---

name: evaluation description: "Evaluate and summarize Boltz prediction outputs, benchmark result tables, and legacy OpenStructure-based evaluation workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

Evaluation

Use this sub-skill when the task is to interpret Boltz output metrics, summarize prediction result files, diagnose benchmark/evaluation folder layout, or reason about legacy CASP/PDB evaluation workflows.

Start Here

• For local prediction-output summaries, use the bundled script: scripts/boltz_evaluation_summary.py.
• For benchmark reproduction planning, read references/evaluation-workflows.md before running anything external.
• For confidence, affinity, structural, ligand, and physical-validity metric meanings, read references/metrics-and-outputs.md.
• For missing dependencies, unavailable Boltz-2 evaluation assets, folder mismatches, metric mismatches, overlap caveats, and affinity interpretation traps, read references/troubleshooting.md.

Routing Boundaries

• Use this sub-skill for: evaluating Boltz outputs, aggregating evaluation CSVs, interpreting affinity/confidence metrics, CASP/PDB benchmark layout, OpenStructure requirements, and top-1 versus oracle summaries.
• Route raw boltz predict setup, YAML/FASTA input authoring, MSA server use, model checkpoints, and inference options to the prediction sub-skill.
• Route training/evaluation jobs launched through PyTorch Lightning or Hydra training configs to the training sub-skill.
• Route CCD, MSA, mmCIF preprocessing, dataset split construction, and benchmark input layout preparation to the data-preparation sub-skill.

Safe Defaults

• Treat Boltz-2 benchmark evaluation files, setup, and scripts as not yet published when using the documented repository state.
• Treat the available evaluation scripts as legacy Boltz-1 benchmark references unless newer Boltz-2 assets are supplied by the user.
• Do not claim a local confidence or affinity JSON summary is a reproduced structural benchmark; structural benchmark metrics need targets, aligned benchmark folders, and OpenStructure.