Video production: plan and research
The phase after the script and the voice-over, turning them into a video, has been planned and researched, not built. The plan lives in research/video-production/; start at its README.md. The video is arranged to the voice: about 220 scenes for 1,217 seconds of Hinglish audio (the script has 14 graphic notes), each one a data scene from the fact ledger, a piece of typography, a flat diagram of the script's own metaphors, or a cropped passage of a real source, with a thin layer of licensed photographs. Claude writes a structured storyboard; tools check it, fetch and build the assets, compile a frame-exact timeline, render, mix and verify.
- Proven here: a Node renderer ran a heavy 1080p scene at 22 frames per second in 186 MB with no frame files (this machine has 6 GB RAM and 7 GB free disk); ffmpeg has every filter needed; Wikimedia Commons, the Internet Archive and headless Chrome work without keys.
- Not proven: every pacing, load, density and loudness default, whether viewers feel any of it, how YouTube treats a synthetic narrator, and anything about how the audio sounds (Claude has not heard it).
- Folklore found: "1.6 second shots", "15 to 20 seconds static kills retention" and "65% at 3 seconds" have no primary source; Google AI Mode said so itself. They are not used as targets.
- Waiting on the user: 14 decisions, each with a recommendation, listed below; the first slice proposed is beats B0 and B4 (183 seconds).
1. The four questions, answered
| Question | Answer | File |
|---|---|---|
| Which visuals, and how to get them | Five honesty classes (evidence, data, metaphor, typography, atmosphere); a closed catalogue of 26 scene types; a status grammar that draws verified, alleged and "not found" differently; rights and routes for each source; a 32-idea bank | 01-visual-assets.md |
| How to arrange them with the voice | A paragraph contract (anchor, evidence, figure, landing); scene per sentence; cuts on sentence ends and pauses; visual load opposite to verbal load; transition meanings and defaults; text and number rules | 02-visual-grammar.md |
| Sound | Every sound signals a visual event, mirrors what the audience remembers, or is silence; functional sounds synthesised in code; CC0 foley for the recognition moments; four music beds the user auditions; two-pass loudness | 03-sound.md |
| What else matters | YouTube's policy on templated and synthetic content; Indian fair dealing after ANI v RSY News (29 September 2026); honesty on screen; privacy and allegations; this machine's limits; how to review without hearing; measurement; delivery | 04-other-areas.md |
2. The pipeline, V0 to V9
Lock
Hash the script, ledger, 14 WAVs and word times. Every beat must be aligned (only B1 and B2 are).
Meaning map
Per sentence: numbers with ledger status, dates, quotes, names, questions, pauses, hard or easy window.
Storyboard (Claude)
One scene object per sentence or clause, in beat-sheet.json. The creative step.
Beat-sheet lint
37 rules: honesty, traceability, sync, pacing, load, text, sound. Hard gates and warnings.
Assets
Commons, Internet Archive, headless capture, pdf.js, Freesound. A licence and source row for each.
Timeline
Integer frames at 30 fps, SFX events, holds, landings, a complexity score per second.
Render
Per beat, cached by content hash, Node canvas piped into ffmpeg.
Sound
Synthesised effects, foley, bed, ducking from word times, loudness normalisation.
Check
ffmpeg detectors, text and contrast, flashes, a contact sheet for Claude to read, a verdict.
Publish and learn
Captions, credits, an OpenTimelineIO export for DaVinci Resolve, retention joined to scenes.
Contracts, rules and the twelve tools to build, in order, are in research/video-production/BUILD-SPEC.md. Nothing in it exists yet.
3. Facts measured on the way
- The Hinglish v7 script has 58 paragraphs, 210 sentences, 3,519 words, 69 number mentions, 2 direct quotes and 14 graphic notes; the audio is 1,217.0 s at 173 words per minute (beats run from 162 to 186).
- Free photographs exist for the CD-R, the 5.25-inch floppy, the 1982 Philips and Anefo CD launch, the Swiss clock maker Moser-Baer AG and the Padma Shri insignia. None exists for the Moser Baer India plant, the founder, the workers or the tribunal.
- 11 of the 28 URLs in the script's fact ledger have a Wayback Machine snapshot, so source pages have to be captured live and the capture recorded.
- Remotion is free for an individual or a team of up to 3 (official FAQ), but a Chromium-based engine is risky on 6 GB; the plan uses Node with
@napi-rs/canvasinstead.
4. Decisions waiting for the user
| # | Decision | Recommendation |
|---|---|---|
| 1 | Adopt the honesty classes and the never-list (no news video or article photographs, no logo, no photoreal AI imagery, no staged evidence, no faces) | Yes |
| 2 | Renderer: Node plus @napi-rs/canvas in a separate package | Yes |
| 3 | Renders and caches outside OneDrive | Yes |
| 4 | Captions | A closed-caption file from the script and word times; key-term typography on screen |
| 5 | Music | Library bed per act that the user auditions, plus synthesised drones; AI music off |
| 6 | Stock footage | Off |
| 7 | Pace (173 wpm) | Keep for slice 1 |
| 8 | Gap between beats | 0.8 s |
| 9 | Mahesh Sharma on screen | Text only; blocked until the 2019 article is opened |
| 10 | A private benchmark of 6 to 10 reference videos | User decides (YouTube's terms restrict downloading) |
| 11 | Unlisted test upload with the synthetic voice | User, 10 minutes |
| 12 | A free Freesound token in .env | Yes (never in chat) |
| 13 | First slice | B0 plus B4, then B7 plus B11 |
| 14 | Trial the deadline clock, pivot ladder, status grammar and pivot motif | Yes; drop what is disliked |
Fifteen threads were answered. It gave useful structure and vocabulary (sound as a signalling cue, a hierarchy of cut, reframe and in-scene change, a measurable proxy for each attention principle, the QA filter list), but contradicted itself on the pacing numbers between two questions and, asked to plan the opening two minutes, invented footage that matches no claim. Digests and verdicts are in research/video-production/sources/RESEARCH-LOG.md; they are digests, not verbatim threads.
5. Files
research/video-production/README.md: front door, decisions, what is proven.01-visual-assets.md,02-visual-grammar.md,03-sound.md,04-other-areas.md: the four areas.05-moser-baer-plan.md: the measured beat table and the beat-by-beat plan for Moser Baer v7 Hinglish.BUILD-SPEC.md: stages, contracts, rules, tools, acceptance tests.sources/RESEARCH-LOG.mdandprobes/: every source, check and probe, with the scripts that produced the numbers.
Written 2026-10-31 from the files in research/video-production/ and the probes run that day. If a page and a file disagree, the file is right: fix the page.