Skip to main content

Module joint

Module joint 

Source
Expand description

Cross-question / joint coverage.

Conformal coverage guarantees are marginal: a 90% prediction set per question does not imply 90% of batch rows have every question covered. A decision batch answers many questions against the same state, and downstream consumers (confidence-filtered joins, SQL composition) compose the answers — so the joint number is reported separately from the marginals, alongside the product-of-marginals reference that independence would imply. This module owns that report; it takes per-row coverage indicators produced by any conformal method.

Structs§

JointCoverage
Joint-coverage report for a batch of multi-question rows.

Functions§

joint_coverage
Compute joint and marginal coverage from a row-major coverage matrix: covered[i][q] is true when row i’s prediction set for question q covered the realized label.