Analysis-code supplement — input data are not redistributed

Scope: current source scripts for conventional models, source calibration, paired comparisons, representation analysis and aggregate metric verification. These files were reviewed against saved results. This is not a standalone raw-data-to-paper executable package. The release does not prove every historical executed version and does not include data acquisition, case-level inference outputs, model weights or credentials.

Run only in an isolated working copy. Scripts use work/ input paths and outputs/ result paths relative to that copy and can overwrite aggregate outputs. They must not be run against the authors' frozen result directory.

Required authorized local inputs:
work/pangi/expanded_cohort_candidate.csv: eligible VitalDB records, including caseid, subjectid, event14, anestart, aneend and common fields.
work/pangi/folds.csv: caseid, subjectid, fixed patient-disjoint fold assignment.
work/mover/candidate_patientlinked_features_NOT_FROZEN.csv: external harmonized features, LOG_ID, corrected_MRN, outcome and common fields. The historical filename is retained; analysis inputs were hash-checked subsequently.
work/mover/candidate_cohort_NOT_FROZEN.csv and sensitivity flag files where requested by scripts.
work/common_formal/prompts.jsonl and representation_prompts.jsonl: local keyed prompts generated according to the supplied generic specification. Do not publish patient-containing prompts.
work/common_formal/{qwen14,gemma12}_pilot_responses.jsonl: baseline JSON responses, schema_ok, finish_reason, probability and keyed operation identifier. Despite historical filename, these are the formal 3700-case outputs.
work/common_formal/{qwen14,gemma12}_representation_responses.jsonl: 11100 variant outputs per model, with caseid, variant and seed.

Suggested dependency order after authorized harmonization and inference:
1. common_oof.py and run_common_classical_external.py; run_common_boosting.py.
2. calibrate_common_classical.py; analyze_common_llm.py.
3. compare_common_uncertainty.py; common_calibration_diagnostics.py; common_subgroups.py.
4. analyze_representation.py; representation_uncertainty.py; verify_final_predictions.py.
Inspect each script's Path/read_csv/read_json calls for its exact intermediate-file dependencies. Data acquisition, phenotype construction and inference must be completed separately. The dictionaries and generic prompt document those specifications but do not substitute for the required licensed source data.

Environment snapshot: Python 3.14.7, NumPy 2.5.2, pandas 3.0.5, scikit-learn 1.9.0, SciPy 1.18.1. threadpoolctl is also required. This package was syntax-checked and checked against saved aggregate results; full authorized data/inference rerunning is not claimed in manuscript preparation.
Patient-group bootstrap: 1000 draws, seed 20261004, fixed predictions/maps, no multiplicity correction.
For privacy, personal workstation paths are not included in metadata. Script paths are project-relative. No data-access passwords are needed to read this supplement.
License: authors must approve a code-sharing license before publication. No license is inferred from the supplied unrelated manuscript.
