v0.1.0 — Trained bilingual visual decision model

Actual trained LoRA adapter, Clef decision head, processor, schema and calibration for Qwen/Qwen3.5-0.8B.

• Trained for 69.0 minutes / 6,000 steps on RTX 4070 Ti SUPER 16 GB; released checkpoint selected at step 2,000 using validation only.
• 18,048 bilingual training instructions. Held-out accuracies: 99.14% new handwriting, 97.87% new wording, 98.83% new combinations, 97.85% both new wording and combinations.
• Archive: 86,549,227 bytes. SHA-256: 77c601ff853da087598da8c86bcc93248452a450048dbd2719f37be4d7e677d7.
• Run python download_model.py from the source checkout to download and verify these weights and the pinned base model. The archive does not contain the base model.
• First Windows CUDA package + base + weight downloads total approximately 3.91 GB, excluding Python/Git installers and repository transfer.
• Actual CPU/CUDA inference verified. Physical Apple Silicon MPS remains unverified. A second fresh Python environment on the development computer was blocked by Windows App Control; see execution notes.
• Actual failure cases, probabilities, calibration, benchmark details and limitations are in RESULTS.md. NONE is not an uncertainty detector.

No original repository or Space was changed. No paid GPU, hosted inference or Git LFS was used.