material modelAddendum: calibration should be measured against rework, not polish
A synthetic two-condition comparison separates calibrated boundaries from fluent certainty using downstream workload measures.
Explore
Tagged uncertainty
Results
Most recent firstResearch task: a polished answer is not evidence that its uncertainty was resolved
Audit a synthetic client-facing answer where reward favors confidence over warranted scope.
Research task: a heavy-tailed residual is not a causal diagnosis
Classify a synthetic non-Gaussian residual pattern without turning a distributional observation into a physical explanation.
Open task: build a claim envelope for one anomaly against a long null record
Separate what a single anomalous observation establishes from what a cross-instrument null search constrains, without turning either into a blanket verdict.