SAME FRAMING, projector rectangle now DARK.
        Re-use the beat-1A generation with the projector element removed/relit
ON_SCREEN_TEXT (post): NONE. Zero cards, zero slabs, zero marks. This is a hard instruction, not an
        omission — the beat's function is to be the board's one quiet frame before the product
SFX: paper, calendar pages, room tone under the final shot. The bed continues; it does NOT drop out
VO DIRECTION: the wager restated, and it is a question that is not asked as a question. Flat.
        "what a body does with the two percent a month it is still losing" is the ad's last cold
        line before the turn; deliver it as a finding, not as an appeal.
DURATION: 16s (51 words @ 191 wpm — slightly under pace; the surplus is the quiet, and the quiet is
        picture and bed, never silence)

PRIMARY_LAYER: practical macro (product) + higgsfield-generate (explanatory field) + hands-only pickup
SHOT_1 (3:10-3:13): clinical-white field, a single step drawn between two points — the CONVERSION,
        held in beat 5's explanatory register. THIS IS THE WELD: the film is still explaining at the
        moment the product enters. Do NOT cut to the bottle from a narrative frame (MC_06)
SHOT_2 (3:13-3:16): the [pause] lands here as ONE TIGHT <=0.8s HOLD on the last explanatory frame.
        Dead still. This is the board's only bracketed token and there must never be a second
SHOT_3 (3:16-3:20): THE REAL KINDLED BOTTLE, practical macro, clinical white, label toward camera,
        SLOW DELIBERATE PUSH. First appearance in the film. Label legibility is a hard gate
SHOT_4 (3:20-3:24): the dropper drawn and held; then the drop held under the tongue.
        CROPPED TO LIPS AND DROPPER ONLY — never a face, never a chin-to-forehead frame.
        REAL-SHOOT PREFERRED (mouth actions are the board's artifact risk #2)
SHOT_5 (3:24-3:28): label macro, the three named actives, one per cut, on the REAL printed label
CUTTING: ~3s cuts THIS BEAT ONLY. Everything before and after runs 1 to 2s. Donor-exact:
        it slows only for the packaging shot
ON_SCREEN_TEXT (post): MARKS "SELENIUM" t=14.0 / "ZINC" t=15.2 / "IODINE" t=16.4, each on the label
        frame carrying the printed word | SLAB "THE RATIO" t=17.0-beat end
SFX: a dropper draw; a bottle set-down; the bed THINS a half-step here (its only change) and holds thin
VO DIRECTION: the register does not become an ad. She still reports — she reports the product.
        "Kindled is a thyroid-support supplement. A berry-flavored liquid, two droppers under the
        tongue every morning with a meal" orients a COLD viewer: category, form, how it is taken,
        THEN the actives. Selenium RUNS the conversion, zinc LETS the hormone dock — hard verbs,
        never "supports", never "helps" (NH6 support-verb species: 0, hold it).
        CONVERSION and RATIO are pitch-drops.
DURATION: 18s (78 words @ 260 wpm — dense despite the slow cutting; the picture slows, the read does not)

PRIMARY_LAYER: hands-only practical pickup (handwriting) + practical macro (product) + post (badge)
SHOT_1-3 (3:28-3:37): the notebook's RIGHT-HAND column, three cuts, ONE handwritten entry per cut,
        each written on in real time. REAL-SHOOT PREFERRED (artifact risk #1). Carol-chained props:
        same notebook, same pen, same sleeve, same table as beat 3
SHOT_4 (3:37-3:42): the Kindled bottle ON THE KITCHEN TABLE — the only frame where the product and
        Carol's world share a surface. Locked. Warm domestic transitioning toward clinical
SHOT_5 (3:42-3:46): a clean guarantee card, clinical white
SHOT_6 (3:46-3:50): requisition macro, the range column, upper bound 4.5
SHOT_7 (3:50-3:53): the bottle, clinical white, LOCKED AND DEAD STILL, hold ~1.5s to black
ON_SCREEN_TEXT (post): HANDWRITING "Made it to eight without the couch" t=1.6 /
        "Wore short sleeves in March" t=4.4 / "Recognized the woman in the mirror" t=7.2
        — HANDWRITING, NEVER TYPESET (review C2) |
        BADGE "30-DAY MONEY-BACK GUARANTEE" t=15.0, ~4s hold, NO seal / ribbon / laurel |
        SLAB "4.5" t=20.0-22.0 | END FRAME bottle, no card, no URL, no logo, no button, no arrow
SFX: pen scratch per entry; a bottle set-down; a soft stamp on the badge; the bed RESOLVES OUT under
     the end frame (its only resolution in 3:53)
VO DIRECTION: pure sell and reveal, and still reported. Say the guarantee WHOLE ("a thirty-day
        money-back guarantee") and never clip it. "if your own right-hand column reads exactly the
        way it reads today, you get every dollar back" resolves inside its own sentence — do not let
        a take break it across a breath. Scarcity is DEMAND-DRIVEN ("It sells out because women who
        feel their mornings change start ordering for their sisters") — never supply-weakness, never
        a count. "Go see if it is in stock." keeps its uncontracted form (F1) and carries NO link
        register, ever. Resolve, not warmth: she does not soften at the end.
DURATION: 25s (100 words @ 240 wpm — the largest beat on the board, donor-exact at 10.7% vs 10.6%)

• NO PERFORMER EXISTS ON THIS BOARD. There is no presenter, no face, no piece-to-camera and no talent identity to keep continuous across 3:53. This is the format's defining constraint (D10: "film-grade investigative delivery with NO continuous on-camera presenter") and it removes the single largest continuity risk that S1 and S2 both carried. VH2 is therefore satisfied by world, prop and grade continuity rather than by speaker continuity, and each of those three is mapped explicitly above.
• Five declared worlds, and every beat is assigned to one or two. No beat invents a location. The two institutional worlds (W1, W2) and the system world (W4) share one grade family; W3 and W5 are the two declared departures and each has a stated owner set.
• The colour arc is an ownership map, not a drift. Institutional beige owns beats 1, 2, 7, 8, 9, 11. Warm domestic is an EXCURSION owned by 3, 4, 6, 10 and it always returns on the cut out. Clinical white appears once early (beat 5) as a plant and then takes the film from 3:10 to black. The single largest grade jump in the film is beat 11 → beat 12, and the board spends beat 11's card-free quiet buying it.
• Carol continuity is object continuity, and it is chained from one keyframe. The same hands, the same cardigan sleeve, the same spiral notebook with its ruled two-column pages, the same pen, the same kitchen table grain, across beats 3, 4, 6 and 13 — roughly fifteen clips from one canonical keyframe. She is never seen in full face in any of them. Any face drift, any second pair of hands, any second notebook = cut the clip; never ship a second Carol.
• Prop loops, all closed: the bound guideline volume (1A → 1B → 8 → 11) · the requisition form and its range column (1A → 3 → 6 → 10 → 13, and the film's last document is the same document as its second) · the printout wall (1C → 8) · the notebook's two columns (3 → 6 → 13, where the right column pays off) · the 2003 conference room (1A → 11, same framing, projector dark) · the Kindled bottle (12 → 13, and nowhere before 3:16).
• The film returns to its own first frame. Beat 11 shot 6 is beat 1A shot 1, re-lit. The film opens on an emptied room where a decision was made and comes back to it 3:06 later with the projector off. That return is why the close can say "The range is still 4.5. It is not moving back on its own." over a document the viewer has already seen twice.
• OST grammar constant: two classes only (NUMBER SLABS, DOCUMENT MARKS), never sharing a frame for more than a second, across all 13 beats and all 3 hook variants. No continuous caption track anywhere — the donor's grammar is blocky phrase-level burns with numbers massive, not subtitles. Exactly one beat (11) carries no OST at all, deliberately.
• Cutting grammar constant: hard cuts exclusively, 1 to 2s, across 3:53. No dissolves anywhere. The board contains exactly four non-cut camera moves and all four are declared: the beat-1A conference-room push, the Hook-C rack focus, the beat-9 push into the empty space, and the beat-12 product push. The one tempo change is beat 12 and it snaps back immediately (VS4 satisfied).
• Hook seams: all three variants are 12s, all three end on the identical seam state (a paper macro, institutional beige, caption field cleared, camera already drifting), and the body does not shift by a frame between variants. Any hook cuts straight onto beat 2's first line. Zero variant-conditional body elements.

• First 5s: PASS on all three variants via Fork 4 strategy 3 (hook visual asset before any speaker) — which on this board is not a choice but a necessity, because there is no speaker to cut to at all. A no-presenter documentary is either mute-legible or it is nothing.
• A: an emptied 2003 institutional room with "2003" massive at t=0.3s and the category noun THYROID on screen at t=0.9s, handing at t=2.1s to "2.5 → 4.5" sitting on a printed guideline page for a full three seconds. Strongest muted hook of the three.
• B: hands turning the pages of a clinical volume under a desk lamp, with the board's most viewer-implicating card up at t=0.4s ("THE RANGE THAT DECIDES YOUR THYROID IS NORMAL"), handing to "RECLASSIFIED" at 4.2s. This card is the F3 disposition and it is the reason B qualifies on mute even though its spoken line one qualifies the narrator.
• C: "13,000,000" full-frame from t=0.0s, racking onto its own printout wall by t=2.0s, then "TOLD YOUR LABS WERE NORMAL" at t=2.6s. Most arresting muted frame on the board; thinnest on qualification timing (t=2.6s vs 0.4s and 0.9s). Declared as the trade for its cold open, not a defect.
• Hook OST word counts (VS2, ≤12 words each): A "2003" (1w) / "THE THYROID RANGE MOVED" (4w) / "2.5 → 4.5" (1w) / "13,000,000" (1w) / "TWO YEARS. THEN PERMANENT." (4w) ✓ · B "THE RANGE THAT DECIDES YOUR THYROID IS NORMAL" (8w) / "RECLASSIFIED" (1w) / "13,000,000" (1w) / "TWO YEARS INSIDE IT" (4w) ✓ · C "13,000,000" (1w) / "TOLD YOUR LABS WERE NORMAL" (5w) / "TWO YEARS IN" (3w) ✓. Every card on the whole board is ≤12 words; the longest is Hook B's 8-word qualifier and the longest body card is beat 3's "AFTER TWO YEARS, GONE FOR GOOD" (6w), tied with beat 4's "SAME BODY. SAME JOB. SAME ALARM." (6w) and beat 9's "FOUR PRESCRIPTIONS. FOUR SPECIALISTS. FOUR COPAYS." (6w).
• VS5 (asset survives the full 5s): A's chain is continuous with two overlapping cards and a 3s slab dwell inside the window ✓. B's qualifier card runs t=0.4 to 3.6 and hands to a slab at 4.2 ✓. C's slab runs t=0.0 to 2.4 and hands to a card at 2.6 ✓.
• U12 scan of every viewer-visible on-screen string on this board: 0 em dashes, 0 en dashes. Full swept inventory, all dash-free by construction: "2003" · "THE THYROID RANGE MOVED" · "2.5 → 4.5" · "13,000,000" · "TWO YEARS. THEN PERMANENT." · "THE RANGE THAT DECIDES YOUR THYROID IS NORMAL" · "RECLASSIFIED" · "TWO YEARS INSIDE IT" · "TOLD YOUR LABS WERE NORMAL" · "NORMAL" · "TWO YEARS IN" · "ONE NUMBER MOVED" · "3.6" · "3.8" · "3.4" · "3.9" · "FINE." · "Made of lead" · "2% A MONTH" · "AFTER TWO YEARS, GONE FOR GOOD" · "SAME BODY. SAME JOB. SAME ALARM." · "THE CONTROL GROUP" · "THE ORDER" · "THE PACKAGE" · "2014" · "2016" · "2018" · "2021" · "I don't recognize who I've become" · "NOT REIMBURSED" · "REJECTED" · "12 MINUTES" · "2.5" · "4.5" · "THE PAPERWORK DID" · "FOUR PRESCRIPTIONS. FOUR SPECIALISTS. FOUR COPAYS." · "2.8" · "YOU ARE ALREADY IN IT" · "GO LOOK" · "SELENIUM" · "ZINC" · "IODINE" · "THE RATIO" · "Made it to eight without the couch" · "Wore short sleeves in March" · "Recognized the woman in the mirror" · "30-DAY MONEY-BACK GUARANTEE" — plus the four one-word specialty labels at beat 9. Remaining em/en dashes: 0. ("30-DAY" and "MONEY-BACK" are compound hyphens and stay; "→" is an arrow glyph, not a dash; "%" is a glyph.)
• On-screen numeric set audited against the no-invented-numbers rule: every number on the board is spoken by the narrator. No price, no discount, no stock count, no review count, no p-value, no sample size, no percentage other than the spoken 2%, appears in any frame.

production_limitations.md is a sparse skeleton (§2a to §2d all "empty — populates via update sources in §4", well under the 3-entry threshold) → formal YS6 cross-check silent-skipped per the file's own heuristic, as recorded in the Research Brief. Tool-knowledge flags applied from avatar_tool_prompt_syntax.md and the runtime routing instead:

1. Fork 1 rationale — the avatar stack is DELETED, not chosen against. D10 forbids a continuous on-camera presenter, so there is no Heygen call, no Soul Cinema render, no /higgsfield-soul-id training pass, no Marketing Studio presenter, no talent day and no wardrobe continuity problem on this board. This is the lightest talent load and the heaviest asset load in the Echo chain. The trade is explicit and it is the right one for a first D10: a board with no performer cannot fail on performance, and it can only fail on picture.
2. The one casting decision is a VOICE. Speaker 3.7 Investigative / Forensic Narrator, female re-cast, UNTESTED — and the archetype × delivery-mode combination is also untested (this is the account's first D10). Direction is fully specified in every beat's VO block: forensic, absolute facts, flat-fast at 221.8 wpm, no warmth asides, no peer texture, no company-"we", uncontracted throughout (F1), one fear spike at beat 3, one register turn at beat 10, resolve rather than warmth at the close. Cast for the read, not the demographic — the register is the true-crime documentary narrator with the male voice swapped out and nothing else changed.
3. 221.8 wpm cannot be delivered at a house pace and this is a casting gate, not a direction note. It is the winners' p90 and the donor's measured rate. A narrator who reads this at 190 will run the film to 4:20 and past the 251s ceiling. Record a pace test before booking.
4. Hands and mouth are the three highest AI-artifact actions on the board, and there are exactly three of them: the notebook writing (beats 3, 6, 13 — five separate handwriting inserts), the page turn (Hook B), and the sublingual dropper (beat 12). Every one is a hands-or-mouth action generated models handle worst. Recommendation: a half-day hands-only practical pickup. It needs a table, a lamp, a notebook, a pen, a book, the real bottle, one pair of hands and no face — the cheapest shoot in the batch, and it de-risks the board's two load-bearing payoffs (the "Made of lead" entry and the three closing entries).
5. The canonical Carol keyframe gates roughly fifteen downstream clips and is the board's one hard sequencing dependency. Generate it FIRST, QC it, LOCK it, and Nano-Banana-chain every later Carol asset from it. Character drift in a documentary is worse than in a testimonial because the viewer is being asked to believe these are one person's real objects.
6. All on-screen text in POST — nothing rendered in-generation, and specifically: every NUMBER SLAB, every DOCUMENT MARK, all handwriting overlays, the ink circles and underlines, the drawn-curve marks, all four specialty labels, the guarantee badge, and every pixel of the beat-10 patient-portal UI (no video model renders legible interface text reliably, and that beat cannot afford a legibility miss).
7. Legibility gates (seven of them): the beat-1A guideline range line and requisition column · the beat-1B tube label · the beat-1C results-column word "NORMAL" · beat 3's four requisition readings · beat 7's EOB status field and unticked order box · beat 10's portal results row · beat 12's Kindled label with its three actives (the hardest, and the one where a composited fallback is barred). Build documents at 2× and crop tight; never film a full page.
8. INSTITUTION LAW is a QC gate on every single frame. No real organization's name, mark, letterhead, logo, masthead, payer name, portal brand, drug name, pharmacy mark or page appears anywhere on this board. Every document is generic and unbranded. This is what lets a documentary about an institutional decision be produced without forging an institutional record, and it is also the F4 disposition.
9. No invented document, no invented study, no invented citation, no charted curve. The beat-8 bell curve is drawn by a hand on paper with two marks and no axis. A motion-graphic chart there would convert the board's most honest beat into its least.
10. The period look must survive QC. Beat 1A's 2003 room and beat 1B's volume carry the entire dating of the angle. Generate several and reject any with anachronism (flat-panel monitors, modern chair geometry, contemporary typography on the documents). Anchor B is the evidence that generated period imagery holds at this quality bar.
11. The beat-10 phone must never be shot as UGC. Locked, slightly overhead, object-as-evidence. A hand-held phone frame in a D10 film imports an entire competing grammar and loses the editorial-television register that is the archetype's defining constraint.
12. Pause and runtime discipline: exactly one bracketed token exists (beat 12) and it is boarded as a single tight ≤0.8s hold. Delivered 233.8s boarded / 235.3s word-derived against the 251s ceiling → headroom 17.2s / 15.7s, identical on A, B and C. No second pause may be added, the hook must not be padded past 12s, and the four under-dense beats (5, 8, 9, 11) spend their surplus on picture and bed, never on silence.
13. Staging constraint, beat 4: no selenium, iodine or thyroid supplement of any kind may appear on the medicine shelf. Their presence would fight the beat-12 reveal. QC gate, not a preference.
14. No frame anywhere shows a lab number improving (brand_bible §9.2). Every lab number on the board is an in-range historical reading or an unchanged range bound. The close's payoff is the handwriting column and that is load-bearing copy (review C2).

Asset totals: roughly 50 distinct visual assets — about 32 generated (period/institutional, Carol-world, system-world, clinical-white graphic), 11 photographic or practical (of which 7 are the real-shoot-preferred hands and mouth inserts), 4 product/label macros, and 3 re-uses re-framed and re-timed — cut into roughly 150 cuts across 233 seconds at the donor's 1 to 2s rate. Zero presenter performance, zero avatar renders, zero talent days for picture. One VO session.

Per-beat prompts collated above in higgsfield-generate plus real-shoot syntax (avatar_tool_prompt_syntax.md hybrid and real-shoot conventions, extended with the WORLD / GRADE / SHOT_N sequence fields this format requires and the v2 per-beat schema does not model). Layer stack: higgsfield-generate (GPT Image 2 keyframes → Seedance 2.0 image-to-video for all period, institutional, Carol-world, system-world and clinical-white footage; Nano Banana reference-chaining for the Carol continuity spine and for beat 4's relight chain) · hands-only practical pickup (half a day: five handwriting inserts, one page turn, one sublingual dropper, the paper substitution, the drawn curve) · practical product macro (bottle, dropper, label) · post (every OST, every ink mark, all handwriting overlays, the four specialty labels, the beat-10 portal UI, the guarantee badge) · one VO session (female investigative narrator, 3.7 re-cast). No avatar tool appears anywhere in this stack — no Heygen, no Soul Cinema, no Soul ID, no Marketing Studio presenter — because D10 has no presenter to render. Copy-paste ready as a generation queue plus a one-table practical shot list.

• VS1 (B-roll density): roughly 150 cuts across 233s ≈ one new visual every 1.6s, which is an order of magnitude ABOVE the 1-per-10s ceiling. CONFIRMED-INTENTIONAL and donor-mandated: the donor's measured grammar is "very fast, 1 to 2s cuts, no pauses," and in a no-presenter documentary the imagery is not B-roll interrupting a talking head — it is the entire film. The VS1 band does not model a format with no A-roll.
• VS6 (cross-adaptations): 4 logged. ① Anchor B — its ANCIENT framework, 373.7 wpm text-scroller pacing and wordless construction rejected wholesale; borrowed at archival-look and typography-weight level only. ② Anchor C — its entire on-camera credentialed-persona staging rejected (D10 forbids it, product_facts §5 forbids it again); borrowed at H04 formula level and re-staged as hands-and-volume, which review C4 cleared explicitly. ③ Anchor D — its mortality imagery rejected on standing doctrine (never invent mortality; our terminal is file-sourced and non-mortal), and per review F2 our Hook C's copy routes to H14 rather than the cited H07, so the visual treatment is built for the H14 reading and the mismatch is recorded rather than papered over. ④ Anchor E — the faceless-VO carrier imported, the resilia persona lane and its counted scarcity explicitly not imported (product_facts §5).
• VS3 / VS3b: every beat cites Anchor A at its matching structural turn, and every insert carries its own anchor or an inline rationale. Four declared original-direction calls: beat 4's relight-not-regenerate technique (the persuasion goal is literally sameness, which is a frame-match problem first), beat 9's empty fifth space (the copy's argument is that nobody is looking, and an absence is the only honest picture of nobody looking), beat 11's card-free treatment (the board buys a clean field for the product beat's first card), and beat 13's handwriting-not-typeset ruling (review C2 named the entries-versus-number distinction load-bearing, and typesetting it would have simplified it visually while leaving the words alone).
• Hook archetype spread (review F2, carried not fixed): with Hook C reading H14 by the catalog's own tie-break, the board's real spread is H14 / H04 / H14 — two archetypes across three hooks, which still clears the ≥2 gate but means A and C will fatigue as one door. The board gives them three genuinely different visual strategies precisely because the copy-level differentiation is thin. The human should know this before the test reads.
• Review F5 (product-entry depth): boarded and stated on both bases — beat 12 at 3:10 = 79.9% of boarded runtime, 78.9% word-depth like-for-like against the donor's 80.5%, 82.0% on the brand-name basis. No compensating visual delay added, per the review's explicit ruling against structural change here.
• Donor honesty flag (carried from the Research Brief, operator-level): this donor's ws is 52.24, below the 60 winner bar, and it executes EUREKA's measured non-winner structural signature (13 beats with a full disqualification arc where EUREKA winners run 8 to 9 beats with none). It is an exemplar of shape and delivery, not a proven winner. The board reproduces its D10 grammar and its 221.8 wpm with confidence and records the structural caveat rather than acting on it — but if this concept under-performs, the 13-beat structure is the first hypothesis, not the visual execution.
• Format-novelty flag (operator-level, not a rule violation): this is the account's first D10 cinematic documentary, on the delivery catalog's #1 carrier (38.2% winner share, +11.5pp) which the account has never once run. The EVIDENCE CONTRACT above is written to be executable with no in-house precedent, but the first cut should be QC'd specifically for the three failure modes this format has: (a) a NUMBER SLAB that ended up more than one cut from its document and now reads as an invented statistic, (b) any frame that has drifted social-native (the beat-10 phone is the exposure), and (c) a held frame anywhere outside beat 12.

Concept ID: V3 · CID: KD_C0086 (pre-allocated) Production Target: AI-GENERATION-PRIMARY hybrid, NO on-camera presenter and NO talent day — higgsfield-generate (GPT Image 2 -> Seedance 2.0, Nano Banana chaining) + a half-day hands-only practical pickup + practical product macro + all OST in post + ONE VO session. No Heygen, no Soul Cinema, no Soul ID, no Marketing Studio presenter Visual Style: donor-derived institutional-forensic documentary (five declared worlds, two OST classes, numbers massive, 1 to 2s cuts, one slow-down at the packaging beat) Comp Library Anchors: 27357369823856563 (DONOR/SEED CORNERSTONE), 1269484834872935, 1336103951757183, 1957809038129580, 26907486858952165 (no-presenter geometry ref) Runtime: A/B/C all 3:53 boarded (233s); word-derived 234.5s; delivered 233.8s boarded / 235.3s word-derived; band 185-251s, headroom 17.2s / 15.7s, identical on all three variants Mute-test: PASS (all 3 hook variants) Production-limitation flags: 14 (avatar stack deleted, VO-only casting, 221.8 wpm pace gate, 3 hands/mouth artifact actions, Carol keyframe sequencing dependency, all-text-in-post, 7 legibility gates, INSTITUTION LAW, no-invented-document rule, period-look QC, anti-UGC phone rule, pause/runtime discipline, beat-4 shelf staging, no-improving-lab-number rule) Production weight: ~50 distinct assets, ~150 cuts, 15 boarded units, ZERO performer — heaviest asset load and lightest talent load in the chain

**Adapted-convention / compat skips (flagged, not silently dropped):** Beat 1A/1B/1C hook variants · cumulative INTEGER M:SS contiguous timecodes taken from the script's own beat markers on a shared 12s hook baseline (all three variants identical, body does not shift) · a **density table** carrying word-derived seconds alongside the boarded markers per the review's check-10 recommendation · CID KD_C0086 stamped at storyboard time (PRE-ALLOCATED per the E3/R3/S-series convention — no mint-at-ship collision) · donor PINNED as anchor cornerstone with NO fresh comp query (run-order directive; anchor set inherited from the script's cited lineage per R3 finding 3, plus the S2 finding-4 same-mode geometry slot) · thirteen-beat discipline with donor order and shares preserved · **a new EVIDENCE CONTRACT section** (two OST classes, the one-cut sourcing rule, the provenance ladder, INSTITUTION LAW) which no v2 schema field models · per-beat `World / grade` and `Shot sequence` fields replacing the presenter-assuming `Shot type` / `Camera` pair · SKIPPED per echo convention: rotation-log append (§2g), concept_pipeline append (7c), library_changes style-family log, Notion/board/publish (§7d) — "echo run — handled at package" · save path redirected to the Echo session tree.

1. The per-beat schema assumes a performer exists, and D10 is the case that proves it does not have to. v2's per-beat "Shot type / Camera / Avatar-tool prompt" triple is written for a frame containing a speaker: shot type describes how the speaker is framed, camera describes eyeline and blocking, and the avatar-tool prompt renders the speaker. On a no-presenter documentary all three are pointed at nothing. This board replaced them with World / grade plus an explicit Shot sequence sub-list with internal timings, and re-pointed the tool prompt at an asset-generation queue. S1's finding 1 predicted the shot-sequence need for long-form; D10 forces it for any runtime, because a 9-second beat with no performer still contains three shots. Recommend promoting World / grade + Shot sequence into the per-beat schema as the presenter-free variant before the next D10 or D04 board.
2. A documentary that cites a real, dated fact needs an EVIDENCE CONTRACT, and it should be a named, reusable pattern. The hazard is specific and it is not a compliance hazard: a blocky number on a coloured field is visually indistinguishable from a fabricated statistic, and an ad whose entire angle rests on a real institutional decision cannot afford to look fabricated for one cut. The two-class OST split (NUMBER SLAB / DOCUMENT MARK), the one-cut sourcing rule, the provenance ladder (decision → consequence → scale), and INSTITUTION LAW (no real organization's name or mark anywhere, so the film can indict a system without forging a record) together solved it, and they are format-portable to any donor whose grammar renders numbers massive. Recommend pinning the contract as the named documentary-evidence pattern. It also generalizes S1's finding 5 (the two-class OST grammar) by giving the second class a behaviour — marks behave like ink on paper, not like typography over it — which is what makes the distinction survive an editor.
3. "Generate long and wide, cut short and tight" is the reconciliation a fast-cut generated board needs, and it should be stated up front. At 1 to 2s cuts a 233-second film needs ~150 cuts, and a board that lists ~50 assets will read to a producer as either under-specified or impossible. Naming the ratio explicitly — generate 5 to 8s at max resolution with edge headroom, pull 2 to 4 cuts per generation by re-framing and re-timing — converts an apparent contradiction into a production instruction. Every board in a fast-cut delivery mode should carry it in its global spine.
4. Deleting the avatar stack is a legitimate Fork-1 answer and the chain now has all four. E1/E2/E3/E5 routed hybrid AI stacks; R3 and S2 routed pure real shoot; S1 routed hybrid with a real-shoot-preferred presenter and a priced degraded fallback; V3 routes hybrid with NO presenter layer at all. That is not "AI presenter" or "real presenter" — it is a fourth answer in which the presenter question does not arise, and it inverts the load: zero talent days, zero avatar renders, zero performance risk, against the heaviest asset count in the chain. Recommend Fork 1 gain an explicit "no presenter (delivery mode forbids one)" option so a future D10 or D04 board does not have to argue its way there from a menu that assumes a speaker.
5. comp_select.parse_anchors truncates a hyphenated advertiser name, and this is the first board in the chain to carry one. The v2-compat twin parses all five anchors correctly (v1's bolded headers parse zero, which is exactly why the twin exists), but the block regex captures the anchor's name field as [^—\-\n]+ — it stops at the first hyphen. Anchor B's try-forge therefore lands downstream as try. The ad_id is unaffected (it resolves exactly, and _resolve_ad_id is a no-op at this id length), so comp selection, vision enrichment and the intel links are all correct; the damage is display-only, and it will show as "try" in the editor brief's §10 comp list because Anchor B sits inside the default cap=3. Every prior board in the chain drew all-lymphoria / sereneherbs / trysculptique sets and never hit it. Recommend widening the name capture to allow internal hyphens ([^—\n]+? with a trailing-space trim, or an explicit [a-z0-9][a-z0-9_-]* brand class) before the next board draws a hyphenated advertiser. Recorded here rather than worked around in the twin, because the twin's contract is "ONLY the comp-anchor grammar is adapted" and inventing a hyphen-free spelling of a real advertiser would be a content change wearing a formatting disguise.
6. A first-of-mode board should reserve an anchor slot for a same-mode reference even when the mode's #1 evidence is the donor itself. S2 established this for geometry; V3 confirms it under a harder condition — the donor here IS a D10, so the mode was nominally covered, and Anchor E (a D04 no-presenter VO) still earned its slot by answering a question the donor could not: how a presenter-free film sustains attention over its explanatory beats. Its "deliberate slow-down on the explanatory beat" instruction is what told this board how to spend beat 5's 6.3 surplus seconds. Recommend widening the S2 rule: reserve the slot for a same-family reference (presenter-free, animation-led, composite, and so on), not only the same delivery-mode code.