Pinned Loading

1. judgecliff judgecliff Public

   Which judges survive optimization pressure? Exploitability of image-generation QC judges (rules, API VLMs, open VLMs, trained reward models) under BoN, DPO, and SRPO attacks.

   Python

2. llama-chess-uci llama-chess-uci Public

   Full fine-tune of Llama 3.1 8B to play legal UCI chess. 45 to 98 percent legal moves, median centipawn loss 255 to 55, Stockfish-verified.

   Python

3. research-agent research-agent Public

   Autonomous deep-research agent with a model-driven control loop and a cross-provider eval panel

   Python

4. soho-coop-rag soho-coop-rag Public

   RAG Q&A over every co-op building in NYC's SoHo historic district, from public data, with a hybrid SQL/vector router and a control-group evaluation.

   Python