Sophia — the Wisdom Gate
Provenance-aware, verifier-gated reasoning layer for LLMs. End-to-end training, inference, and evaluation infrastructure with bilingual benchmarks and CI-gated measurement.
- Attribution hallucination 36.1% → 23.6% on local 8B (0% FP cost)
- 528 bilingual EN/ZH evaluation examples (HuggingFace)
- Defensive publication — Zenodo DOI 10.5281/zenodo.20930874
- SimpleQA calibration uplift at 20% coverage (cross-model)