Video DNA: original evidence before adaptation

Owner correction, 2026-09-10: the three subvideos inside targ's opening 0–9.7s

are an **example of the required depth everywhere**, not a special opening task.

Apply the same decomposition to the entire duration of targ, sisterbrazil and

svekr, and to every future original. This is the evidence layer of steps 2–3 in

`generation_contracts.md`, not a second production workflow.

What the harness must preserve

One frozen original → a complete original passport → an explicitly changed

product/region/cast adaptation passport. Share software and product facts;

never share another original's creative structure, music or character images.

Current work analyzes originals. It does not resume cancelled films or buy new

adapted songs, images or video. Existing execution controls remain authoritative.

Level	Unit and required detail	Evidence
0	Original, native frames, presentation timestamps, audio/video tail difference	Original SHA256, ffprobe, complete decoding
1	Every adjacent-frame visual change; whole image, scene area, caption area, spatial tiles	Compressed per-frame numerical trace; actual source frames remain retrievable
2	Every separate subvideo/shot, including brief inserts and returns to the same composition	Exclusive native frame interval, original clip, entry/middle/exit images
3	Transitions with start/end spans: hard cut, fade, white flash, wipe, paper aperture, zoom blur, morph	Pre/inside/post original frames and actual moving transition; distinguish effect from a new shot
4	Within-shot events: onset, action, reaction, resulting state; expression, gaze, hands, object contact, camera, text	Timed microevents, visible change, source frame/clip links, observation confidence
5	Audio and visual relationship: exact word occurrence, speaker/delivery, music/stem/onset/beat change, visual response	Original sound, raw ASR, listening correction, measured acoustic trace; separate uncertainty
6	Identity, relationships, problem progression, props, locations, style and story causality	Passports using original images and exact antecedent → action → consequence references
7	Adaptation obligations for a new product and region	Preserve/replace/forbid/uncertain decisions, supported product fact IDs, localized cast and problem-result target

Numerical detection is not semantic understanding. A decoded frame count proves

measurement coverage only. Contact-sheet sampling must report its actual rate;

a video model's claim that it watched every frame is not proof. Candidate cut

counts are not shot counts. A line, lyric, 20-second transport window or broad

story chapter must never silently stand in for the actual subvideo inventory.

Numbered execution contracts

Step	Input	Output	Validation and transition gate
1. Freeze	One catalog original	source ID/SHA, probe, complete decoded native PTS table	Source hash matches; every frame decodes; audio tail explicitly accounted for
2. Measure all frames	Frozen original	Per-frame scene/caption/tile deltas and candidate boundaries	Decoded frame count equals timestamp count; zero silently skipped frames; no semantic acceptance yet
3. Inspect the complete picture and sound	Full sequence in source-bound A/V packets, dense sheets, candidate pairs	All-shot inventory, transition spans, within-shot event ledger, actual audio observations	Every transport interval inspected, ambiguous/missed edits refined at native boundaries; merged mini-shots rejected
4. Decompose the sound	Original full mix and independent source separation	Actual vocal/drum/bass/other stems, loudness, envelope, beat/chroma/pitch estimates, raw word ASR	Original hash bound; measurement method/version retained; model timbre estimates cannot override loudness measurements
5. Join the timelines	Reviewed visual events, captions, lyrics, music events	Event-level crossmodal map and story causality	Distinguish word onset, caption change and shot change; audio inference cannot create an unseen shot; every event has an original time
6. Passport the original	Corrected event map plus images/audio	Full recurring cast, relationships, style/camera/edit grammar, props, problem/action/result, voice/music passports	Source role images exist; age/identity variance and source unsupported claims explicitly recorded; no fixed cast count
7. Prepare adaptation obligations	Original passports, target product/region	Source-to-new-product/locale/role mapping and required generation attachments	Ukrainian casting requires regional identity, speech, setting and plausible relationships; translated names alone fail; owner target appearance remains a separate required lock
8. Freeze reusable packets	Actual reviewed source intervals	Text/structure/plot, original role frames, original audio and original video plus hash manifest	Exact ranges, actual bytes, single source hash; unresolved semantic inputs remain unavailable for final generation acceptance
9. Audit and show	All above, raw critiques and corrections	Browsable timeline, passports, measurements, packet index, coverage/gap report	No hidden unresolved defect, invented perfect score or averaged approval; source-specific readiness reported separately

Canonical time and event contract

Shot intervals use `[startFrame, endFrame)` with **exclusive** end frame. `start`

and `end` are source presentation seconds, not transport-local seconds. Hard

cuts use one boundary; dissolves/other transitions carry their own overlapping

`start`/`end` range. Transport segments may split a shot: merge on original

continuity, while retaining observations and receipts from both packets.

Each shot records identity, framing/composition, camera movement, location,

characters and their state, subject/object action, gaze/addressee, caption

changes, audible words/delivery, music, incoming/outgoing transition, causal

antecedent and later consequence. `subevents[]` records meaningful changes

inside the shot. Stable holds are explicit observations, not invented actions.

For an event retain: ID, time/range, channel/type, observed before→after,

actual reference binding, observation method, confidence/uncertainty and

separate interpretation. A model's raw critique is kept unchanged; the director

adjudication identifies its defect, source counterevidence and correction.

Adaptation passport and generation attachment contract

An original passport explains what the source does. A new adaptation passport

explains why and how a new character/product/geo changes it. These are different

artifacts. New cast portraits require their actual own-original role frame plus

the localized adaptation role. New voice/music requires actual own-original

sound. New shots require own-original video/frame plus their causal beat and

the accepted same-branch character/problem-state images.

Every subsequent request retains the exact script, structure, plot, photo,

frame, voice, music and video evidence applicable to that modality, and records

the actual provider attachment fields. Packet names without attached bytes fail.

`tools/reference_contract.py` remains the production attachment validator;

analysis evidence does not bypass it or the problem-passport gate.

The problem passport freezes the owner's intended visible defect and supported

target appearance. Do not replace it with dryness to make acceptance easier.

Compare the same identity/age in before/application/intermediate/resolved

states, hold clean comparison frames, record deliberate fades and verify the

export. A source's extreme slimming, medical mechanism or apparent de-aging is

not an allowable goPure claim. Preserve its narrative purpose through an

explicit supported adaptation. Owner rejection reopens the gate, and known

failed candidates require versioned corrections while work stays authorized.

Evidence honesty

Measurements, detector estimates, listening/visual observations, interpretation,

and proposed adaptation choices must be visibly distinct. Preserve ASR errors

separately from corrected heard words. Meter, chord labels, pitch and estimated

beats carry algorithm caveats. Stem leakage means low RMS cannot establish

physical instrument absence. A regular grid cannot identify a music generator.

Attention mechanisms are hypotheses; the owner calls these champions backed

by millions in spend, but the harness does not measure campaign performance.