# The demo video Two minutes, eight recorded chapters. It shows the deployed product doing its job, then reviews a card. Watch: https://tomscot-keel.hf.space/video (embedded, 1080p original), or the 1440p mirror on YouTube: https://www.youtube.com/watch?v=EP5-Xh9t1GY. ## Chapters | time | chapter | what it proves | |---|---|---| | 0:00 | Open | The fleet is real: twelve checkpoints, verdicts, validity, fingerprints. | | 0:13 | Verify | The card is recomputed from committed responses and returns MATCH. | | 0:25 | Tamper | One number changes the fingerprint, and verification returns DIFF. | | 0:36 | Families | Same ablation, one judge: Qwen 0.51 to 0.97, Mistral 0.81, Glimmer 0.16. | | 0:57 | Scan | Upload your own response files and get a card. | | 1:06 | Re-score | The judge runs live on the same responses; the rate and fingerprint hold. | | 1:29 | Card review | The card itself: verdict, drift, harmful, validity, structure, limits. | | 1:51 | Close | A measurement, its evidence, and its limits. | ## How it is made - `scripts/record_live_demo.py` drives the live Space with Playwright. The page is captured at 2560x1440 with the page zoomed 1.4, so the text is native 1440p rather than an upscale. A visible cursor and click feedback are injected because headless Chrome does not render the system cursor. - `scripts/build_live_video.py` dubs each chapter with Piper `en_US-ryan-high` at a slower length scale and a slight pitch drop, burns a lower-third title, muxes, and concatenates. It strips any display-matrix rotation so the file plays horizontal. - Two encodes come out of the same cut: 1920x1080 for the site, and 2560x1440 at 14 Mbps for the YouTube master (`VIDEO_W`, `VIDEO_H`, `VIDEO_BITRATE`, `VIDEO_OUT`). - Captions: `video/captions.srt`, served at `/demo.srt`. ## What it must not do - No static screenshots standing in for the product. - No frozen tails: every recording runs past its narration. - No number that is not on a card.