What this shows
Backtested on an 80-patient synthetic cohort (assimilate 2 scans, forecast the rest).
A well-calibrated 90% interval should contain the truth about 90% of the time.
Prior-only and trained models were badly overconfident (~31–37% coverage);
adding horizon-growing process noise brings coverage to ~90% without hurting
accuracy. Reproduce on your machine with make analysis.
Real-data check · NSCLC-Radiomics
Tested against real patient survival (TCIA NSCLC-Radiomics, 399 patients):
the twin's risk scored a concordance index of
0.505 — essentially chance. An honest negative
result: the growth model was not trained on survival, and the prognostic signal
lives in imaging features (tumor volume, heterogeneity) not yet loaded per patient.
Real CT ingestion works; extracting tumor radiomics across the cohort is the next
step. Reproduce with python analysis/real_data_analysis.py.