Visual stages, comparisons and decisions from the original run. Internal prompts, operating documents and downloadable implementation files are omitted. Some interactive comparisons are shown as separate images.
Work both sides
of the transition.
A paired native gesture lifts tone from the dark side and deposits on the light side. The longer version makes a more continuous transition on these three canvases, while passing checks for new light pockets and boundary movement in either direction.
What the comparison establishes
All branches start from the same source for their case, share the discovered guide, contrast objective, six decision rounds, protection and cumulative change limits. All use the new two-sided integrity guard. Short pairs cover approximately the old local span; long pairs increase that span fourfold. The lift-only control exactly reproduces the previous round.
Long pairs reduce the measured contrast loss more in all three cases. Short pairs improve the figure more, but the field and interior less, than lift-only. The benefit therefore cannot be attributed simply to adding deposition: spatial coverage matters. Native command counts and footprints differ; this is not a matched-effort benchmark or a human-quality rating.
| Canvas | Policy | Commands | Contrast loss reduction | Added reversal | In / out shift, px |
|---|---|---|---|---|---|
| figure | guarded-lift | 18 | 17.0% | 0.045 | 0.17 / 0.00 |
| figure | paired-short | 12 | 22.5% | 0.027 | 0.00 / 1.99 |
| figure | paired-long | 8 | 30.4% | 0.031 | 0.90 / 1.98 |
| field | guarded-lift | 18 | 13.5% | 0.058 | 0.44 / 0.00 |
| field | paired-short | 12 | 12.9% | 0.002 | 0.08 / 0.25 |
| field | paired-long | 12 | 33.7% | 0.016 | 0.20 / 0.43 |
| interior | guarded-lift | 18 | 34.8% | 0.039 | 1.53 / 0.00 |
| interior | paired-short | 12 | 28.7% | 0.000 | 0.63 / 1.00 |
| interior | paired-long | 12 | 55.6% | 0.000 | 0.37 / 1.63 |
Authored limits: added reversal ≤ 0.06 normalized ink; inward and outward shift ≤ 2 pixels. All final outputs pass. The long figure branch stops after four pairs because no remaining proposal earns admissible gain; other branches reach the six-action experiment cap. These are controller stops, not artistic completion judgments.
Figure
Field
Interior
Process & methods
- Find a continuous contrast edge inside an authored rectangle. Keep its guide fixed throughout the trial.
- At each candidate location, estimate a local tangent and normal. Read which side is darker and center the two paths around a local tonal midpoint.
- Trace a lift path on the dark side and a deposit path on the light side, each offset by 0.45 probe radius. Scale the two pressures from observed ink values.
- Render the pair into a trial canvas. Accept or reject the entire pair together against protected pixels, cumulative change limits, tonal integrity and contrast gain.
- Record both native commands, every candidate decision and exact replay evidence.
This is digital removal and deposition, not conserved pigment transport or simulated physical smudging. The guard measures cross-edge non-monotonic variation and motion of a baseline midpoint crossing. Its tolerances are explicit, resolution-dependent proxies; it has no anatomical or compositional understanding.
Visual reading and decision
The longer paths reduce the punctuated look of local lifting and give the transition more continuity. A narrow gray rim remains a risk, especially where a firm silhouette matters. The figure’s outward movement is close to the permitted limit. We should not read a passing guard as proof that the form is untouched.
Keep the paired gesture as an explicit experimental choice. The next useful batch should test unfamiliar edges, including edges that must stay firm, so we evaluate when to use or refuse the action. No new painting sitting is warranted merely because this local score improved.
The retained artwork and closure
Evidence
Tests cover actual native lift and deposit, useful contrast gain, whole-pair veto, outward expansion rejection, exact replay and policy validation, alongside the existing controller, studio, localization and integrity checks. All nine output hashes, source hashes and code dependencies were verified.