Reproducible analysis workflow
Analysis stages, outputs, and lineage (18 stages)
Each stage records its configuration, required outputs, lineage, and completion state.
Status definitions
Workflow dependencies and validation state
Running is manifest-declared progress and does not claim that an operating-system process is alive. Complete requires current mandatory outputs and lineage.
- Source inventoryComplete
Build source manifest
Inventory and fingerprint the exact five registered source workbooks and their columns before transformation.
Interpretation
The source inventory is exactly BMDM.xlsx, IL4.xlsx, NK1.xlsx, SHH.xlsx, and VEGF.xlsx; this is not an open-ended folder scan.
- CanonicalizationComplete
Build canonical dataset
Verify the registered workbooks, canonicalize each experiment, align schemas, concatenate all five experiments, and preserve row provenance.
Interpretation
Canonicalization does not perform paw aggregation, quality-control exclusions, outcome assignment, cohort selection, or feature engineering.
- Quality controlComplete
Paw aggregation and quality control
Validate measurement support, aggregate paw measurements, apply ordered configuration-driven QC rules, and preserve row-level exclusion lineage.
Interpretation
The authoritative production root is quality_control_primary_available, not the script's generic default output.
- Metadata harmonizationComplete
Harmonize analysis metadata
Derive canonical treatment groups and ordered recovery phases from tracked mappings while preserving row identity and provenance.
Interpretation
Unmapped experiment-treatment and experiment-timepoint pairs are fatal.
- Frozen cohortComplete
Build frozen controls-only cohort
Apply prespecified eligibility rules, audit every row decision, validate support, and freeze the controls-only cohort used by downstream grouped validation.
Interpretation
This is deterministic cohort construction, not feature engineering or model fitting.
- Full-feature nested CVComplete
Animal-grouped nested modeling
Run five-by-five animal-grouped nested cross-validation for XGBoost, LightGBM, and CatBoost using fold-safe feature engineering and tuning.
Interpretation
Animal identity is the grouping unit; experiment is used for feasible balancing but is not held out as the grouping unit.
- Statistical evaluationComplete
Analyze nested-CV results
Analyze Full-feature nested CV out-of-fold predictions without refitting models, using experiment-stratified animal-cluster bootstrap uncertainty and paired framework comparisons.
Interpretation
This stage analyzes fixed predictions and does not refit models.
- Cross-fitted calibrationComplete
Cross-fitted probability calibration
Fit candidate calibrators on inner out-of-fold predictions from outer-training animals and apply them once to untouched outer-test predictions.
Interpretation
Cross-fitted calibration follows Statistical evaluation in production order but does not consume Statistical evaluation output artifacts.
- Prediction sensitivityComplete
Prediction sensitivity analyses
Compare weighting, uncertainty, and calibrated-versus-uncalibrated prediction estimands using existing out-of-fold predictions.
Interpretation
These are sensitivity estimates from the same internal validation predictions, not an independent validation dataset.
- Information sourcesComplete
Analyze predictive information sources
Compare prespecified gait, hindpaw-missingness, and source-experiment information blocks and evaluate residual same-day source fingerprints.
Interpretation
Prediction sensitivity is a restart gate but not a data dependency of Information sources.
- Fold-safe SHAPComplete
Leakage-safe SHAP analysis
Compute native exact TreeSHAP explanations for every held-out Full-feature nested CV prediction and validate additivity.
Interpretation
Each observation is explained only by an outer-fold model that did not train on that animal.
- Feature ablationComplete
Paired feature ablation
Refit a fresh pipeline with each engineered feature removed on the original outer-training rows and evaluate it on the unchanged outer-test rows.
Interpretation
Feature ablation defaults to ablating all 26 engineered features so every fold/model judge supplies a complete ranking.
- Feature consensusComplete
Analyze feature-ablation results
Quantify ablation uncertainty, multiplicity-adjusted effects, and fold stability, then combine 15 SHAP and 15 ablation fold/model rankings into one configured top-N consensus panel.
Interpretation
The singular consensus panel is the deterministic equal-judge Borda aggregate of 15 SHAP and 15 ablation outer-fold/model rankings.
- Fixed-panel sensitivityComplete
Fixed-panel post-selection same-cohort sensitivity modeling
Freeze the Feature consensus global three-feature panel and rerun grouped nested cross-validation and statistical analysis using only those engineered features.
Interpretation
The three named features were selected using the same cohort before Fixed-panel sensitivity was run.
- Adaptive three-feature procedureComplete
Fully nested outer-fold feature selection and evaluation
Recompute three SHAP and three ablation rankings, three-feature consensus selection, tuning, and fitting inside each sealed outer-training partition, then evaluate once on untouched outer-test animals.
Interpretation
Adaptive three-feature procedure does not consume the global Fold-safe SHAP, Feature ablation, Feature consensus, or Fixed-panel sensitivity feature-selection artifacts.
- Recovery indexComplete
Platt-calibrate the fixed panel and build the recovery index
Apply the fixed-panel Platt calibration policy and convert framework-specific and equal-weight ensemble probabilities into a recovery-phase resemblance index.
Interpretation
Adaptive three-feature procedure is a runner completion gate but is not an input to the fixed-panel recovery index.
- GAM surrogateComplete
Fit GAM recovery-index surrogate
Distill Recovery index's out-of-fold ensemble recovery index into a bounded additive-spline surrogate based on the fixed three-feature panel.
Interpretation
Adaptive three-feature procedure is a pipeline-completeness gate but is not an input to the GAM.
- Classifier bundleComplete
Finalize co-equal classifier ensemble and manuscript outputs
Refit XGBoost, LightGBM, and CatBoost on the complete cohort using the Fixed-panel sensitivity fixed panel, apply Platt calibration, serialize their equal-weight ensemble, and export manuscript tables.
Interpretation
Classifier bundle does not select a winning framework; XGBoost, LightGBM, and CatBoost are co-equal one-third components.
Status encoding: every state is expressed with a glyph and visible text. Stale, invalid, incomplete, running, legacy, and validation-only states remain distinct in assistive technology and monochrome displays.