# Automated production evidence, version 1

Protocol ID: `kpb-automated-v1-20260924`. Prepared September 24, 2026, before new renders or rechecks. The accompanying `freeze.json` records the exact UTC freeze time and input/code hashes. Execution must refuse a changed hash.

## Amendment and interpretation

This adds an independent automated study. The September 23 `kpb-bottle-v1` human-acceptance protocol, its 27-slot ledger and its thresholds remain unchanged. These results do not complete any of those 27 slots. Objective results may be published while human acceptance, human editing time, customer validation and cost per accepted ad remain unknown.

The purpose is to establish which retained files meet explicitly listed delivery properties, provide reproducible conventional exports, and show what technical success cannot establish. There is no new paid generation, no provider access and no rate-card repricing. No model ranking is planned.

## Cohort A: three conventional bottle exports

Use the existing fictional Kingy Field Bottle image and the three briefs in `fixtures/briefs.json`: launch, detail, retarget. Run the existing FFmpeg/Pillow `render.py` once for each brief, in that order, with `demo=False`, no supplied generated shots and no human edit. This reuses the published conventional method rather than replacing its renderer. One output per brief is a technical fixture, not three independent observations of human production speed or creative success.

Before each run retain its start record; afterward retain the exact output, any error, render time, input hash and technical measurement. A failed run is retained as failed. Do not silently rerun or alter the renderer in this cohort. A correction requires another version or a separate diagnostic explicitly excluded from the study.

Technical gates, fixed before execution: one video stream; MP4 container; H.264; width 720; height 1280; frame rate 24; duration 15 ± 0.1 seconds; exactly 360 decoded frames; zero audio streams; full video decode without FFmpeg errors. Sample the output at seconds 1, 6 and 11 for a separately labeled agent visual check. Source text and visual samples cannot substitute for human acceptance, full OCR or a measured readability study. Copy, fidelity and creative acceptance remain unscored by the technical gates.

## Cohort B: recheck every currently retained Creative Work Lab original

Snapshot attempts 001–011 from the separate `kingy-creative-pilot-001` study, inclusive. This is the complete set of retained original MP4s discovered for the current recheck, not its completed 18-run design or a balanced provider comparison. Preserve the source protocol, frozen input manifest, source attempt records and recorded bills without modifying the originating project.

Each recheck must verify the original attempt's recorded SHA-256, decode the entire file, and measure duration, size and audio presence. Apply the originating raw-media mechanical checks: one valid video; positive duration no greater than 10 seconds; aspect ratio within 0.03 of 16:9; no audio, since these originals precede separately added narration. Use the existing source checker, retained with its hash. These raw clips do not satisfy the bottle cohort's 15-second portrait brief and are never scored against it.

Publish attempt-level results including missing files, hash mismatches, decode errors or failed properties. Preserve every included original. No repair, recrop, trimming, overlay or narration is added in this cohort. The source records say final overlays/narration were not completed; human acceptance remains unknown. Retained earlier agent frame-sample observations may be quoted only as dated agent observations, with their still-sample and incomplete-playback limitations. They are not new human judgments or provider pass rates.

## Accounting

New paid requests and new provider consumption for these local executions are zero. Do not describe this as zero operating cost: electricity, machine allocation, agent service costs, source-image creation cost, research labor and human review are unmeasured. Record automated render and check duration separately from human time, which stays null.

For cohort B sum the exact `provider_cost_usd` recorded per source attempt using decimal arithmetic and show the billing-source description. These are previously recorded request-level billing observations; this recheck does not log into a billing account or establish cash purchases, current balances, refunds or taxes. Never count these amounts again as new spending. The September 5 bottle pilot's $4.11 is a third historical cohort and remains separate.

Cost per accepted deliverable is null in both cohorts because no qualifying human acceptance exists. A technical pass must never populate the calculator's accepted-count input.

## Reproduction and publication

Retain protocol, freeze time, source hashes, tool versions, started/completed times, all outputs and errors, CSV/JSON results, source receipts, input assets and script hashes. Publish a baseline workflow ZIP and a retained-video recheck ZIP with instructions, dependencies, limits and file manifests. The checks run locally and make no network calls. Users may run the baseline on the supplied fictional image or adapt it to their own rights-cleared input as a new, unscored workflow.

Keep original results immutable. A changed assessment or correction becomes a dated new report. Public technical results must retain separate fields for agent observations and human acceptance; null is not success. Publish all rows without selecting attractive examples or averaging incompatible cohorts. Link the finished files to Kingy's output-preflight tool for additional user checks, with its scope visible.