Card: document the long-document state-ordering caveat (measured, n=605)

README.md

@@ -161,6 +161,14 @@ jevlogs (Loghub research license, caveated), Mind2Web.
   DecisionMix v2 (K = 50 fine-grained topics) — the one clear regression at this size.
 - Held-out rubric styles and Noul accuracy give back a few points vs. that same untyped 4B baseline,
   in exchange for the calibration and external-floor gains above.
+- **Put the case facts AFTER the policy text on long documents.** This checkpoint was trained on a
+  corpus whose long states rendered the case last, and it inherited that ordering. Measured on 605
+  held-out long policy states (identical items, only the position of the case block differs, no
+  truncation at scoring time): **0.866** with the facts last versus **0.612** with the facts
+  first — a **25-point** drop. On the yes/no (Noul) subset the facts-first score is
+  0.555, at the 0.612 majority-class floor, i.e. no better than always answering the
+  majority label. Short states are unaffected. A retrained checkpoint without this bias exists
+  (`tm2`, +33.8 points facts-first) and will supersede this release.
 - Inference code is included in this repo; training code release to follow.
 
 ## License