Four studies. Six deciders. 2.3 million decisions.
Each one tests whether a decider's past decisions can predict its future ones, on cases the model never saw. Five subjects are public strangers. The sixth is one private operator, on his own work. Every test is blind. One returned nothing, and that is published too.
Case Study Zero
The first subject who is not a stranger. A host shoots the same scripted line four or five times and keeps one, and the rival takes are word-for-word identical. Nothing inside a take predicts the choice. The taste sits in how a take stands against its rivals.
Predicting Judgment from Behavioral Exhaust
Reasons pulled from each decider's own published record, ranked twice: how often a reason is raised, and how much it moves the decision. The two rankings disagree. The second one predicts.
Judgment Without Text
No language at all, and a target written into the rulebook that all 124 deciders aim at identically. A personal signature survives anyway, and it lives almost entirely in the contested calls.
The Refusal Engine
One method, run identically against three unrelated deciders. A refusal is a stack of separable reasons, and the odds climb as they accumulate. On the third decider the signal was not in the public record, and the method returned nothing.