Six self-contained research notebooks, each written to be readable without prior background in its field, and a formal paper written from each one.
| Project | Question | Headline result |
|---|---|---|
Heart Disease Cross Hospital Research.ipynb |
Does a clinical risk model survive a move to another hospital? | 0.910 internal AUC; directed transfers range 0.571–0.888, so a single "generalization gap" hides the structure |
bankruptcy_crisis_stress_test/ |
Does a bankruptcy model trained in calm years survive a recession? | A random forest improves under the 2008–09 crisis (0.755 → 0.798 AUC) while linear and neural models slip |
clip_forgetting_forecasting/ |
Can we forecast how much CLIP will forget, early in fine-tuning? | A 5-epoch warm-up forecasts final forgetting at R² = 0.92 — but a zero-parameter persistence rule already reaches 0.85 |
exoplanet-model-comparison/ |
What is the exchange rate between accuracy and interpretability? | Transparency costs 0.029 AUC on detection and 85% more error on sizing; a label-free physics-informed model lands in between |
raddose-phytodosimetry/ |
Can gene expression identify radiation exposure in an unseen study? | 84.2% accuracy and 0.883 AUC under leave-one-study-out, after rejecting exact-dose regression on design grounds |
eeg-controls-audit/ |
Are the perturbation controls used to validate EEG decoders themselves reliable? | False-alarm rates span 1.3% to 68.8% across decoders — a control's trustworthiness depends on what you point it at |
papers/ holds one single-column, IEEE-style manuscript per notebook, with LaTeX
source, a figure-generation script, the compiled PDF, and a source zip. See
papers/README.md for the build instructions and for which papers
recompute their numbers versus transcribe them from an executed notebook.
cd papers && ./build.shEach project directory carries its own README.md and requirements.txt. Two projects
need data that is licensed and therefore not committed:
bankruptcy_crisis_stress_test/needsmirror/american_bankruptcy.csv.raddose-phytodosimetry/ships its harmonized matrix and runs as-is.
The rest are self-contained. Every notebook is committed with outputs rendered, so all of them can be read without being run.
- Preprocessing, feature selection, and hyperparameter choice are fitted strictly inside the training boundary, and the notebooks say where that boundary is.
- The independent unit of evaluation is stated explicitly — a hospital, a dataset, a study, a subject — and uncertainty is quantified at that level, not at the sample level.
- Every project separates definitions, observations, results, and interpretations, and ends with a claim boundary describing what the evidence does not support.