Popular repositories Loading

1. llm-jury llm-jury Public

   Training-free committee-of-6 LLM judge for RAG hallucination detection: AUROC 0.88 / F1 0.77 on RAGTruth, reproducible at $0 from frozen verdicts — and it surfaces provable label errors in the benc…

   Python

2. driftbet driftbet Public

   Label-free drift attribution (model / noise / world / annotator) with anytime-valid guarantees and a pre-registered change-ledger — closes the zero-label misattribution gap 0.50→0.00 on the recover…

   Python

3. whest-teardown whest-teardown Public

   Reproducible teardown of ARC WhiteBox Estimation 2026: plain Monte Carlo beats the analytic reference ~9×, plus a 4.4× FLOP-accounting arbitrage in the scoring harness.

   Python