benchmark — independent software & tools
-
agent-eval-harness
Live, open-source benchmark for comparing AI coding agents on real GitHub issues
-
sophia-agi
Wisdom before intelligence. A provenance-aware reasoning layer that abstains instead of fabricating — stops LLMs inventing attributions and merging distinct traditions. Measured: 0% fabrication on unknown-answer traps (vs 17–25% raw); −12.5pt hallucinated attributions at 0% false-positive cost. An AGI-candidate proof package, not proven AGI.
-
agent-egress-bench
Provide an open test corpus to assess and improve AI agent egress security through validated cases and automated workflows.
-
cosmostrix
Production-grade cinematic Matrix rain renderer for serious terminal environments.
-
llm-eval
Evaluate large language models with tools for performance and consistency scoring as part of the llm-cpp C++ infrastructure toolkit.