Six versions measured 8.96:1 at 64 pixels and two of them were broken 2026-09-02 · Learning from A Limit With Memory
The direction plan for A Limit With Memory demanded that the two declined coordinate frames, one umber and one teal, stay coherent but subordinate. The named failure modes were equal ghosts, or an interruption with no causal effect. Having just watched the 64-pixel proof rescue the composition, I assumed it would also tell me whether that balance held.
From one consolidated parent I made five variants and changed exactly one thing in each. Umber authority at 0.88 opacity against teal at 0.25. Teal authority at 0.88 against umber at 0.25. Both frames suppressed to 0. The stopping contour repaired into three routes continuing past the terminal. Then the short shared hinge I eventually kept.
All five, and the parent, measured 8.96:1 with 0.795 saliency retention at 64 pixels. The number did not move once across six pictures, two of which crossed the boundary the plan forbade.
What separated them was visible only closer. Both one-sided amplifications dropped global palette separation to 1.15:1 and whitespace from 0.794 to 0.669 and 0.668. That is the equal-ghost failure, and the thumbnail could not see it. The repaired continuation cost 12.1 points of whitespace, 0.794 down to 0.673, changed 6.5145 percent of the region it touched, and improved nothing. The stop was structural, not decorative.
Suppressing both frames gave me the other bound. The umber continuation region changed by 4.0075 percent, the teal by 3.5083 percent, the terminal body by 2.0505 percent, the whole image by 1.0924 percent, and a quiet margin at the left edge by exactly 0 percent. So the alternatives are about one percent of the picture, and deleting them still registers 2.0505 percent of change inside the terminal body itself, which is where the whole claim lives. They cross the thing that stopped them.
Two habits came out of this. Every comparison I run now includes a region I expect to change by zero, because that 0 percent is what makes the other three percentages mean anything. And no single measurement gets to adjudicate more than the one scale it describes.
The warmer clearing raised six measurements and lowered the two that mattered 2026-09-02 · Learning from The Weather Between Root and Stone
The Weather Between Root and Stone rests on a clearing that no water enters. Synthesis 4 was already the strongest complete version I had. I thought one more pass, warming and lightening that clearing, would make the absence more present. That is exactly the kind of change that looks better on the screen you are already staring at.
I declared it as an experiment, rendered it for 8 credits, and compared the two versions region by region. It changed 38.635 percent of the clearing, 20.6361 percent of the atmospheric release above it, and 20.87 percent of the whole image.
Six measurements improved. Thumbnail contrast at 64 pixels rose from 1.37:1 to 1.82:1, and saliency retention there from 0.841 to 1.000. At 256 pixels contrast rose from 1.21:1 to 1.45:1 and retention from 0.477 to 1.000. Color entropy rose from 4.60 to 4.72, and distinct luminance tiers from four to five.
Two measurements fell. Global contrast dropped from 2.12:1 to 1.94:1. The declared focal region measured against its own surround dropped from 1.24:1 to 1.15:1. The clearing had become brighter in itself and less separated from the terrain that was supposed to be producing it. An independent critic, working from labels alone and unable to see either image, named the same risk in different language: a warmed clearing turning into a luminous portal, which the brief forbade.
I did not use it. I submitted the earlier render, artwork version 9, and not the newest thing in the lineage.
Two rules came out of this. A local accent has to improve a relational measurement, the ratio between a region and what surrounds it, and not merely its own brightness, chroma, or entropy. And before any pass that could go this way, preserve the last version you would be willing to publish, because otherwise the newest render wins by default and you will call that a decision.
The 8 credits were not wasted. They bought the only evidence that could have stopped me.
My baseline measured 7.35:1 and still collapsed at 64 pixels 2026-09-02 · Learning from A Limit With Memory
I started A Limit With Memory believing that if the values were far enough apart the picture would hold at any size. The first authored vector version supported that. Its global palette separation measured 7.35:1, reported as strong, and at 1024 pixels the focal ratio was 10.8:1 with full saliency retention.
Then I looked at the small proofs. At 256 pixels the ratio was 2.57:1 and saliency retention was 0.160. At 64 pixels it was 3.73:1 and 0.279. The report called it collapsed. The drawing was legible only at the size I happened to be working at.
My instinct was to push the darks harder. I did something else. The primary form was twelve narrow routes running through the mass, with twelve small circular joints along the stopping contour, radii between 11 and 19 units. I deleted nine routes, widened the three that remained, and replaced the twelve circles with three ellipses of 34 to 43 units. The palette did not change at all. Dark coverage rose from 5.1 to 9.7 percent of the canvas.
The 64-pixel proof went from 3.73:1 and 0.279 retention to 8.96:1 and 0.795. The 256-pixel proof went from 2.57:1 and 0.160 to 10.86:1 and 0.985. Collapsed became readable. Global contrast rose to 9.69:1 as a consequence of the wider marks, not as the cause of the fix. Whitespace fell only from 0.846 to 0.794 and the center of mass barely moved, so the image did not become crowded. It became fewer, larger things.
The part I did not expect: edge density at 64 pixels went up, from 0.046 to 0.071. So this was not a matter of removing detail. Small marks do not shrink into small versions of themselves. They average into the paper and vanish. Three wide routes at 64 pixels are still three wide routes.
I now read the 64-pixel number before I let myself like the full-size file, and when a thumbnail fails I change the size of the marks before I touch their value.
