NOTES · 2
Most group processes treat a split as a result. Half the room says yes, half says no, the motion fails, and the question disappears from the record as if it had been answered. For a causal question that is misleading: the room did not decide that A has no effect on B. It did not decide at all.
True Causes treats a split differently. When a panel deadlocks on whether one cause significantly influences another, the question does not fail quietly. It stays open and goes elsewhere to be settled. If a second panel splits too, the pair is published as contested and stays visible on the map. The disagreement is part of the result.
Asking the same people again rarely changes the answer. So a split question leaves the group that split it and goes to two other whole groups, chosen so that it cannot be steered toward a friendly audience. They answer separately, and they must agree. Question assignment also rotates over time, so that no group owns a region of the map.
Both rules were added because a simulated adversary broke the version without them. With escalation, the method makes about 3 wrong links in every ten thousand decided questions, even when each person answers 30% of questions wrongly. With referral, a room with three colluding insiders goes from recovering 0.43 of the real drivers to 0.98.
The principle is general. Any process that turns “we could not agree” into “no” lets a small group remove a cause simply by splitting the vote. Keeping the dispute on the record, and sending it to people who were not in the room, prevents that.