card: add Adaptive Classifier specialist baselines (MiniLM and ModernBERT)

README.md

@@ -199,6 +199,40 @@ To be scored as a generalist, train on other workflows entirely and evaluate
 here zero-shot. Say which mode you used. A specialist number sitting next to a
 generalist number, unlabelled, misleads the reader.
 
+## Baseline results
+
+Two specialist baselines, fitted on the `train` split and scored on `test`. Both
+use [Adaptive Classifier](https://github.com/codelion/adaptive-classifier)
+0.2.0, one classifier per question, with the encoder frozen.
+
+| Model | Acc | Soft acc | Macro F1 | KL | TV | Brier | ECE | Score MAE | Within 1 level | ms/case |
+|---|---|---|---|---|---|---|---|---|---|---|
+| MiniLM-L6 (22M) | 0.587 | 0.506 | 0.424 | 0.262 | 0.267 | 0.143 | 0.108 | 0.515 | 0.864 | 22 |
+| ModernBERT-base (149M) | 0.646 | 0.542 | 0.469 | 0.223 | 0.249 | 0.119 | 0.179 | 0.444 | 0.931 | 349 |
+| Prior (input ignored) | 0.494 | 0.446 | - | 0.339 | 0.313 | 0.184 | 0.075 | - | - | 0 |
+| Uniform | 0.311 | 0.312 | - | 0.444 | 0.381 | 0.238 | 0.169 | - | - | 0 |
+
+Read these against the floor of 0.520 and the saturation point of 0.735.
+ModernBERT lands 12.6 points above the floor and 5.8 below what perfect
+scenario understanding would give.
+
+The gap between the two encoders is the trade-off this benchmark exists to
+measure. Six points of accuracy cost 16x the latency.
+
+Both are specialists. Neither can answer a question it was not fitted for, so
+neither can be scored zero-shot.
+
+Configuration, tuned on a held-out quarter of the `train` split and never on
+`test`: mean pooling, `max_length` 512, 30 epochs, `prototype_weight` 0.3.
+
+One harness detail matters for reproducing these. Adaptive Classifier trains on
+hard labels, so the gold distribution is normally thrown away at fit time. Each
+case is instead entered four times, apportioned across labels in proportion to
+its gold, which carries the soft target into a learner that cannot represent one
+directly. That single change cut KL by a third and score MAE by 15%, while
+barely moving accuracy. The argmax was already right. What improved was the
+shape of the predicted distribution, which is what this benchmark is for.
+
 ## Splits
 
 Two splits, generated independently. `test` is the benchmark. `train` comes from