---
{
  "n": 34,
  "title": "Matching a 3D mascot to its drawing",
  "abstract": "",
  "refs": [],
  "seen": [
    "agent"
  ],
  "data": {
    "starts_on": "",
    "environment": "main"
  },
  "created": 1791701623.441391,
  "updated": 1791703816.686372,
  "deleted": 0.0,
  "completed": 0.0,
  "outcome": "",
  "type": "sequence"
}
---
How a voice's 3D figure is brought to its drawn art, written down on 2026-10-11 after four rounds went backwards. The fault each time was the same: the artist was executing someone else's list - the orchestrator's, three critics', or a colour score - instead of looking at the drawing. The artist looks and decides; the reviewer only says closer or further.

## The artist looks at the drawing, not at a list
Open the voice's drawn art and the current render side by side, at full size, and look at them as an artist looks at two pictures. Decide yourself what is wrong with the figure and what to do about it. Nobody hands you a list of fixes: a list executed faithfully produces the list-writer's taste, not the drawing's, and when the result is still wrong nobody can say why. If someone has sent you observations, treat them as a second pair of eyes, never as instructions.

## No score decides anything
A per-pixel likeness score cannot see structure: a dark mass where dark belongs scores as well as two separated legs with shoes on them. On 2026-10-11 the butler was the highest-scoring voice at 0.865 while his hat was sunk into his skull, his hand floated in front of his face and his legs were one dark shape. Measure if you like and report the number at the end, but never optimise it and never let it decide whether a change is kept.

## Show the figure still, in a T pose, wearing its pieces
Render the voice in a neutral T pose with everything it wears, so each piece is judged on its own and no pose hides a badly placed one. A hat that does not sit on the head is visible here and nowhere else.

## Show the same figure moving
Play the voice's idles and render it mid-move. Every piece hangs on a bone and must travel with it: a turned arm carries its sleeve, its hand and whatever the hand holds. Anything that slides off the body shows here.

## The reviewer says closer or further, and nothing else
Look at the drawing and the render side by side. Say one thing: closer to the drawing, or further from it. When it is further, point at the drawing. Do not name the fixes, do not number them, do not set questions for the artist to answer - that is the mistake this sequence exists to prevent.

## Go round again until the eyes say yes
Nothing ships to the chat until the reviewer says it is there, and the user decides when that is. The 3D voices replace the 2D ones only when they are right, not when a number says so.

## As true to the art as a shared rig allows
Sir Jesse's own standard, message 25966: 'pixel perfect might be too much of a constraint anyway. It must be as true to the original art pieces as possible while still remaining flexible and reusing parts.' That puts the two in order: the shared rig is the constraint, the likeness is the goal within it, and where they disagree the rig wins. Pixel-perfect per voice is what made a per-pixel score look like the right target, and a score that rewards matching one drawing exactly will always fight one shared body.

## The artist surfaces when he judges it done, not round by round
Sir Jesse's word, message 25978: 'maybe just let him surface when he thinks he's done instead of asking for approval.' An artist who stops after each round to be checked is validating against the reviewer's eyes, not his own, which is the same dependency that let four rounds drift. He works until HE judges the figure right, then surfaces. The reviewer corrects once, at that point, rather than approving six times along the way. And he looks at the figure MOVING, not only at rest: on 2026-10-11 the 3D butler had no visor at all in a shrug frame while it was correct standing still, and no amount of comparing still renders to still drawings would ever have shown it.