# SEEDS recorded-TV ad muting demo
Completed locally on 12 October 2026 (Vancouver).

## Use
Use the hosted player. Click Jump to next ad, then compare Ads muted with Original sound. Both complete recordings are streamed in short HLS segments at the source resolution. Reviewed timestamps control the player's mute state; this is an offline analysis demonstration, not live AI inference. Master downloads and internal scripts remain local and are not part of this shared page.

## Result
- HGTV: 28:57.32; three reviewed break intervals; 436.1 seconds muted.
- Weather Channel: 23:39.93; nine reviewed intervals, including standalone network promos; 439.2 seconds muted.
- All12 two-second samples inside the mute intervals measured zero PCM RMS; all14 programme samples retained audio level within the specified15% tolerance after AAC re-encoding.
- Compressed video packet SHA-256 hashes match each original: picture data unchanged.
- Both complete output videos fully decoded without FFmpeg errors.
These are media-processing checks, not independent detector precision/recall results.

## Method and corrections
Initial Gemini2.5Flash native audiovisual five-minute chunks returned unusable timing and were rejected. A second pass used Gemini3FlashPreview with 27 clips of <=120s, reduced-size pictures plus original-source audio, and an overlaid clip-second clock. Prompts asked for visual, speech and music/sound evidence. The model was not given a list of known ad positions or downloaded captions. Raw responses are in model-results.json, including errors.
Four second-pass clips still returned out-of-range timestamps. A single end-to-end automatic pass was therefore insufficient. technical reviewed full-recording contact sheets at10s sampling, one-second sequences around candidate boundaries, and four targeted audiovisual rechecks. Final corrected intervals are explicit in mute-intervals.json; nearby visual cuts refine some endpoints. Remaining non-snapped boundaries are approximate to about one second, not certified frame-accurate.

Specific corrections: restore HGTV project-instruction card1370–1389, lodging acknowledgement1697–1706 and production credits/logos; recover Weather Channel Motorola commercial1160–1190 and preserve sponsor logos over programme content; preserve Weather Wisdom and local forecast bumpers; fix mistimed Denny’s, prescription-plan, Pacific Update/Cablevision and business-planner/book breaks. Muting policy includes separate full-screen channel/website promos, but not normal programme transitions or sponsor credits.

This is an AI-assisted, technically reviewed offline demonstration. It is NOT a validated unattended live detector, physical-remote test, or independent human-labelled accuracy benchmark. The same reviewed timestamps drive playback; no AI is running inside the HTML player. Two old recordings do not establish robustness on modern streaming, sports, drama, overlays, native ads or household equipment. Next engineering test should freeze the algorithm, process unseen clips sequentially with no future context, compare to independent annotations, and measure false mutes, missed ads and start/end latency. The analysis feed must remain audible when the output is muted.

## Sources and evidence
Source links/checksums: sources.json. Reviewed windows: mute-intervals.json. Raw model outputs: model-results.json. Processing QA: verification.json. Original rights remain with their owners; this local evaluation does not grant redistribution rights.
## Hosted preview — 13 October 2026
H.264 baseline/yuv420p video and AAC audio were re-encoded for streaming. Original packet-hash and sample measurements describe local masters, not these compressed streaming copies. The player applies the same reviewed mute windows during playback.