Today a run compares one candidate against the original. Users selecting between two cheaper models want a
head-to-head. Add an A/B mode (e.g. tjb run --candidate A --candidate B) that runs both against the same
tasks and reports paired statistics for each versus the original (and against each other).
Acceptance criteria: two candidates run on identical tasks; the report shows paired deltas + McNemar for
each; verdict language stays honest; tests cover the offline path.
Today a run compares one candidate against the original. Users selecting between two cheaper models want a
head-to-head. Add an A/B mode (e.g.
tjb run --candidate A --candidate B) that runs both against the sametasks and reports paired statistics for each versus the original (and against each other).
Acceptance criteria: two candidates run on identical tasks; the report shows paired deltas + McNemar for
each; verdict language stays honest; tests cover the offline path.