STEM-Bench
A closed-form reasoning test where every answer must be computed, not remembered — distractors are the results of plausible mistakes, so recall alone fails. Two hundred tasks across four difficulty tiers and five disciplines, keys re-derived by script before scoring.
Professional-Bench
Models produce real work products — an M&A memo, a strategy call, a legal review, a systems design — graded against strict expert rubrics, then are put to work in a live database with tools. Eleven models across five labs, judged on what they produce, not what they recall.
Agentic-Bench
Models are dropped into a messy production database they've never seen — amounts in cents, duplicate rows, soft-deleted accounts, mixed currencies — and must investigate with live SQL, then answer twelve finance questions. Graded objectively against ground truth, no LLM judge. The field spreads from 92% to 25%.
More benchmarks are on the way — including harder, longer-horizon agentic tasks and expert-authored domain evaluations. This list will grow.