# Ki33 — One Sheet / Talking-head Reels

Independent concept by Ki33. Not a client commission. No claims about views, promotion, earnings or endorsement. Publication on the existing Ki33 portfolio was explicitly authorised by the owner on 1 November 2026. No Freelancehunt account upload is included in this authorisation.

## Deliverables

- `exports/Ki33-One-Sheet-Reel-1080x1920.mp4`: 34.4-second master, 1080×1920, H.264/AAC, 30 fps. Includes a 5.4-second attribution card.
- `exports/Ki33-One-Sheet-Reel-Freelancehunt.mp4`: the same edit at 1080×1920, encoded in two passes below **10,000,000 bytes**.
- `exports/Ki33-One-Sheet-Before-After.mp4`: sequential comparison of the same recorded passage, original then edited. Original framing, colour and audio level are retained in “before”; no defects or noise were added.
- `captions/reel.en.ass`, `.srt`, `.vtt`: editable English captions. ASS contains colour emphasis and soft fades; SRT/VTT are plain-text versions.
- `project/timeline.json`: actual source in/out points, crops, B-roll placements and audio parameters used by the build.
- `tools/build.py`, `tools/comparison.py`, `tools/finalize.py`: actual rendering, packaging and verification tools.
- `portfolio-stage/`: an isolated complete copy of the existing Ki33 portfolio with the new `/talking-reel/` case and a new home card. The original portfolio checkout is retained; publication adds the new case and its home card to the existing site.
- `delivery/freelancehunt-RU-EN.txt`: copy for the category “Аудио и видео монтаж”; analogous simple edits start at $10. This is a proposed future service price, not payment for this concept.

## Source and permission

Original creator: **VPRO / The Mind of the Universe**.

1. **Erik Demaine, the genius that plays for a living**, interview dated 21 April 2016: <https://themindoftheuniverse.org/play?id=Erik_Demaine>.
2. **Erik Demaine – Extra footage**: <https://themindoftheuniverse.org/play?id=Erik_Demaine__br>.
3. Publisher's explanation: <https://themindoftheuniverse.org/about>.
4. Source and adaptation licence: **Creative Commons Attribution-ShareAlike 4.0 International**, <https://creativecommons.org/licenses/by-sa/4.0/>.

The source pages explicitly permit copying, changing, distributing and exhibiting the material, including commercial use, with attribution and the same licence. Saved copies of the source/permission pages are in `sources/`. Attribution appears visibly in the exports, in MP4 metadata and in the case page. The adaptation is shared under CC BY-SA 4.0. No endorsement by VPRO or Erik Demaine is implied. This licence does not make Ki33 the original filmmaker or speaker.

The publisher's automatic transcript contains errors. It was used as a reference, not treated as an infallible transcript. The 98 selected words were reviewed in text against local speech recognition and the source transcript. Recognition is an aid, not evidence of human listening.

## Preserved footage

Only the necessary excerpts were retrieved with FFmpeg and packet copying:

- `sources/interview-0000-0058.mp4`: original interview 00:00–00:58, stream copied, 1920×1080, 25 fps, original AAC. Its container duration is about 58.08 seconds because of packet boundaries.
- `sources/broll-0000-0130.mp4`: extra footage 00:00–01:30, original video stream copied, no source audio track retrieved (the B-roll is used silently under the interview).

Full download links are preserved in `sources/source-manifest.json`. The 1.6 GB full interview was not downloaded. These excerpts are included in the editable project ZIP; they have not been graded or captioned. Container metadata may differ after extraction.

## Editorial decisions

The opening is taken from the end of the same answer. The qualifier **“In many senses”** is retained. The explanation follows: interest in origami, the sheet of material, rules, folding, complexity, geometry and algorithms. Longer pauses and the “And so it's very kind of…” filler are removed. Speech remains at 1× speed; no voice cloning or replacement was used.

Two modest crop sizes, 608×1080 and 580×1032, produce restrained changes in shot scale. Two accompanying VPRO cutaways show hands and a paper structure. English phrase captions use white text and pale green keyword emphasis; large safe margins leave the face and common mobile UI clear. Gentle brightness/contrast and colour balance adjustments are applied consistently.

Dialogue processing: 75 Hz high-pass, 6 dB FFT noise reduction, 2:1 compression, measured two-pass loudness normalisation targeting −16 LUFS and −1.5 dBTP. Actual encoded measurements are recorded in `qa/QA-REPORT.json`. No music was used.

## Actual tools / editable project

**FFmpeg + Python/Pillow** were used to render the edit. **faster-whisper 1.2.1 / Systran Whisper base.en** was run locally to check speech and timing. No speech was uploaded to a transcription service. No paid material or service was purchased. No Premiere Pro, CapCut, Shotcut or DaVinci Resolve project is claimed or fabricated.

The build is an editable code-based project, not an NLE timeline. `project/timeline.json`, ASS subtitles and the scripts are the actual project files used. FFmpeg and Python binaries, Microsoft fonts, the speech model and installed Python packages are not redistributed in the delivery ZIP.

## Rebuild

1. Install Python 3.12+ and Pillow, and an FFmpeg build with libx264, AAC and libass.
2. Extract the editable project ZIP, retaining its directory structure and included footage excerpts.
3. On this Windows machine the existing binaries are `C:/AI/ScreenCutter/runtime/ffmpeg.exe` and `ffprobe.exe`. Elsewhere set `FFMPEG` / `FFPROBE` or pass the `--ffmpeg` / `--ffprobe` options to `build.py`.
4. The supplied design uses installed **Segoe UI**. Set `REEL_FONT` and `REEL_BOLD_FONT` to regular/bold font paths if those fonts are unavailable. Set `REEL_ASS_FONT` to the corresponding installed family name for caption rendering. Using another font changes the appearance; recheck line lengths and the credit card.
5. Run:

```powershell
python tools/build.py
python tools/comparison.py
python tools/finalize.py --package
```

The working folder contains the complete portfolio staging copy. To avoid bundling unrelated old projects, the editable ZIP includes only the new case HTML/CSS and the shared stylesheet/favicon. Rebuilding supplies its video assets. Home/back links require the complete portfolio staging copy, which is kept separately in the working folder.

Rendering from included data needs **no speech-model download or network connection**. The saved `qa/original-asr.json` supplies timing. Optional ASR reproduction uses `tools/transcribe.py`, the dependencies in `tools/requirements-asr.txt`, and the freely downloadable Systran model; it is not needed to render the delivered edit.

For timing/crop/audio changes edit `project/timeline.json`, then run a full build. Caption phrases and keyword choices are in `tools/build.py`; individual word text is in `captions/words.reviewed.json`. For direct ASS/SRT/VTT editing, preserve the edited files with:

```powershell
python tools/build.py --skip-camera --keep-captions
python tools/comparison.py
python tools/finalize.py --package
```

`--skip-camera` is only safe if source cuts, crops and B-roll source ranges have not changed. For timeline changes, omit it. `--keep-captions` assumes caption timing still matches the new edit. Recheck after changes. Output dimensions and graphic composition are fixed at 1080×1920 in the renderer.

## Preview the portfolio locally

On this Windows computer, double-click `preview.cmd` in the working folder; it starts only a local loopback server. Otherwise, from the project directory:

```powershell
python -m http.server 8791 --bind 127.0.0.1 --directory portfolio-stage
```

Open <http://127.0.0.1:8791/talking-reel/>. Native players do not autoplay sound. Playing a second video pauses the other player. The page includes technical verification notes. Publication was subsequently authorised by the owner.

## Verification and remaining review

Technical results are in `qa/QA-REPORT.json`; visual inspection boards and browser captures are in `qa/`. Exports are decoded end to end, checked for expected codecs/dimensions/duration and measured for audio level. Caption timing/overlap, selected transcript text, local links, image loading, mobile overflow and actual browser playback are checked.

**Auditory review limitation:** owner listening was agreed, but its completion has not been independently confirmed to the assistant. The owner subsequently authorised publication. The assistant in this session cannot receive audio input and does not claim to have listened to the soundtrack or manually verified every spoken word by ear. Please watch the 34.4-second master and the comparison with sound, especially: the opening qualifier (0–6 s), the quick rule explanation (12–17 s), the filler cut before “A simple setup” (about 20 s), and the final word “fold” (about 29 s). Listening remains an owner review item; this package does not claim auditory approval by the assistant.

This is a demonstration of a small editing task, not a promise that every future source will need the same time or price. Starting price of an analogous simple edit: **$10**, scope agreed separately.


## Revision — 2 November 2026

Removed header/footer service labels and the name plate from the speech section in all video exports. The comparison retains only large BEFORE / AFTER markers; source attribution is on its own final card. Large spoken captions, speech timing and audio processing are retained. Posters, case frames, the cover portrait and archives are regenerated from the new MP4 exports. Small captions and labels adjacent to the page players are removed.
