Measurement science for real enterprise problems.
We believe progress in AI depends on our ability to measure it well. Our work spans foundations, prediction, validity, incentives, and real-world impact — a scientific footing for enterprise evaluation and governance.
Predictive evaluation
Predict how a model will perform in your business, from as few observations as possible.
Ability prediction from sparse observations
Extrapolating overall performance from a handful of business samples: methods, limits, and failure conditions.
Cost–capability curves on enterprise tasks
Joint laws of capability, latency, and cost, grounding model selection under budget.
Validity and reliability
Does the benchmark measure what it claims to measure?
Failure modes of public benchmarks in enterprise selection
How leakage, saturation, and gaming drain public scores of their decision value.
A validity checklist for private benchmarks
An actionable validity review tool for in-house evaluation.
Incentives and governance
When evaluation results decide contracts and budgets, incentives start bending the measurement.
Evidence standards for vendor capability claims
What evidence is enough to support a procurement decision.
The leaderboard illusion in procurement
How well do public rankings correlate with in-house measurements?
Evaluation in the real world
Evaluation should not stop the night before launch.
Continuous evaluation in production
Closing the loop from pre-launch tests to post-launch behavioral monitoring.
Evaluation protocols for high-stakes domains
Safety, fairness, and robustness protocols for finance, healthcare, and government.
Enterprise AI Evaluation — the whitepaper
A systematic map of how AI evaluation fails in model selection, deployment, and compliance, with actionable method recommendations. A draft for comment ships in 2026 — real cases and co-creation partners welcome.
Measurement science is a community effort.
Many of these questions can only be answered by researchers and enterprises together. Bring us your scenario — join the co-creation, or just write.