# Recheck the retained nine-ad cohort offline

This supplement does not generate videos or call an external API. It assesses the retained September2026 files. Do not run the original archive's bench.py or provider-generation code for this recheck.

1. Download and extract the existing original bundle: 
https://kingy.ai/wp-content/plugins/kingy-production-benchmarks/assets/kingy-nine-output-cohort-v1.zip
2. Extract this supplement separately. Install/locate Python 3 and FFmpeg/ffprobe; Pillow is needed only for --frames. Our run used Python 3.12.14, FFmpeg/ffprobe 8.1.1 and Pillow 12.3.0 on macOS 26.3 arm64. Installing dependencies may require network access; the recheck itself is offline.
3. From the supplement directory, with verify_media.py beside reproduce.py, run:

```sh
python3 reproduce.py --source /path/to/extracted-originals --output /path/to/new-recheck --frames
```

Output must be a new directory outside the originals. The adapter checks 33 archived manifest entries, all 9 final exports and 12 raw decodes, and extracts 135 exact frame indices with the published alias map. It reads source files without importing or running their production scripts. Result JSON is written to the new output directory.

verify_media.py is the unchanged frozen checker. reproduce.py adapts its checks to the public ZIP layout. Do not invoke the frozen checker's historical-audit CLI against the public ZIP: that mode expects private historical records. The published technical-results.json describes the larger 61-hash original audit, whereas reproduction-results.json describes the portable 33-entry public check. These counts intentionally differ.

Compare output frame PNG hashes with reproduction-results.json. Matching FFmpeg versions are recommended; encoding or library changes may affect extracted bytes or contact-sheet typography. The diagnostic criteria and source hashes should still agree. A hash mismatch needs investigation, not a quality failure label.

frames/A01...A09 contains every contact sheet and seven cited full-resolution PNGs. The portable recheck extracts all 135 PNGs. Review path tokens such as [assessment] and [workspace] replace private local roots; no verdict wording was changed by that substitution. Original reviewer prompts are in each review's provenance. The coordinator's final decisions are in results.json. AI passes are restricted to the cited inspected samples, and human fields remain null.

SHA256SUMS.json covers all files inside the supplement ZIP except SHA256SUMS.json itself. bundle-sha256.json on the page covers the ZIP. Original videos are linked in results.json and in the unchanged original bundle. No new video copies or hidden attempts are included here.