Metrics

Metrics do not all measure the same object, and they do not only run at the end of generation. This page covers how to read the metrics system as a whole: what a bad number means depending on which family produced it, how binder and target selections resolve, and where in a run a metric can stop further work.

The metrics guide covers turning metrics on and running them, inline or offline; the metrics reference lists every metric, oracle, pass criterion, and filter that ships, with what each one outputs and the tag that turns it on.

What each metric is looking at

Metrics do not all measure the same object. Four different things get measured, and knowing which one a metric looks at tells you what a bad number means.

Family	Looks at	Answers	Examples
Design	the design alone	Is this a physically sensible structure?	clashes, Ramachandran
Adherence	design against your input	Did it build what you asked for?	motif RMSD, sequence adherence
Refolding	the refolded structure alone	Does a folding model believe it?	pLDDT, PAE, pTM
Cross comparison	input, design and refold together	Does the whole story hold up?	insertion RMSD, chirality

A bad design metric means the model built something strained. A bad adherence metric means it ignored part of your specification. Those need different fixes, which is why the split matters.

For the refolding family specifically - the oracle, the AF3/OF3 setup it needs, and how a design can carry more than one refold — see the refolding guide.

Binder and target selections

Binder RMSDs use is_chain_of_interest; target alignment uses conditioned atoms outside that chain. Ambiguous inference requires an explicit chain_iid or chain_id query. The backbone-aligned variant aligns target backbone atoms but still measures all binder atoms.

Metrics can stop a run early

Metrics are not only scored at the end. Three points in a run read them and decide what proceeds, and that is where most of the compute savings live.

Gate	Sits	Decides
Pre-oracle filters	before refolding	which designs are worth an oracle call at all
Pass criteria	after refolding	what counts as a success, judging design, oracle and refold together
Search pruning	inside the trajectory	which candidates survive to the next checkpoint

Pre-oracle filters save refolding compute by skipping designs that already fail a criterion. Filtered designs remain in ordinary and clustered pass-rate denominators, so filters can stay enabled for benchmarks. Report the filters used; disable them only when you want to refold every design.