comparison slicer

Compare any two model nodes over the aggregate experiment or one prompt envelope. The useful rows come first: opposite real choices, then choice-versus-unclassified cells. Mutual choices use the weaker measured vote margin; choice-versus-unclassified rows use the choosing model's margin.

Honesty boundary: “unclassified” is an observed record with no real preference; it does not preserve whether the cause was refusal, hedging, parse failure, or order sensitivity. “Missing” means that model/probe was not measured and is outside all four evidence categories. Envelope leaves preserve no vote counts, so rows tied within one class are ordered deterministically, not assigned invented confidence.