# Production Benchmarks · Report 002

Published September 24, 2026. Study `kpb-automated-v1-20260924`. Automated technical checks and agent visual observations; no human acceptance decisions.

## What we learned

The existing conventional template produced 3 valid exports from 3 different bottle-ad briefs. Each is a silent 15-second H.264 MP4, 720 × 1280, 24 fps, with 360 decoded frames. All ten fixed technical checks passed. New provider consumption was $0. The original still image and conventional editor preserve a simple repeatable format; these files do not contain generated product motion. Whether this format meets a client's brief needs human review.

Separately, all 11 retained raw outputs from the incomplete Creative Work Lab study passed their own duration, aspect ratio, audio, decode and recorded-file-hash checks. These are approximately eight-second landscape clips from different fictional product briefs. They are not comparable trials of the bottle-ad task. None has the required final overlays; the filter brief also needs fixed narration. A technical pass is not a finished deliverable.

The historical request records total $8.54384 in provider consumption. These are prior request-level billing observations, not new charges, today's rate estimates, or provider-issued invoices. The provider account was not rechecked. Tax, credit commitments, source-image cost and human editing time are unknown.

The sampled frames expose useful failure questions even when a file passes: hands appear despite a no-people instruction in attempt 001; several invoice interfaces contain pseudo-text; an off/on indicator sequence makes the kettle action ambiguous in attempt 008. These are agent observations at fixed sample times, not blind human judgments. See agent-observations.md and the original files before drawing a quality conclusion.

## Scope and chronology

The protocol and executable/input hashes were frozen at 2026-09-24T18:46:36.275449+00:00. The first render began at 2026-09-24T18:46:36.416919+00:00. Freeze SHA-256: `e07817b179811269cb05d9e04b6964577058241364318fed92525be4dfdc5f56`. The local timestamps and hash manifest document ordering; there is no independent timestamp attestation.

The three baseline files are one run per brief, not three independent operators or repeats per brief. Every retained raw clip available at the freeze was included; no output was removed for quality. Cohort A and B use separate technical gates and denominators. There were no runtime failures in these fourteen local operations. Original human study: 0 of 27 sessions complete. Human acceptance decisions: 0. Human editing minutes and cost per accepted ad: unknown (null), not zero. No provider ranking is supported.

## Reuse and check

- `baseline-workflow.zip`: the original bottle fixture, briefs, truth sheet, exact conventional renderer, three exports, run records, protocol and checksum verifier. Requires Python, Pillow, FFmpeg/ffprobe and an available Arial or DejaVu Sans font. Run `python3 verify.py`, then `python3 render.py launch reproduced-launch.mp4`. This local command makes no paid provider request. Platform/font versions can change encoded bytes; assess technical properties as well as hashes.
- `retained-video-recheck.zip`: all eleven original MP4s, safe extracts of the historical request/receipt records, the original mechanical checker, current results and checksum verifier. Run `python3 verify.py`, then `python3 score_media.py originals/attempt-001-original.mp4 --attempt receipts/attempt-001.json`. This rechecks a retained file; it does not generate a new video. Exact private source-record hashes remain in automated-freeze.json; public extracts are explicitly labeled and have their own manifest hashes.
- `automated-results.json`, `automated-attempts.csv`, `protocol-automated-v1.md` and `automated-freeze.json` preserve the study record. Display contact sheets may be optimized by the website; byte-exact PNGs are inside the ZIPs.
- Use [Output Preflight](https://kingy.ai/output-preflight/) to inspect a file before editorial review. It does not establish creative acceptance.

## What remains

A matched new generated-workflow comparison needs approved provider access and consumption. The original planned 27-session comparison also needs measured operator time and independent human reviewers. Neither is implied by this local study. The older September 5 feasibility records and $4.11 consumption are preserved in Report 001 and excluded from all counts and totals here.

Technical environment: ffmpeg version 8.1.1 Copyright (c) 2000-2026 the FFmpeg developers; Python 3.12.14. Original files and machine-readable results support corrections. A new protocol version is required if the fixed gates change.