Research

A 7B-parameter code model fine-tuned via SFT, GRPO, DPO, and Self-Play. MBPP+ 73.3% — the clean, uncontaminated benchmark. Full disclosure of HumanEval training-data overlap and evaluation methodology.

A 3B-parameter model specialized for Prisma ORM schema and client-code generation, fine-tuned via SFT, DPO, and GRPO on a single free-tier T4 GPU. 0.9992 pass@1 on the 1,200-task PrismaBench, with a documented and source-verified limitation in the static evaluation gate disclosed alongside the results.

Gating writes before they enter memory using source provenance, evidence corroboration, and bi-temporal conflict detection. 69.9% → 98.2% accuracy on a 113-case evaluation set across four configurations. Inter-annotator agreement and full failure-mode transparency.