# Coordinate control: a falsifiable implementation claim

Frozen before evaluating outputs. Three existing library reproductions: Matisse Piano Lesson, cropped Degas Dancers Pink and Green, Gauguin Vision of the Sermon. No generated substitute artwork, training or scene reconstruction. Sources remain unchanged.

## Claim to test

Within a bounded color-adjustment family, spatial chroma coordinates can become executable targets. A constraint solver should move the chromatic centroid while preserving six hue-family chroma totals, per-pixel OKLab lightness, local hue direction, and a fixed protected region. It should decline targets outside its reachable set. This is a narrow control claim, not proof of artistic intent, aesthetic improvement, exact perception, model superiority, or a new optimization algorithm.

## Representation and controls

Use the user's Color Kernel v0.4.8 color conversion functions, extracted verbatim from notebook cell 11, plus the existing ISR fields for independent readout. Work at maximum dimension 512, preserve aspect ratio, and retain a lossless normalized baseline. Keep signed normalized x/y coordinates. Chroma mass uses raw OKLab sqrt(a²+b²), not per-image percentile-normalized ISR fields or a changing top-percentile mask.

At each pixel hold L and hue direction fixed and multiply a,b by a positive gain. Gain is a smooth 6×6 spatial basis with 72 bounded positive/negative controls, scaled by available sRGB gamut headroom. Gains lie within [0.35,1.65]. The central [0.4,0.4,0.6,0.6] rectangle is protected exactly with a 0.04-frame transition outside it; this is a geometric test region, not automatic object recognition. Pixels do not move; chromatic emphasis moves. This does not disentangle material from illumination or edit semantic objects.

Six source hue bins (60° each) define fixed chroma budgets. A bounded linear feasibility problem precedes a quadratic minimum-change solve using SciPy HiGHS/SLSQP. Constraints use actual per-pixel contributions. No language-model call participates in the solver. Edits are reversible states evaluated from the baseline, not cumulative destructive color conversions.

## Comparison

Both methods receive identical sources, controls, gamut bounds, protected region and target. Baseline: minimum-change coordinate fit without the six hue-budget constraints. Candidate: same objective and target plus those constraints. The comparator is an automated version of spatial color adjustment, not a timed expert workflow or a generative-model benchmark. Include no-op. Do not claim that imposing a constraint is itself novel.

Requests, fixed for every image: chroma-centroid displacements (+0.025,0), (-0.025,0), (0,+0.025), (0,-0.025), (+0.06,0), (-0.06,0), (+0.025,+0.025), plus a deliberate contradiction that locks all controls while requesting (+0.025,0). Do not replace failed cases or shrink targets to manufacture successes. Report attainable horizontal/vertical intervals subject to the candidate's other locks.

## Acceptance and evidence

Check saved 8-bit PNGs by reloading them, independently recomputing OKLab fields and actual centroids. Target error <=0.002 normalized-coordinate Euclidean distance; each hue-bin chroma budget drift <=1% for bins contributing >=1% of source chroma; max absolute per-pixel OKLab L drift <=0.004; protected rectangle pixels exactly unchanged; no spatial resampling after baseline normalization. Report hue-angle error above source chroma 0.02, gain range, gamut safety, ISR structural deltas, solver time and evaluations. State quantization limits plainly. A mathematical feasible point and an accepted exported edit are separate outcomes.

Visual review must inspect all exported candidate outputs and representative comparator outputs for contour artifacts, unnatural color and visible discontinuities. Numerical success is not artist approval. Report impossible requests as refusals, not successes on the image-edit task. Report false rejections/failures where detected. Test selective rollback: keep a second coordinate request while removing the first; it must equal direct rendering of the remaining state, not merely a hash of copied files.

## Competitive boundary

Image-space optimization and geometric constraints have prior art (DiffVG, saliency-driven manipulation, point-based diffusion editing). This work's initial contribution, if it passes, is a concrete integration of the user's field coordinates with a bounded executable edit and an explicit feasibility/verification contract. No novelty or superiority claim beyond the observed test is warranted. The important next boundary would be adding geometry/material/semantic operations without losing these guarantees.