tda-diagnosing-computational-defects

What problem does it solve? Computational defects in research pipelines often pass every software test while producing mathematically wrong, statistically invalid, or provenance-incomplete results. This Skill provides a disciplined procedure to reproduce, isolate, and fix such defects without closing them prematurely. ## Core Features & Use Cases - Structured defect diagnosis: A 10-step procedure covering symptom statement, lane classification, red-capable reproduction commands, minimal fixtures, falsifiable hypotheses, and post-mortems. - Known failure-class catalog: Checklists for Monte Carlo p-value denominator errors, label-shuffle null mistakes, PCA/scaler refit drift, silent truncation, and cross-version pickle label collapse. - Use Case: A permutation test returns suspicious p-values. Use this Skill to classify the defect lane, build a failing reproduction command, check the known failure classes (e.g., wrong p-value denominator), fix only after reproduction, and add a regression test with a contract binding. ## Quick Start Ask the assistant to diagnose why a result JSON changed unexpectedly or why a null model gives suspicious p-values, following the defect diagnosis procedure.