ReferenceAI skillsInterpreting metricsReferences

Numbers and evidence levels

Explain evidence levels without confusing structural confidence, comparison error, candidate scores, and experimental measurements.

Skill version: 1.1.0. Resource: skill://nadoobio-metric-interpretation/references/evidence.md.

Separate predictions, model confidence, agreement with references, actual binding, biological function, and clinical effects. High scores at an earlier stage do not establish success at a later stage.

Percentages depend on denominators: success per target differs from success per generated candidate. Report error exclusions alongside the original sample size. Keep the performance of candidates chosen after seeing the answer separate from candidates chosen by an actual selection rule.

Even identically named metrics are not directly comparable if regions, alignment, normalization, or units differ. Leave missing units or definitions unconfirmed and consult current documentation and result schemas. Do not convert confidence into success probability without justification.

Report in this order: inputs and conditions → returned values and units → valid comparison scope → conclusions requiring further experiments.

Items to record before calculating

In the input table, check target and candidate identifiers, measurements, units, success criteria, missing/failure states and selection rules. In the output summary, record numerator, denominator, exclusions, the scope where criteria were applied, and whether data are hypothetical, computed or experimental. Do not replace unavailable failure values with 0.

Hypothetical teaching example: with 20 candidates for each of 10 targets and one success per target, target-level success is 10/10=100%, while candidate-level success is 10/200=5%. If 20 of 200 candidates failed processing and 10 of the remainder succeeded, report 10/200=5% over all requests separately from 10/180≈5.56% over the 180 evaluable candidates. Do not hide the 20 processing failures or relabel them as confirmed biological failures.

Tabulate the best candidate chosen after seeing the reference structure separately from candidates selected by a prespecified rule without seeing it. Improvement only in the best value after increasing candidate count does not establish better per-candidate success. If candidate-level raw data are unavailable, report that the denominator and selection bias cannot be verified.

These calculations are arithmetic under the stated definitions, not actual research performance. Evaluation definitions comparing contacts and arrangements with reference structures exemplify the basis of computational metrics; they do not define target-level success, candidate-level success or clinical effects.

On this page