# Project-wide rule: replicate the winning reference

## Latest owner rules — no subtitles; timing tolerance,2026-09-10

**Never generate or add subtitles.** This applies to Omni outputs, repairs and later
editing across all project sessions. Do not generate dialogue captions, translations,
transcript/karaoke overlays or subtitle tracks. Any newly generated take with subtitles
fails retention even if their spelling is correct. Original references and historical
rejected takes remain evidence; their overlays do not authorize new subtitles.

If the scene needs it, actual duration may be **up to25percent shorter or longer** than
its authored timing: `[planned×0.75, planned×1.25]`, inclusive. A planned4s shot may be
3–5s. Preserve every intended word, action and meaningful reaction; timing tolerance
never excuses omissions, speed-up or lost payoff. Rebuild the edit clock from actual
accepted media. The narrative still has no fixed total-runtime target.

Bind `harness/owner-output-policy.json` as `output_policy` in new native plans and use
`scripts/production_policy.py` for actual timing/subtitle checks. These latest rules
supersede older caption requirements or exact-duration assumptions, including those
inside the preserved owner-approved story snapshot. Its characters/story remain approved.


Owner instruction, 2026-09-10. This rule applies to **every future session and every
generation in this project**, including revisions and repairs. It supersedes conflicting
instructions that favor invention without a matching source reference.

The task is to replicate the selected reference's winning DNA with our product,
geography and story inputs. The owner identifies these references as creatives backed
by millions of dollars in spend. Preserve that as owner-supplied provenance; no ad-account
data or causal performance measurement has been independently inspected here.

## Every generation receives its matching original fragment

Attach a concrete, relevant fragment of the **original selected reference** to each
generation context. A whole-video link, remembered analysis, generic style description,
the first fragment repeated everywhere, or a previously generated descendant is not a
substitute for the corresponding original. Localized passports and approved intermediate
assets are additional inputs, not replacements for the source fragment.

| Generated output | Required original material in the generation context | What to validate |
|---|---|---|
| Scenario and plot | Matching scene/excerpt with original dialogue, action, cause and consequence | Recognizable event/function and earned next action survive localization. |
| Structure | Ordered source-turn map plus the corresponding original ranges | Coverage and order; explicit product/geo reasons for departures. |
| Photo / identity / concern passport | Relevant actual source frame(s), alongside our product and identity inputs | Visual register, age, casting role, texture, concern ownership and permitted change. |
| Storyboard frame | Actual frame/excerpt from that source scene plus accepted local passports | Staging, scale, eyeline, light, depth, gesture and correct local identities. |
| Spoken delivery / voice | Corresponding actual original speech/audio excerpt and localized stressed text | Audible contour, intention, reaction, pace and every Ukrainian word occurrence. |
| Music, when the format includes music | Corresponding actual original musical section | Musical role, evolution, transitions and native-language performance. |
| Video / motion / repair | Corresponding actual original moving excerpt plus approved local board/passports | Scene action, blocking, acting, camera/edit grammar and actual synchronization. |

This narrative format has **no song or music stage**. Record that stage as not applicable;
do not generate music merely because this project-wide table includes it.

Bind each fragment to source path/hash, source start/end (or frame time), local fragment
path/hash, extraction recipe and source role. Record the exact provider field carrying
it and the exact submitted request hash. A remote URL must have an upload receipt for
those bytes. A preset ID is not original audio. A typed path or URL in prose does not
prove the provider received or decoded the media.

If a provider cannot accept the required modality, mark the stage **blocked by transport**
with the documented capability limit. Do not silently replace the fragment with prose,
send unsupported fields, or claim direct conditioning. Resolve a supported path or obtain
an explicit owner decision about that concrete limitation before dependent paid work.

## Numbered pipeline, one inspectable contract per step

**Latest speech ordering override, 2026-09-10:** the owner explicitly chose Omni's
native speech. Prepare and preflight Ukrainian text, then generate picture and voice
together in Omni Flash360p using the original AV excerpt. Actual occurrence-level
listening and mouth review follow generation and gate retention. Steps6–8 below are
one linked native AV production loop; external TTS and a previously accepted external
WAV are no longer prerequisites. Native audio becomes the final soundtrack after review.

Every production has a numbered pipeline. Each step declares its ID, purpose, exact
inputs and dependencies, matching original fragment(s), output, validation criteria,
actual validation result and evidence, and failure/repair behavior.

1. **Identify the originals.** Input: owner assets/URLs. Output: source hashes, probes,
   complete audiovisual access and provenance. Gate: correct original and complete coverage.
2. **Analyze picture and sound.** Input: original AV and locally timed fragments. Output:
   observed turn/visual/voice maps, interpretations and corrections kept separate. Gate:
   full source including ending; model timing/claim errors adjudicated.
3. **Compile the reusable DNA.** Input: observed source maps. Output: callable function,
   source adapter and unrelated transfer tests. Gate: evidence traceability and usable inputs.
4. **Adapt structure and scenario.** Input: source scenes, function, product facts and geo.
   Output: localized story, dialogue and explicit source mapping. Gate: recognizable
   source sequence, causal/emotional completeness and product-appropriate claims.
5. **Create and review passports.** Input: corresponding original frames, local identities,
   wardrobe, product and concern requirements. Output: actual reviewed visual references.
   Gate: identity, ownership, state and source-style fidelity in real pixels.
6. **Generate and accept speech.** Input: matching original audio, contextual Ukrainian
   stress ledger and vocal score. Output: retained takes, alignment and accepted master.
   Gate: actual occurrence-level listening, expressive source comparison, complete joins.
7. **Create and review storyboards.** Input: matched source frames/excerpts, accepted
   passports and actual speech clock. Output: unique scene boards. Gate: actual staging,
   action, wardrobe/concern and source grammar; no passport-as-scene shortcut.
8. **Compile and generate motion.** Input: v2 plan, accepted speech/boards, matching original
   moving fragments and supported transport. Output: Omni Flash 360p takes and receipts.
   Gate: reference attachment verified before submission; real output reviewed afterward.
9. **Assemble, validate and repair.** Input: retained approved assets. Output: complete
   film with sound and selective versioned repairs. Gate: full-film source comparison,
   narrative completeness, actual mouths/actions/states, technical and playback checks.
10. **Publish the complete evidence.** Input: accepted film and every step's records.
    Output: hosted HTTPS film/report, branch/PR and Chief report. Gate: playable complete
    media; all inputs, outputs, criteria, results, rejected attempts and unknowns inspectable.

These numbers describe dependency order; independent preparation can run concurrently.
No downstream acceptance is inferred from a passed upstream contract. Preserve historical
media and receipts, and invalidate only dependencies whose inputs changed.

## Enforcement and scope

`scripts/reference_generation.py` validates source attachments and numbered step records.
The shared Kie submission client invokes its generation gate before authentication/network
for new jobs. Historical task retrieval/resumption stays available without new generation.
External tools and assistant-authored text still require the same reference packet and
review record; the shared-client gate is not a claim of intercepting arbitrary clients.