A graph memory for Monty, built from a STEP part.

Monty learns an object by sensing it: a patch of surface here, another there. Its pretrained memories come from episodes — an agent explores an object and stores what it perceived — which works, and is slow, uneven (the agent explores what it happened to touch), and cannot say why a region is called what it is called.

A CAD part has been explored already. This package samples its exact B-rep at a chosen spacing, takes the percept Monty's own sensor modules would take at each visit, and writes the labelled graph memory a Monty run loads.

The three steps, and where each lives

• MontyCad.Percept — a Quiddity.Frame becomes the observation a sensor module reads, and Monty's own ObservationProcessor answers the message. Nothing here re-implements a percept.
• MontyCad.Training — visits over the surface, one percept each, with the labels quiddity's recognition gives: an exact face and feature per node, concluded from nothing.
• MontyCad.Library — the file a run loads: gzipped JSON lines, one record per node, in the shape GraphLM reads.

{:ok, part} = Quiddity.open("bracket.step")

{:ok, written} = MontyCad.Training.write(part, "bracket", "bracket.jsonl.gz",
  spacing: 2.0
)

written.nodes
#=> 186

The same job from a shell:

mix monty_cad.train bracket.step --object bracket --out bracket.jsonl.gz

What is faithful, and what is not

docs/bridge.md states both, and the shortest version is: the percept is Monty's own code, the labels are exact, and the geometry is a drawing rather than a measurement — a CAD frame has no sensor noise, no occlusion from a real probe, and no colour.