v0.1.2

• Model: added the v4 segmentation model, candidate24 selected at step 8,000, with matching PyTorch and NumPy weights. The Python package uses v4 by default. The previous v3 model remains available in the repository and v0.1.0 release.
• Pitch recognition: keeps staff tracks from following nearby beams, correcting a group of pitch errors while preserving note timing.
• Training and integration: added portable fine-tuning and export tools. Training is optional. The Python wheel provides image/PDF-to-JSON/MIDI inference, and the JAR provides the standalone Java decoder.
• Validation: 1,096 Java tests, 19 Python tests, installed-package PNG/PDF pitch and duration checks, and scoped real-score regressions.

See the README for installation and integration. SHA256SUMS.txt lists checksums for the downloadable artifacts. Code and weights remain Apache-2.0 licensed, including use and modification in closed-source products under that license.

Recognition is experimental. Dim-paper development accuracy regressed, and missed or extra notes and rhythm errors remain. The new held-out model split had not been evaluated at this release. See models/evaluation.json for coverage and limitations.