Uses / Direct Use / Out-of-Scope Use, per the card template

README.md

@@ -222,7 +222,7 @@ the block above mirrors it, kept here so the card stands alone.
 population it stands for. Under the 1:2 design ratio a predict-all-illicit guess scores about
 50% F1 unweighted and 0.25% weighted, which is the whole distance between the two frames.
 
-## Contents
+## Files and licences
 
 `load_dataset` pulls only `data/`, which is 31.7 MiB. A `git clone` pulls everything,
 which is 1.11 GiB, almost all of it supervised model weights. Six file families were missing
@@ -341,11 +341,15 @@ weight that maps a metric computed here onto the full split.
 
 ## Uses
 
-**Direct use.** Scoring a detector on a transaction-graph AML benchmark whose weighted
+### Direct Use
+
+Scoring a detector on a transaction-graph AML benchmark whose weighted
 metrics stand in for a million-edge split; measuring where a prompt change acts, using the
 four-step rubric rather than a single F1 number; and building selective ML-and-LLM routers
 against a fixed evaluation design. It is a research benchmark and a measurement instrument.
 
+### Out-of-Scope Use
+
 **Not for deployment, and the paper's own numbers are the argument.** Restoring the native
 prevalence by Horvitz-Thompson weighting puts every one of the fourteen evaluated LLM cells on
 the trivial predict-all baseline: F1 between 0.23 and 0.25 on HI-Small against a floor of 0.25,