Metallicity

Whether a material conducts. Not written into an input file — a fact about the material that changes the mesh density and the smearing choice together.

Quantity	What it is	Model
is_metal	True when the DFT band gap is zero	CGCNN
band_gap	The gap itself, in eV	none

The second is the same relationship k-distance has to a mesh: predict a continuous quantity, let the consumer apply the rule that turns it into a decision.

The other artifact under this setting

There are two published CGCNNs here, and only one of them answers anything.

	What you get	Can load_model serve it?
is_metal/cgcnn	metal or insulator	yes
representation/cgcnn	64 numbers	no

The second supplies one block of the k-distance model's feature vector and answers no question of its own. Its page explains what that means, why it cannot be served as a classifier, and which paper trained it.

The two are not two attempts at one task, and their numbers do not belong side by side: they were fitted on different datasets, under different splits, for different purposes.

Core's side

This package can now serve a metallicity decision through the inference seam. Core has nowhere to put the answer yet: it carries no metallicity field, and its own rule is a heuristic — smearing is applied only when every element is metallic, so metallic oxides get none. Tracked in stfc/goldilocks-core#175.