Say what a score means, so a user can tell whether theirs is good

README.md

@@ -126,6 +126,24 @@ needs the repository as well.
 The weight column sums to the full split size by construction. That is the property the dataset
 exists for, and it is asserted at build time and again on every load.
 
+## What a score means here
+
+So you can tell whether your number is any good. All HT-weighted, so all describing the full
+split rather than these few thousand rows.
+
+| | HI-Small | LI-Small |
+|---|---:|---:|
+| Supervised ensemble, detection F1 | 70.9 | 29.6 |
+| Seven open-weight LLMs | 40.7 to 50.7 | 40.6 to 50.7 |
+| Their typology macro-F1 | 5.6 to 18.4 | 2.5 to 13.0 |
+| Best ML and LLM hybrid | 87.4 | 84.9 |
+
+The LLM band is the finding, not a weak result to beat: under the 1:2 design ratio a
+predict-all-illicit guess scores about 50, so every model in that band is at the floor. Recall
+reaches 100% and precision sits near a third. Per-model tables, the prompting variants and the
+hybrid's construction are in the
+[repository](https://github.com/nat-nischw/AMLCompact#what-you-are-up-against).
+
 ## Quickstart
 
 ```python