# A direct raster-steering trial

The experiment at `/image.html` connects the Conservatory's observe–predict–check loop to the image domain. It edits actual 64 × 64 RGB arrays, with separate errors inside and outside a selected elliptical region. The pixels represent linear RGB values in [0,1]; the UI applies an sRGB display conversion. The task brightens each selected channel by 0.12 and preserves every other pixel. PNG export contains the displayed current image without the selection overlay; JSON retains the linear arrays and full action history.

## Observation and control

The controller receives original pixels, current pixels, target pixels and the selection mask. It does not receive the hidden actuator condition, gain or shift. Two supplied controls adjust a local footprint and global brightness. Four conditions make the local control truly local, add uniform spill, shift its footprint, or add a nonlinear outside response to larger edits. These are authored experimental mechanisms, not models of a particular image editor.

The learned controller applies a 0.02 probe to each control and physically undoes each probe. The resulting pixel differences estimate two response columns. A two-variable weighted least-squares calculation proposes a combined edit, giving outside preservation weight 5 and target fitting weight 1. Individual commands are bounded to ±0.3. Each proposal records predicted target and outside errors before the actual operation.

The guarded policy retains an edit only if target error decreases and the measured outside changes meet both preservation tolerances. Rejected trials are undone and the next proposal is halved. The unchecked policy keeps every proposed edit. The fixed policy assumes the local control has unit gain and no spill. A no-op control leaves the original unchanged. All have a maximum of 24 physical commands and eight steering proposals; an early finish uses fewer actions. Probes and undos count. They are not free simulations.

## Acceptance and failure

Success requires all three numerical conditions:

- Target-region RGB root mean squared error ≤ 0.015.
- Outside RGB root mean squared error ≤ 0.008.
- Maximum absolute outside channel change ≤ 0.025.

These thresholds are declared engineering choices, not human perceptual measurements. The no-op satisfies preservation while failing the target. A shifted actuator can make the desired image unavailable through the two controls. A locally measured response can mispredict larger nonlinear edits. Guarded editing can preserve the surroundings yet spend its budget without reaching the target. Report all these failures.

Preservation applies to the committed final image, **not every temporary image**. Calibration probes and rejected trials can exceed the limits before an exact snapshot undo. Transient outside excursion is included in each report. This workflow assumes reversible local editing; it is not a control-barrier certificate for irreversible physical actions.

## Research connection

The local `russell/image-steering-accreditation` capability map and chapter 35 distinguish observable response from chosen structure and emphasize testing useful predictions even when hidden parameters are not uniquely identified. This implementation measures the control-to-pixel response directly; it does not claim to recover the hidden actuator's full mechanism.

The research library's preservation and conformance distinction is also explicit here: a numerical metric and a declared acceptance rule are implemented, while perceptual quality and institutional accreditation remain untested. The experiment does not reproduce the library's profile-likelihood, spline, topology, or spatial-field studies. No files in that research repository were changed.

This is now an image-domain experiment rather than only an analogy to watering beds. It remains a synthetic pilot: full pixel observation, noiseless arrays, two supplied controls, a supplied target mask, and reversible operations. Real photographs, user-drawn regions, semantic quality judgments, noisy observations and editor integration need separate evidence.

## Frozen evaluation

Twelve independent image seeds × four actuator conditions × four policies, frozen before execution in `experiments/frozen/conservatory-v7/manifest.json`. Primary outcome: fraction of trials meeting all three requirements. Conditions are equally weighted; cases within a seed are correlated. Report target and outside error separately, alongside action cost and rejection count. Predicted pixel-channel values are counted separately from physical image commands. Development seeds 9301, 9311 and 9341 are excluded. Results are accessible through the page and linked in `README.md`.