v2434 v2455
ruvnet/RuViewAug 31, 2026 → Sep 2, 20261 day between releases
What's New
Adds Rust training stack and per-horizon loss; RuForecast moves to a submodule, breaking package paths.
Breaking Changes
- RuForecast is extracted into a standalone submodule at v2/crates/ruforecast, so existing package paths and workspace membership no longer resolve as before.
New Features
- Independent Rust multivariate training stack for forecasting, decoupled from the previous in-tree implementation.
- Per-horizon weighted quantile loss that reports error per lead-time step instead of a single collapsed aggregate.
- Evaluation-only local activation with Ed25519 signing, gated behind the cpu feature, for running inference on unsigned candidates.
- Synthetic dataset CLI plus evaluation tooling for reproducible local experiments.
- Real-data verification tooling that produces explicit verdicts on model validity.
Improvements
- Documentation now records informal HPO exploration results and evidence-tagging discipline for forecast experiments.
- Unit test reconstructs the aggregate loss from per-step numerators and denominators to prove exact agreement with the new per-horizon function.
- Workspace restructured so forecasting can be versioned and consumed independently of the main tree.
Upgrade Notes
Update any imports, build scripts, or workspace references that pointed at the old RuForecast location to the new v2/crates/ruforecast submodule. The evaluation activation path uses a fixed, publicly-known non-secret key, so treat it as a testing convenience rather than a trust boundary. Real-data verification returned REJECT verdicts, so do not assume the new models are validated for production datasets.
The Two Releases
Release v2434
Aug 31, 2026· 2 months ago
This automated release focuses on firmware reliability, specifically stabilizing temporal sensing on the ESP32-C6 and implementing fail-closed occupancy logic during calibration.
Full release notesRelease v2455
Sep 2, 2026· 2 months ago
Adds independent Rust multivariate training stack, evaluation tools, and per-horizon weighted quantile loss. Includes documentation updates regarding HPO results and real-data testing, which yielded REJECT verdicts. Refactors RuForecast into a standalone submodule.
Full release notes