# Two guided revisions: control record

The intended evidence is selective preservation under a substantial transformation, then preservation through a second revision. Visual polish alone is not success.

## Fixed before generation

The active frozen-protocol.json and protocol.sha256 identify the source, prompts, guide pixels and timestamps, seed, model, SDK, permitted changes, protected decisions and review samples. Stage two takes stage one's actual uncorrected output as its video input. It receives the original identity/performance anchors, graded to the second direction. This is external reassertion of decisions; it does not demonstrate internal model memory.

The five guides are at source frames83,101,125,143,155:0,.75,1.75,2.5,3seconds. Other reviewed frames are92,110,119,134,152,166,171,174. Every output frame is retained. No post-result prompt/seed/guide retuning or successful-output rerolls are permitted.

## Preflight amendment

Two identical requests were rejected with HTTP 400 before any generation task was returned or credits spent. The server imposed a maximum keyframe timestamp of 3 seconds on the 3.833333-second input. The final guide was therefore replaced with the source frame at 3 seconds, before viewing any generated outcome. The rejected protocol, validation response and amendment remain available. Earlier visual preflight strengthened the conventional amber grade before any API submission.

The maximum of two video calls means two accepted generation jobs. The two unpaid validation rejections are recorded separately. No hidden successful generations or candidate selection are permitted.

## Exposed controls

Official SDK 5.20.1 exposes the Aleph2 source video, prompt, up to five timed images, optional edit ranges, seed, ratio/target aspect, output encoding/profile and moderation. This test uses MP4, seed 20261001, no range or aspect override, default public-figure moderation. There is no exposed regional preservation mask, CFG, sampling-step or denoising-strength control. Internal model weights/preprocessing are not fixed by this project. Same seed is documented to produce similar results for an identical request, not an exact replay; the two stage inputs and prompts differ.

References:
- https://help.runwayml.com/hc/en-us/articles/51683104370451-Creating-with-Edit-Studio
- https://docs.dev.runwayml.com/assets/inputs/
- https://github.com/runwayml/sdk-python

## Observed sequential input difference

The initial input is 1280×720; stage 1 returned 1920×1080. Stage 2 therefore receives the 1920×1080 file, preserving its exact bytes and compression as required by the protocol. Resolution, accumulated encoding, input content and edit brief differ across stages. They are not independently isolated variables.

## Comparison boundary

The deterministic reference uses the existing restored character/prop and generated frozen plate, followed by fixed linear-RGB grades. It does not reconstruct physical illumination or improve known source-mask/edge/contact limitations. This is an agent-built conventional reference, with access to all 92 native masks. The API receives five guide images. The resulting comparison describes workflows with different information access; it cannot isolate an algorithmic advantage or establish that Afterimage beats a skilled compositor.

The older text-only generation used a different prompt and stricter appearance prohibition. It is historical context, not an isolated ablation. One development shot, one seed, and agent visual inspection cannot establish generalization, artist preference, saved labor, or production readiness.

## Reproduction

prepare.py builds the fixed reference clips and guides from the existing native source/cleanup assets. It refuses to overwrite a frozen protocol. review.py decodes complete clips and produces timestamp-matched comparisons, preserving all 92frames without retiming. The API runner is held outside this portable folder to keep account execution separate; the public run receipts provide all non-secret settings and lineage. Downloaded raw outputs are retained because model reruns are not exact reproduction.

Source: Settlers, Blender Studio / Blender Foundation, CC BY 4.0. https://studio.blender.org/projects/settlers/?asset=2533