Edgeless Lab / Pen Plotter Autoresearch FILED 2026·04·07
RUN № 18 FACTORIES 40 PIECES 37,571 BEST 89.1 STATUS ACTIVE

Pen Plotter
Autoresearch.

Thirty‑seven thousand five hundred and seventy‑one specimens of generative line art, scored by an autonomous loop of forty procedural factories and five language‑model judges. A field journal of what the algorithms loved — and what they got wrong.

moire-0005
Plate I. moire · variant 0005 composite 89.1

Each cycle, an autoresearch loop audits every factory, identifies the single weakest metric across the catalog, edits the corresponding generator, regenerates a thousand new variants, scores them on six algorithmic axes and five vision‑model judges, and either commits the change or reverts.

The loop has been running since the third week of March 2026. Forty factories, thirty‑seven thousand five hundred and seventy‑one pieces — spanning color dot factories, botanical gardens after Zancan, watercolor polygon deformation after Tyler Hobbs, Fidenza‑style hatched softshapes, and Owen Jones Islamic geometric knotwork. Most of them are forgettable. A few hundred are extraordinary. And then there are the grey rectangles — pieces the algorithms loved and the human eye refused.

Scoring moved to v3 midway through Run 17. The new system applies Galanter's complexity theory — rewarding the complex middle of the entropy spectrum rather than raw density. Vision‑language judges now follow Feldman's four‑step art criticism protocol (describe, analyze, interpret, judge), and an intra‑factory uniqueness penalty prevents any single factory from concentrating the leaderboard.

Two pieces.
Same factory.
Same algorithmic score.

moire-0421
Plate II. moire · variant 0421
Algorithmic composite 88.3
Rank in catalog Top 5 of 35,849
Claude vision judge 88.3

“Reads as uniform fine dot/grid texture with no discernible composition, focal point, or visual interest. Appears to be a degenerate parameter combination.”

moire-0062
Plate III. moire · variant 0062
Algorithmic composite 87.0
Rank in catalog Top 12 of 35,849
Claude vision judge 87.0

“Multiple radial interference lobes create a striking starburst with exceptional depth and Op Art energy — reminiscent of Bridget Riley’s radial works. Gallery‑ready.”

The algorithm cannot tell the difference between a star burst and a piece of woven fabric. That is what the judges are for.

One thousand
two hundred and
ninety‑six
specimens.

A representative sample of the catalog, sorted by composite score. The first row is the strongest work the loop has produced; the last row is the weakest still kept. Each plate is annotated with a score bar running its composite from zero to one hundred. Hover any plate to magnify; click to inspect.

1,296 shown · 34,553 withheld ⚠ very slow · view all 35,849 plates

The loop produces.
The judges score.
The artist still picks the ones that get pen on paper.

§05  ·  Colophon & Continue

Set in Boska & JetBrains Mono.
Printed on aged paper.

Display
Boska · Indian Type Foundry, distributed via Fontshare
Body & data
JetBrains Mono · ligatured monospace, ofl licensed
Substrate
A3 Fabriano cotton for the physical edition · warm specimen paper for the screen edition
Plotter
iDraw 2.0 with archival pigment ink · paths optimized via vpype
Generators
40 procedural factories in Python · seed-driven, deterministic per variant
Scoring
6 algorithmic signals + 5 vision-language judges (Claude · Gemini · Qwen · Cerebras)
Loop runtime
marimo + Python 3.11 · audit, edit, regenerate, score, commit, repeat
Edition
Run №18 · April 2026 · 37,571 specimens
Designed by
D. Murray · Edgeless Lab

  D. Murray  ·  April 2026  ·  Edgeless Lab