Read the evidence, then decide.
Start with a benchmark board, then choose a test. AI Intelligence, Harness Rank, and Mem are peer boards. Each board has its own comparisons, studies, cases, and original evidence.
A board groups tests of one capability
AI Intelligence contains AIQ Core. Harness Rank studies model–harness combinations, beginning with Codex, Hermes, and Pi. Its first protocol is design only. Mem contains native memory, reader, and writer comparisons. A leaderboard belongs to a particular benchmark and its recorded conditions. New capability boards can have their own tests; scores do not combine across boards.
Start with the task
A model replay, native retrieval test, extraction replay, and complete Agent workflow measure different behavior. Compare subjects within the same recorded study and workload. Model version, provider, effort, output allowance, parser, embedding width, and system version can change the result.
Separate quality, time, and cost
Complete answers, literal quote bindings, and per-quote checks use different denominators. Read all three. A matching quote can still describe the wrong project or fail to support a claim. Missing and infrastructure-invalid results remain unavailable; they are not semantic zeros.
Settled USD is actual reported API spend for the stated calls. An estimate is labelled as an estimate. Neither includes local hardware, operator time, or subscription usage unless the study says so. A single non-streaming request time is not a stable provider speed ranking.
Open a case
Follow the comparison to a saved question, answer, citations, and returned context. Check whether a failure happened before retrieval, during extraction, or in the final answer. Keep a failed output even when later work repairs the problem.
Inspect the original files
Selected originals keep their exact bytes, SHA-256, and source links. A hash proves file identity, not scientific validity. Private archives and provider credentials stay outside the public library.
Mem compares OpenViking, RAGFlow, and fixed-context model replays. Each study has its own limits. They are not AIQ Core Calibration or Official publications. No universal AIQ score combines these studies.
AIQ Core has its own contract
The fixed task matrix, Rasch measurement, Calibration, and Official publication rules belong to the AIQ Core benchmark. Its Rust runner retains usage accounting, checkpoints, content-addressed packages, and independent evaluator replay. Those rules do not block clearly identified independent study pages.