@magnitudedev/cli@0.2.6 v0.3.29
feyninc/pulpieOct 5, 2026 → Oct 7, 20262 days between releases
What's New
Adds training capabilities like abstention threshold fitting and evidence inspection, alongside documentation clarifications on confidence formulas.
New Features
- Evidence inspection for checkpoint and eval artifacts
- Abstention threshold fitting in training
- train_report.json support
- Three-seed RLCD research panel
Improvements
- Documentation clarifications regarding confidence formulas and hook behaviors
- ONNX dependency cap to prevent NumPy 2 compatibility issues
Bug Fixes
- Fix agent reading state predict_long's questions in one pass
Upgrade Notes
Existing examples may need updates to match corrected confidence formulas and hook behaviors. New training features require familiarity with the --eval flag and train_report.json output.
The Two Releases
@magnitudedev/cli@0.2.6
Oct 5, 2026· 17 days ago
A comprehensive patch release addressing critical bugs across multiple platforms (Windows, Mac, Linux) and significantly improving performance on Apple Silicon devices.
Full release notesv0.3.29
Oct 7, 2026· 15 days ago
This release emphasizes documentation clarity regarding confidence formulas and hook behaviors. It introduces core training capabilities such as abstention threshold fitting and evidence inspection, alongside updates to the research panel for loss functions.
Full release notes