# Preservation-first image edit audit

Tested October 7, 2026 in Aixio Studio. This is a saved private Skill plus a manual inspection recipe, not an automatic pipeline.

## Inputs

1. Put the final composition first.
2. Put product references or localization guides after it.
3. Remove labels, guides and copy that should not survive unless they are themselves part of the test.

## Prompt structure

- `Use image 1 as the canvas.`
- `Use image 2 only as [product identity / localization guide].`
- `Change only [one target].`
- `Preserve [object count, geometry, pose, camera, crop, lighting, background, exact strings].`
- `Do not add [plausible but forbidden objects, controls, copy or styling].`

## Acceptance audit

- Count objects and controls.
- Compare scale, rotations, silhouettes and crop.
- Read every small word and punctuation mark.
- Inspect background, corners and untargeted objects.
- Resize the source to the output and review a 50% opacity overlay.
- Reject the output when the job changes, even if the result looks attractive.

## Handoff

Keep the original, reference, prompt, model/settings, quote, settled charge, output, rejected correction and acceptance note together. Use masks or manual compositing when an exact product, logo or pixel relationship is mandatory. Preserve both the rejected model output and the editable repair source so the manual boundary is auditable.

## Tested scope

Three distinct multi-stage briefs, GPT-Image-2.5 Sunburst, Seedream 5.0 Pro and Nano Banana 2; four paid submissions; one correction; three downstream handoffs. One model output passed and two were rejected then manually finished from locked vectors. Do not infer batch consistency or universal model reliability.