# Painting decisions from trials
Read only when a stage misses its aim. These are observed leads, not style rules. Keep the best scene+PNG; change one cause, inspect the whole image and a relevant crop, then keep/reject/defer. Use `workspace.md` for executable operations.

| Visible mismatch | Next trial | Reject when |
| --- | --- | --- |
| Initial broad wash has unwanted pale seams | Keep all marks/settings; reorder that background stage's water-only strokes before its color strokes. See `prewet`. | White silhouettes shrink or later intended rewetting is lost. This is not “all water always first”. |
| Small focal details spread | From the same baseline, lengthen only the normal-drying wait before details; test a narrow stroke. See `edges`. | Other wet forms degrade. `water:0` does not dry; a wait inside `drying:"hold"` does not resume drying. |
| White cloud or foliage gaps have the wrong large shape | Return to the early wash. Replan brush turnarounds, width and pressure around untouched paper. See `reserve`. | Finer passes create stripes/gaps, or repetitive circular ends replace the intended contour. White paint/lift is not an undo. |
| Pale shadows look like separate pasted patches | Compare a very faint connected underwash beneath the existing shadow strokes. | Lost paper brightness, flattened planes or unwanted bands. One cloud study only; not a validated recipe. |
| Branches are too regular | Revise hierarchy, length, overlap and large negative spaces before contact jitter. See `branches`. | Better twigs damage the whole canopy or make it sparse. A later 17-trial study did not beat the retained tree. |
| Grain distracts from the subject | On fixed standard paper, compare only `brush.granulation`; first confirm the grain is unintended. | Smoother paint loses needed variation. Never tune/remove the fixed paper header. |
| Experimental paper looks whiter but forms blur | Compare full images using the two version tags in `material-v4.md`; retain v3 if worse. | Paper-color improvement masks contour loss. Lower water did not fix one v4 cloud study; avoid endless retries of the same hypothesis. |

## Lifting is a limited tool
Existing `brush.lift` collects mobile pigment with a zero-opacity brush, not a digital eraser. Use `opacity:0`, `lift` in [0,1], no `supply`; water amount and waiting also change the result. Fixed pigment is not directly erased; rewetting can mobilize some, while neighboring wet paint can flow back. Inspect immediately and after a wait. No tissue crease/contact pattern or finite tissue absorption is implemented. A wetness-timing test remains proposed; do not promise crisp tissue-lifted clouds. See `rewet` and schema only when needed.

## Evidence and stopping
The summer study demonstrated one fresh agent could revise its strategy and export a reproducible painting. It did not demonstrate consistent high artistic quality. The cottage study kept standard v3 after 15 rendered candidates; likeness remained partial. Stop adding detail when the large white shapes, value groups or focal scale are still wrong: return to a saved earlier stage, or state the current tool limit.

Schema-valid input, successful execution, reproducible export, local visual improvement and novice transfer are separate results. [Optional evidence summary](/sketch/knowledge/painting-studies-v45.json); raw private references and experiment folders are not published. Finish as `agent.md` requires: deliver the process video and always ask the owner about Gemini music.