Round complete: six transfer studies, findings and full ledger →

Closer is not
always better.

Twelve independent construction tests. Four passages from one figure, three methods, one common starting canvas and target. The retained U03 artwork is unchanged.

Decision

Preserve useful construction. The smaller view performs well for broad passages. A larger view helps the hand only marginally in this run; it is not a universal upgrade. Use it selectively, and inspect the whole.

Reject routine erase-and-rebuild at this budget. All four reconstruction attempts have worse target agreement than their starting passages. The visible result is granular, with broken masses and dark islands. This rejects this reset procedure at this budget—not reconstruction as an artistic practice.

The fixed target still dictates the folds. More observation reproduces clothing detail rather than deciding what to omit. A different artistic direction needs a changed visual plan, followed by a test of whether the renderer can execute it.

Compare actual results

Show common visual target for this passage

Read the set

All twelve whole drawings

Process and methods

1. Load and hash-check the same stage-08 native scalar canvas for every trial.
2. Use the same previously generated lowered-pose target. No new generator call or new critique-led target was made for this batch.
3. Select four rectangular regions: clasp, flank contour, sleeve, and arm–torso interval. Three methods act only within that region.
4. The existing planner uses a 350-pixel-wide local view. The fine variant uses 700 pixels. The reconstruction variant first removes 92% of tone inside an inset polygon with a 24-pixel feather, then uses the same 700-pixel view. This is a manually selected reset, not learned semantic reconstruction.
5. Reset vocabulary familiarity and use seed 11 for each run. Give each the same cap: 900 accepted marks or 16,000 attempts. All twelve reached 900 marks. Equal mark budgets do not imply equal computation or sufficient time to reconstruct erased work.
6. Replay removal, where applicable, followed by all accepted commands. Check exact scalar hashes and zero changes outside the region. Inspect crops and whole images separately from error measurements.

The labels “current” and “fine” refer to experimental 350- and 700-pixel presets of the existing planner. Earlier U03 passes used several scales. The new spacing adapter is deliberately not introduced here, so a renderer change does not confound the scale/reset comparison. Its curved-path and broader mark-family validation remains optional follow-up.

All variants use the same target-error acceptance rule. They test execution, not whether the target itself is a good drawing. Each result starts independently; the twelve outputs are not accumulated into a final image.

Measurements

Error reduction compares each final passage to the original, pre-erasure starting passage. Negative values mean the final result is farther from the common target. The post-erasure starting error is separately recorded in each receipt.

Limits and next checkpoint

This is one figure, one planner seed and one mark repertoire. It does not establish a general resolution policy or rule out better reconstruction methods. In particular, resetting a region without a new structural plan is a weak proxy for what an artist means by rebuilding.

Next: six exploratory transfer outputs—three contrasting briefs with and without critique. Carry forward a conservative renderer, make the critique's structural change explicit, and compare whole images. The batch does not justify more U03 polishing.