General knowledge and reasoning
Use task-level controls and contamination checks before cross-model comparisons.
Benchmarks
Evergreen benchmark framing for capability, retrieval, and reliability evaluation.
No rankings without reproducible methodology + primary sources.
Retrieval and long context
Evaluate retrieval quality separately from generation quality, then test long-context stress behavior.
Tool use and software reliability
Tool-calling claims require deterministic execution logs, test harness disclosure, and replay checks.
SWE-bench ownerSWE-bench paperOwner and method provenance are documented in the linked primary papers.