Kenneth ChengMathematical modeling

Editorial project overview

Intervention-Calibrated Neural ODEs

Do sign constraints improve both intervention prediction and uncertainty honesty?

Reproducible studyUpdated 28 Jul 2026
Technical figure from Intervention-Calibrated Neural ODEs
A reviewed research figure synchronized from the private technical workspace. The original aspect ratio and labels are preserved.

Editorial overview

The central result is deliberately negative: physical structure does not guarantee calibrated uncertainty. Read A Neural ODE Can Be Physically Plausible and Still Miscalibrated.

Key findings

  • Structured RMSE is 0.237 versus 0.259 for the unconstrained model.
  • Structured coverage is only 0.721 despite preventing negative predictions.

Limitations

  • Synthetic toggle switch only.
  • Ensemble spread does not cover missing-mechanism uncertainty.

Technical record

Detailed source, calculations, generated figures, and reproduction instructions are maintained in a private technical workspace. Public articles contain only manually reviewed interpretation and approved figures.

Version history

2026-07-28 — Curated overview reviewed against repository evidence.