Video compression — quality-contract aggregate

Owner-approved public evidence artifact — reviewed 2026-08-09

Substantiates: video-compression metrics batch-outcome and verify-retry.

What was recomputed

A committed extraction script read every retained batch job.json record and recomputed the acceptance contract from the measurement fields. It did not trust the stored accepted flag. For each completed file, acceptance required:

1. independently measured full-output average SSIMULACRA2 at or above 70; and
2. no measured window below the minimum floor of 50.

The script also summed input/output bytes for completed files only, counted terminal states, and counted files whose final encode-attempt number exceeded one.

Redacted extraction result

batches inspected: 2
files: completed=112  kept-original=41  errors=9  total=162
bytes: 6,869,302,702 -> 2,550,282,532  (62.9% reduction)
quality contract: recomputed-pass=112  violations=0
verify-and-retry engaged: 39

Qualification retained with the result

• Fifty-four older accepted records predate the minimum-floor field being written into telemetry. For those records the extractor uses the service’s then-current code constant, 50.0, and reports that assumption rather than presenting it as recorded per-file data.
• Forty-one files were retained unchanged because the encode was not beneficial; they are not counted as accepted outputs.
• Nine failures remain failures with machine-readable reasons; they are not removed from the denominator.
• This establishes enforcement of the chosen metric contract. It does not establish perceptual transparency or suitability for footage outside the reviewed archive batch.

Privacy treatment

No filenames, UUIDs, media, paths, timestamps, host details, or per-file scores are reproduced. The aggregate can be regenerated from private primary records by scripts/evidence/video_compression_metrics.py during owner review.