# How we checked the work

We checked our reading against a hand-labeled set, using a method that does not let the classifier grade itself on its own examples.

**The hand-labeled set**

We hand-labeled 43 markets drawn from the five boundary groups: clean measurements, direct attacks, ceasefire and peace deals, leader and regime survival, and territory and facility control. These served as the answer key.

**Leave one group out**

For each group in turn, we set its markets aside, gave the reading only the examples from the other groups, and asked it to read the held-out group cold. This tests whether the reading learned the rule rather than memorizing the examples it had already seen.

**Result**

The reading agreed with the hand labels on which markets the rule reaches 91 percent of the time. Agreement was 86 percent on the narrow reading and 95 percent on the kind of activity involved. It was near complete on the clear cases, the clean measurements and the direct attacks, and the disagreements clustered in the same boundary categories the narrow-to-broad band already flags. In other words, the reading is most reliable exactly where the stakes are clearest, and least settled exactly where we already report a range instead of a point.

**Limits**

We read each market from its public resolution text rather than the on-chain oracle data that can be more precise, so a small number of genuinely ambiguous markets would benefit from a closer human look. For that reason the entire territory and facility group is flagged for human review before any number from it is used in public.

This is a reading of the proposed rule. It is not legal advice. Close calls are ultimately the Commission's.