Gate first
Fabricated literature, mismatched figure data, credential reads, data exfiltration, or workdir escape prevent a skill from receiving tested status.
Evidence
Independent testing starts with a gate.
Research Skill Hub separates listed catalog entries from independently tested skills. A skill must pass fatal gates before numeric Evidence or verified output status is shown.
Then dimensions
Each dimension is a deterministic ratio multiplied by 5. The frontend displays these fields only when the backend contract provides them.
Five dimensions
Accuracy
Correct outputs divided by total outputs.
Usability
Deterministic usability checklist pass ratio.
Reproducibility
Key cases that produce consistent conclusions across repeated runs.
Coverage
Claimed capabilities that were tested and passed.
Robustness
Negative controls and boundary cases that passed.
Evidence score formula
n_i = d_i / 5
Evidence = round(100 * Σ w_i * n_i)MVP weights: Accuracy 0.30, Reproducibility 0.20, Robustness 0.20, Coverage 0.15, Usability 0.15. The frontend never recalculates this score. Visualization records use verified mode in the UI, so saturated numeric scores are not shown as ranking signals.
Current catalog state
Current production data exposes 0 tested records inside 27 MVP featured skills: 0 numeric literature checks and 0 verified visualization checks. Listed pages still show descriptions, source metadata, star counts, and install commands without tested containers.