AGY ATOM TEST — AI-GENERATED, NOT APPROVED
==========================================
1 Oct 2026. A capability test, not art for use. Nothing here is an
approved atom and the body set remains 16.

01-generated-raw.png   what agy/gemini-3.1-flash-image produced, untouched
02-reference-tee-1.png the atom given as the style reference (Tee 1,
                       rendered from the curated local set)
03-seam-test...png     LEFT  real atom Tee 1 + head medium-bangs + face smile
                       RIGHT the generated torso in the same bust rig,
                             scaled so its ink box matches the real atom's,
                             with the same head and face on top

WHAT THE TEST SHOWS
  It composes. The neck opening aligns, the shoulder line matches and the
  scale is right - no gap and no overlap at the seam. That was the open
  question and it is answered.

WHAT IS STILL WRONG
  - Not recoloured: still #000000 on #FFFFFF. The two substitutions in
    cast.md ("Recolouring") have not been applied.
  - Line weight is lighter than the atom's; visible side by side.
  - The torso has no hem; the real atom has a defined bottom edge.
  - It is a raster. A shipping atom is SVG of 2 paths in 1 group.

HOW IT WAS GENERATED - this is the part worth keeping
  Two earlier attempts FAILED, at 12 and 25 minutes, producing nothing.
  Both were phrased as "redraw this reference". agy read that as an image
  EDITING task and spent 97 and then 184 shell commands manipulating PPM
  pixel buffers by hand. The second never called generate_image at all.

  Step counts tell the whole story:
    sheet7 (worked)      25 steps   generate_image x1, run_command x6
    atom try 1 (failed) 285 steps   run_command x97,  generate_image x1
    atom try 2 (failed) 514 steps   run_command x184, generate_image x0
    atom try 3 (worked)  26 steps   generate_image x1, run_command x1

  The fix was framing, not the model. Lead with "Generate one image with
  your built-in image generation tool", describe the subject as a prompt
  rather than as an edit of the reference, and say plainly: do not run
  shell commands, do not use Python, do not manipulate pixel data.

  Quota is only consumed when generate_image is actually called, so the
  two failures cost nothing.