The tests and evidence that separate a stronger model from a better-looking score.
Benchmarks · Reliability · Cost & performance
5 articles
EvaluationsA contracting benchmark connects model training to business rules. Its partial-credit results expose the distance between an improving agent and a workflow a company can trust.
EvaluationsGDP.xlsx tests whether models can reconstruct a workbook’s logic. Its implications reach back into training data, tool use, and how enterprise capability is measured.
EvaluationsFrontier models beat a small human baseline on structured accounting tasks. The business question is what remains outside the benchmark.
EvaluationsArtificial Analysis adds a hallucination gate to its legal-agent benchmark. The change exposes a model-development problem: covering every requested point and grounding every claim are different skills.
EvaluationsA model that declares success is not necessarily finished. Legacy-Bench shifts attention from code generation to the evidence needed to trust a repair.
Explore the other topics in this desk.
