# Correct the interpretation. Carry the correction.

This Afterimage study makes the meaning of a persistent decision inspectable and revisable in native video. An agent-authored example starts with “keep her isolated,” interpreted as preserving a screen position. The final correction protects a gap between the head/hair and visible rock or steep-terrain surfaces, while allowing screen position to change. The same correction is replayed in two existing Settlers passages.

## Findings

| Final separation predicate | Approach | Crossing |
|---|---:|---:|
| Position interpretation | 0/24 frames | 0/24 frames |
| Corrected interpretation | 24/24 frames | 24/24 frames |
| Direct explicit constraint | Identical corrected camera | Identical corrected camera |

Both interpretations satisfy the existing native checks on all 48 source frames: protected framing, minimum projected head height, sampled head visibility, source-relative character geometry and source-relative ground support. Camera steps remain within the existing limits. These are correlated frames from two one-second passages of one previously used asset, not independent trials.

The direct comparison has the same predicate, inputs, camera pool, guards and ranking. Equality is by construction, not evidence from an independent implementation or a skilled-artist study. The obsolete position interpretation is an ablation, not a competent corrected-workflow baseline.

## A passing score was insufficient

1. **Screen-position v1:** keeps the initial head-center anchor.
2. **Named-rock v2:** checks projected named-rock envelopes. It passed, but visual inspection found the character still against a terrain bank that was missing from the entity binding.
3. **Convex-terrain v3:** adds steep terrain patches, whose convex envelopes overfill empty space. This approximation is rejected.
4. **Visible-formation v4:** intersects actual native triangles and classifies the nearest included surface. The candidate family, 2% width gap and native guards stay fixed.

The correction and all numeric thresholds are agent-authored, not a human teaching session. Protocol amendments and the initially misleading film are retained in `history.json` and `history/`. This is an iterative development experiment, not a preregistered benchmark. The final rule is fixed before its repeat search and has one canonical SHA-256 identity across both passages.

## Measurement and controls

- 108 camera paths per passage; 216 total. Same pool for both interpretations. Fixed 28 mm lens, 8-degree path sweep, inherited 0.8 m root displacement, original performance and 24 fps timing.
- Head/hair screen envelope rasterized at 320×180. Background rays sample the envelope plus an expanded buffer. The intended gap is 2% of image width. Positive results report a lower bound within the tested buffer, not an exact larger distance.
- 1,319,672 native triangles: 78 rock/mountain instances and three terrain meshes. A nearest rock surface is forbidden; terrain is forbidden when its absolute surface-normal z is below cos(45°). Flatter floor is allowed. Plants, refraction, shadows, contrast and aesthetic judgment are outside the predicate.
- The ray monitor uses static environment geometry. Rock geometry is verified in the initial native render records; the three unchanged paths retain those receipts. The new camera replay checks head/hair geometry against the extraction on every frame. Full terrain geometry is checked at passage endpoints. This does not certify every intermediate terrain deformation.
- Failed candidates stop at the first failing frame. Their remaining frames are untested. Both selected paths are measured fully. JSON nulls indicate unavailable measurements, not zero errors.
- Existing native checks and PNGs are reused only for three exactly unchanged camera/placement paths; their source records and hashes are retained. The new approach camera is freshly checked and rendered. All four paths receive the revised ray measurement.
- Common render settings: Eevee, 16 samples, 640×360, fixed foreground-defocus threshold 0.5 and gain 80. No generated/interpolated frames, source retiming, reshaping or environment relocation.

A clear formation gap is one property. It can cost screen-position fidelity, camera angle or artistic quality. These results demonstrate an explicit correction affecting selection and persisting in this scene. They do not demonstrate concept learning, language understanding, novelty, artistic superiority, effort reduction or transfer to an unfamiliar asset.

## Replay the selected camera

Download `replay.py` and `recipe.json`. Obtain the unchanged original Settlers `01_desert.lighting.blend` with its linked libraries and textures. The existing Afterimage revised scene is a different file; the script deliberately checks the original source SHA-256 before rendering.

```sh
blender --background --factory-startup --disable-autoexec \
  --python replay.py -- \
  --source /path/to/01_desert.lighting.blend \
  --recipe /path/to/recipe.json \
  --passage approach --rule corrected \
  --output /path/to/new-render-folder
```

Tested runtime: Blender 3.4.1. Use `--passage crossing`, `--rule initial`, or `--single-frame 17` to inspect another replay. Output is PNG frames plus a receipt; the source scene is not saved or overwritten. `recipe.json` replays selected cameras; it does not rerun the search. The method-source files document the executed local pipeline and require the source extraction records and existing native adapter to rerun the full search.

## Records and lineage

- `protocol.json`: final corrected interpretation, authority, thresholds and amendment history.
- `recipe.json`: both selected camera paths, source hash and common render policy.
- `evidence.json`: complete selected frame observations, native checks and render receipts.
- `candidate-record.json`: all candidates, early-exit observations and native rejection records.
- `history.json`: reasons the earlier interpretations were revised.

Inspired by RFS’s [Changing alphabet](https://un1crom.github.io/computer-week-2026/changing-alphabet/) and [Second chamber](https://un1crom.github.io/computer-week-2026/second-chamber/). Their questions about correctable interpretation motivate this test; no claim is made that this explicit-rule system reproduces learned representational change.

Source: [Settlers — Blender Studio / Blender Foundation](https://studio.blender.org/projects/settlers/?asset=2533), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Source assets unchanged. No API credits used.