The six cards are the top of this, not a selection from it.
A good analyst’s first objection to any dashboard is that you found the finding because you went looking for it. So this sweeps every explanatory field against every slice with the same relate() the cards use, and ranks what comes back.
Why the ranking is not just “lift”. An interaction is invisible to lift alone — passive cooling reads ×1.56 across the whole book.
What gives an interaction away is that the lift moves when you condition: ×1.81 inside hot sites, ×1.08 inside temperate ones. Each cell is scored by how far its lift travelled, weighted by information gain — the bar in the signal column.
The weighting is not decoration. Ranked on movement alone the top of this table was ticket_count=3 inside arctic at 0.82 — a thin slice letting a lift wander. Those noise cells carried info around 0.0002–0.003 while the planted interaction carried 0.056, so the engine’s own information measure separates them without a hand-tuned threshold, and without dropping the noisy fields — which would have been the same curation this page exists to avoid.