Upload app.py with huggingface_hub

app.py

@@ -286,12 +286,49 @@ def _decide_impl(screenshot, goal, ax_tree, app, task_family, options_text, moda
     )
     progress(0.9, desc="Softmax over option letters...")
 
+    # The benchmark's decision rule (eval.scoring / train_4b_v2.group_log_probs)
+    # is PER ELEMENT: among each element's own candidate actions, take the
+    # argmax after re-normalizing over just that element's options. The global
+    # argmax over all letters is a different, stricter question, so both are
+    # reported.
+    by_element: dict[str, list[int]] = {}
+    for i, o in enumerate(options):
+        by_element.setdefault(o["label"], []).append(i)
+
+    element_lines = []
+    actionable = []
+    for label, idxs in by_element.items():
+        if len(idxs) > 1:
+            total = sum(probs[i] for i in idxs)
+            renorm = [(probs[i] / total if total > 0 else 0.0, i) for i in idxs]
+            renorm.sort(key=lambda t: -t[0])
+            p_top, i_top = renorm[0]
+            o = options[i_top]
+            element_lines.append(
+                f'• {o["role"]} "{label}" → {o["action"]}'
+                + (f" (with entity '{o['entity_id']}')" if o.get("entity_id") else "")
+                + f"  [{p_top:.1%} of this element's mass]"
+            )
+            if o["action"] != "skip":
+                actionable.append(element_lines[-1])
+        else:
+            i = idxs[0]
+            o = options[i]
+            element_lines.append(
+                f'• {o["role"]} "{label}" → {o["action"]} (only candidate)'
+            )
+
     best = max(range(len(options)), key=lambda i: probs[i])
     lines = [
-        f"Decision: {LETTERS[best]} — {options[best]['role']} \"{options[best]['label']}\" "
-        f"-> {options[best]['action']}"
-        + (f" (with entity '{options[best]['entity_id']}')" if options[best].get("entity_id") else ""),
-        f"Confidence: {probs[best]:.1%}",
+        "Per-element decisions (renormalized within each element's options — the "
+        "benchmark's decision rule):",
+        *element_lines,
+        "",
+        "Single best option overall (global argmax over the letter softmax):",
+        f'{LETTERS[best]}. {options[best]["role"]} "{options[best]["label"]}" '
+        f'-> {options[best]["action"]}'
+        + (f" (with entity '{options[best]['entity_id']}')" if options[best].get("entity_id") else "")
+        + f" — p={probs[best]:.3f}",
         "",
         "Ranked option probabilities (softmax over the option letters at the final position):",
         _format_options_text(options, probs),