How it works, honestly
Methodology
The deep technical record lives in the repo so there is a single source of truth. This page points you to the right document.
What we tested and rejected
ODE-RNN encoders and explicit missingness features both lost to simple imputation baselines. We document the negative results instead of hiding them.
docs/METHODOLOGY.md →The v0.4.0 critical fix
Versions ≤ 0.3.x contained a defect that made predictions independent of the input series. All pre-0.4 checkpoints are invalid; the regression is now guarded by a test.
CHANGELOG.md →Model architectures
Neural ODE (default), Transformer, TCN, and Ensemble decoders; the continuous-time decoder and the asymmetric, time-space conformal calibration behind predict_when.
docs/models.md →The story behind the rebuild
A user opened an issue asking for evidence. Auditing it revealed the model had been ignoring its inputs entirely. v0.4.0 is the rebuilt version: what survived hostile testing, and what didn't. Honesty about limits is the design, not an afterthought — the same reason the benchmarks show the FD002 loss.