Coding
Implementation, repair, review, refactoring, and executable patch validation.
Practical evaluation · versioned contracts
Benchmarks built to fail honestly.
A request that never reaches inference is not a model failure. A fast answer is not automatically a good one. Every layer is measured separately.
Infrastructure ≠ qualityTransport and parsing failures stay separate
Unsupported ≠ zeroMissing capabilities are reported as N/A
Small samples stay labeledNo confidence theater
Evaluation surface
Real work, bounded claims.
Deterministic checks carry the correctness signal where possible. Qualitative outputs and raw JSONL remain available for human review.
Grounding
Natural and guardrailed hallucination resistance against fixed local evidence.
Repository
Navigation, bounded changes, tests, and appropriate no-change decisions.
Vision
Synthetic diagrams, camera scenes, photographs, charts, and mixed inputs.
RAG
Answerable and unanswerable retrieval at approximately 2K, 8K, and 16K.
Sports
Fictional prediction, recommendation gates, arithmetic, and ledger audits.