Benchmark explorer

A design passes only if it meets every criterion below.Common AF3 evaluation · not the model’s own acceptance filter

Unclustered = passing designs ÷ 200 attempts.

Clustered = distinct passing folds ÷ 200 attempts.

Dots show the mean across targets. Failures count; no replacement attempts.

Example & exact metric definitions

Illustration: 20 passing designs forming 8 structural clusters out of 200 attempts give a 10% unclustered rate and a 4% clustered rate. Clustering happens separately within each model, protocol and target (Foldseek TM threshold 0.5).

Comparing pass rates to design throughput

Click a dot for its protocol

Loading reviewed results…

● Generative■ Hallucination▲ Refinement◇ Search overlay

Compare methods

All workflows are compared together.

Bar = selected pass rate. Select a row for counts, timings and target-level evidence.

Compare methods on each target

Same model colors as the scatterplot. Select a bar for its protocol and exact counts.

Models shown in target chart

Benchmark notes