Benchmarks for epistemic reliability

Measure truth-seeking behavior on a bounded task with ground truth.

Which error matters to the eventual user?

Optimizing the benchmark can reward a proxy rather than understanding.

Construct a held-out task set and report calibration alongside accuracy.

• What’s Important in AI for Epistemics?

Interactive tools load with JavaScript. Original sources retain their language and attribution.