Video production and narration

Guteneo’s public films are editable Remotion compositions backed by local media, not screen recordings of an automatically available delivery service. Replaying an approved film uses stored narration and makes no speech-provider call. Creating new speech, approving a new film and publishing it are separate operations.

Choose the correct production

The active catalog has twelve films: horizontal introduction, vertical introduction and role overview in each of fr, en, de, lb. Its authority is packages/contracts/src/public-videos.json, with actual durations and matching posters/captions. The server’s apps/api/src/public-video-manifest.json records served MP4 byte counts and hashes. Neither a historical filename nor a generated provider result selects a film for the website.

Install and inspect locally

From the repository root, install the locked application and film dependencies. The film package pins Remotion 4.0.526. Keep all Remotion packages at their locked, matching versions; an upgrade needs its own qualification.

npm ci
npm --prefix videos/guteneo-film ci
npm --prefix videos/guteneo-film run lint
npm --prefix videos/guteneo-film run dev

Open the exact local address printed by Remotion Studio. Select a composition in videos/guteneo-film/src/Root.tsx. The Natural-C folder contains dynamic narration-timed introductions and the English/German/Luxembourgish role versions; French roles retain their approved natural timeline. Frame-driven React animation lives in src/scenes/, src/landscape/, src/vertical/ and src/roles/.

FFmpeg, ffprobe and cwebp must also be available for local regeneration. Fonts, images and approved audio are local. The renderer has no customer account, business binding or communication secret. An ordinary Studio preview is a local review, not a publication or provider qualification.

Change a story without changing the product’s promises

Start from complete sentences and the actual workflow. The four roles and independent supervisor permissions come from packages/contracts/src/roles.ts. Preparation, human review, approval, provider acceptance and delivery must retain their distinct meanings. A film does not grant a mandate, make a channel available or certify a marketplace integration.

Use only synthetic product captures. Keep their demonstration disclosures and provenance; do not retouch the DOM to invent a feature. Reconstructed assistant screens are labelled advertising reconstructions, not real chat or published widget evidence. Printed-paper, envelope and fax scenes illustrate uses; they are not delivery receipts.

Review factual claims and the complete script before synthesis. Technical instructions and code stay English; narration, captions and on-screen wording may use their declared language. Localize all three media companions together: film, poster and captions. The player resolves UI/account language and must never silently substitute another language’s voice.

Produce new narration only when authorized

Existing approved speech should be reused. For a new authorized narration batch, follow the actual method C: three contextual blocks rather than synthesizing isolated visual headlines. Introductions group eleven paragraphs 4/4/3; role films group six paragraphs 2/2/2. Horizontal and iPhone introductions share one narration per language.

1. Run the complete script through native ElevenLabs Enhance and accept Keep. Preserve the exact native result in enhancement.json. The connected Creative Studio path did not supply native Enhance; the recorded workflow used authorized Safari account UI.
2. Review the wording and permissions. If needed, record calm-tone edits in editorialEdits; [thoughtful] and [short pause] describe the retained direction. Keep the native enhancement separate from the final edited script.
3. Inspect the live model schema, choose the actual Manon voice and set model_parameters.language_code explicitly to fr, en, de or lb. Use eleven_v4, not an assumed similarly named model. The display name is a selection aid; the approved library retains the actual provider identity.
4. Generate two variations per block through the authorized Creative Studio connection and retain the second by its real session/generation identity. Returned-media order is not selection evidence. If an outcome is uncertain, inspect that existing flow/session instead of generating again.
5. Archive original selected MP3s, exact scripts, filtered receipts, durations, byte counts and hashes. Keep credentials, signed URLs and private project identifiers out of public documentation. Estimates do not prove billed usage.

This is an account-bound, potentially paid phase; the offline render commands below do not perform it. The old local narration:generate adapter remains blocked before network with API_ADAPTER_UNQUALIFIED. Plugin access does not qualify that adapter, and narration/voices.example.json is a historical configuration example, not the authority for the currently published voice.

Fit scenes to the voice, never accelerate the voice

Decode complete MP3s to mono PCM24 at 44.1 kHz. Establish boundaries from word timestamps and acoustic pauses without prompting a recognizer with the expected script. Preserve real transcription differences. In this batch, French, English and German used Whisper word timing; Luxembourgish used native Scribe JSON after local recognition proved inadequate. Recognition does not certify pronunciation or create missing words/timestamps.

Retain every source sample in order at tempoFactor:1. Lengthen scenes to actual speech, with readable minimum holds and interblock pauses. Do not trim a last word, time-stretch a voice or force a localized narration into the old 56-second instrumental timeline.

All films retain the approved five-second silent logo, exactly 150 frames at 30 fps; introductions also retain two silent Signature seconds. Music remains at a constant 0.22 gain during speech, pauses and the logo, with the original score’s cadence/end fade. Voice normalization and peak checks are separate. The old George mix’s between-sentence music ducking is historical, not this policy.

Each of the eight immutable directories narration/releases/{introduction,roles}-{fr,en,de,lb}-natural-c-v1/ retains scripts, enhancement, three MP3 sources, canonical WAV cuts, timing/alignment, receipts and hashes. Together they hold 24 selected MP3s and 68 PCM24 cuts. The builder rejects overwriting an existing snapshot.

MP3, canonical PCM and cut hashes remain strict. Decoder qualification permits only a documented difference of at most eight PCM24 least-significant bits with an unchanged sample count. Scalar FFmpeg decoding uses -cpuflags 0 to avoid CPU-specific optimized paths; this bounded decoder allowance does not permit changing canonical WAVs or published films.

Rebuild approved films offline

Run these commands from the repository root:

npm run test:videos:narration
npm run videos:render

The renderer qualifies narration libraries before public writes, renders musical bases, remixes stored speech while preserving base H.264 packets, then writes posters, VTT captions and the server manifest. It requires no ElevenLabs call. It writes generated public media: review the resulting diff and hashes before including them in a release.

For a bounded subset, pass the existing renderer’s explicit selectors:

npm run videos:render -- --locales de,lb --kind introduction
npm run videos:render -- --locales en --kind roles --skip-existing

--skip-existing still checks the manifest hash and matching poster/captions. It is not permission to reuse an arbitrary file. Normal output masters and render-proof.json are under videos/guteneo-film/out/localized/; public movies and posters are under apps/web/public/videos/.

A new library has a separate initialization path. scripts/build-natural-library.mjs --input <prepared-directory> validates and freezes a prepared library; it does not synthesize audio. scripts/render-natural-sources.mjs --libraries <snapshot-id> writes musical sources only under ignored out/. These are controlled authoring steps, not shortcuts for replacing an approved snapshot or updating the catalog implicitly.

Preserve the approved visual ending

The visual base is the approved V5 composition even when natural narration extends the film. Landscape and iPhone scenes are individually recomposed, not cropped versions of each other. Preserve the simple unpostmarked opening mark and the large final official V5 halftone stamp, postmark and guteneo.com. Do not regenerate or redraw that ending to match a new illustration style.

Use paper #f6f5ef, ink #181b22, cobalt #2450db, local EB Garamond and IBM Plex Sans/Mono. Keep guteneo lettering lowercase and ink black under the existing video composition rules. The stamp already contains the signature; add no duplicate domain. The engineering header’s separate dark-mode exception is not a video-brand change.

Image/font/audio provenance lives beside the assets: public/images/PROVENANCE.md, public/frontend/provenance.md, public/fonts/ licenses and public/audio/README.md in the film project. The original soundtrack is procedural local synthesis, not a third-party music extract. Preserve licenses and generated-asset disclosures; a provider receipt does not by itself establish rights for an unrelated asset.

Verify, then publish through the release process

node --test tests/security/public-video-assets.test.mjs
ffprobe -v error -show_streams -show_format -of json \
  apps/web/public/videos/guteneo-horizontal-v8-en.mp4
ffmpeg -v error -i apps/web/public/videos/guteneo-horizontal-v8-en.mp4 \
  -f null -

Check all changed outputs, not just this example: H.264/AAC stereo, 30 fps, fast-start, intended dimensions, complete decoding, bounded file size below 25 MiB, exact catalog duration, synchronized captions, suitable posters and the unchanged final logo hold. Inspect desktop and mobile framing, active-language media, keyboard controls, sound and reduced-motion behavior.

Listen to every changed language continuously. Automated ASR, peak measurements or successful decoding do not prove a natural accent, complete words or human approval. Only the French role film carries the recorded user-listening approval; other libraries retain pending critical-listening evidence. A phone viewport is not a physical-device test.

Local rendering does not publish files. Follow changes and releases for the exact merged commit, CI, clean main checkout and authorized deployment. Then verify the actual public assets, hashes, byte-range behavior, captions and player on the served release. No film build sends a fax, email or postal letter.

Keep the Gutenberg production separate

The cinematic Gutenberg archive preserves the approved proposal-2 character, François-Louis narration and Manon signature. Do not replace its voice with the public-film default or claim that its revised seven environments have already been rendered. Approved voice files need no resynthesis or speed correction. Its current handoff distinguishes existing short previews from unfinished full exports.

The archive can be recovered and checked locally without a provider account:

python3 videos/guteneo-gutenberg/tools/restore_large_media.py
python3 videos/guteneo-gutenberg/tools/verify_archive.py

Restoration reconstructs exact bytes from tracked chunks without transcoding. Historical localized prompts and voices stay immutable. Read its TECHNICAL_HANDOFF.md, STORYBOARD_AND_PROMPTS.md and DECISIONS_AND_STATUS.md before resuming work; source availability, approved speech, animation quality, finished export and publication remain separate evidence.