When Better Means Less: Quantifying What Benchmarks Miss Between Model Generations. 2,310 controlled comparisons show GPT-5 series lost 6.7x creativity and gained 4.4x false refusals vs chatgpt-4o-latest — invisible to standard benchmarks.

Research Knowledge Acquisition Engine — 5 autonomous campaigns (literature-survey, patent-mining, benchmark-archaeology, meta-analysis, baseline-establishment) with quantitative budget enforcement

llmverify is a lightweight Python tool for externally auditing LLM APIs and verifying whether custom providers and resellers truly serve the frontier models they claim to run, using layered probes and benchmark-based statistics.

How Hard is Microservice Anomaly Detection?