fix(reward): harsh -1.0 for zero JSON output, -0.30 for zero task completion

training/rollout.py

@@ -351,18 +351,34 @@ def rollout_once(
         no_wrong_channels = 1.0
 
     read_runbook_reward = 1.0 if read_runbook_done else 0.0
-    combined = (
-        final_score       * 0.45
-        + no_wrong_channels * 0.15
-        + valid_json_ratio  * 0.15
-        + read_runbook_reward * 0.15
-        + efficiency        * 0.10
-    )
+
+    # --- Reward gating ---
+    # If model never output valid JSON, it never actually tried anything.
+    # Strip all process rewards and apply a harsh penalty.
+    # Process rewards (runbook, no_wrong, efficiency) only matter if model acted.
+    if valid_action_count == 0:
+        combined = -1.0
+    elif final_score == 0.0:
+        # Model tried (valid JSON) but task failed — small process credit, no final bonus
+        combined = (
+            valid_json_ratio    * 0.15
+            + read_runbook_reward * 0.10
+            - 0.30                        # hard penalty for zero task completion
+        )
+    else:
+        combined = (
+            final_score           * 0.45
+            + no_wrong_channels   * 0.15
+            + valid_json_ratio    * 0.15
+            + read_runbook_reward * 0.15
+            + efficiency          * 0.10
+        )
+
     print(
         f"[rollout] steps={step} final={final_score:.3f} "
         f"json={valid_json_ratio:.2f} runbook={read_runbook_reward:.0f} "
         f"no_wrong={no_wrong_channels:.2f} eff={efficiency:.2f} "
-        f"→ combined={combined:.3f}"
+        f"valid_acts={valid_action_count} → combined={combined:.3f}"
     )
     return {
         "prompt_ids": prompt_ids,