Sync from GitHub: 30c5708e27932c87aea32d9d6fc845c55c02d352

app.py

@@ -1181,6 +1181,17 @@ def on_interleaved_benchmark(
     _start_stats = nw_organism.get_stats()
     _start_nodes = _start_stats.get('nodes', 0)
     _start_synapses = _start_stats.get('synapses', 0)
+    # Run 41+ predictive-coding diagnostic — capture cumulative prediction
+    # counters at run start so we can compute delta per run. If predictions
+    # never generate, all per-turn predictions_confirmed/surprised are 0
+    # AND total_predictions_made delta = 0 → confirms prediction_threshold
+    # gating diagnosis. Reads canonical Graph counters directly.
+    _start_total_predictions_made = int(getattr(
+        getattr(nw_organism, "_graph", None), "_total_predictions_made", 0,
+    ) or 0)
+    _start_total_surprised = int(getattr(
+        getattr(nw_organism, "_graph", None), "_total_surprised", 0,
+    ) or 0)
 
     nw_kiss_inst = KISSFilter()
     bl_msgs: list = []
@@ -1478,6 +1489,16 @@ def on_interleaved_benchmark(
             "success_signal": success_signal,
             "pith_same_cat_ratio": _same_cat_ratio,
             "pith_self_retrievals": _self_retrievals,
+            # Run 41+ predictive-coding telemetry — surface what step_result
+            # already carries about predictions plus a snapshot of active
+            # predictions on the graph. If all 0/0/0 across all turns, the
+            # canonical predictive-coding loop is dormant (gated by
+            # prediction_threshold per audit task #12).
+            "predictions_confirmed": int(step_result.get("predictions_confirmed", 0) or 0),
+            "predictions_surprised": int(step_result.get("predictions_surprised", 0) or 0),
+            "active_predictions_count": len(getattr(
+                getattr(nw_organism, "_graph", None), "active_predictions", {}
+            ) or {}),
         })
 
     # ── Heatmap A: ignition-set Jaccard overlap (symmetric) ──
@@ -1634,6 +1655,19 @@ def on_interleaved_benchmark(
         "substrate_nodes_end":      _end_stats.get('nodes', 0),
         "substrate_synapses_start": _start_synapses,
         "substrate_synapses_end":   _end_stats.get('synapses', 0),
+        # Run 41+ predictive-coding diagnostic — cumulative counters from the
+        # canonical Graph. If `predictions_made_during_run = 0` even at
+        # benchmark scale, the predictive-coding loop is dormant (gated by
+        # prediction_threshold per audit task #12) and the surprise-driven
+        # intrinsic reward broadcast (canonical neuro_foundation:2549) never
+        # fires. This is the empirical confirmation gate before any config
+        # graduation work.
+        "predictions_made_during_run": int(getattr(
+            getattr(nw_organism, "_graph", None), "_total_predictions_made", 0,
+        ) or 0) - _start_total_predictions_made,
+        "predictions_surprised_during_run": int(getattr(
+            getattr(nw_organism, "_graph", None), "_total_surprised", 0,
+        ) or 0) - _start_total_surprised,
         "baseline_total_tokens": sum(r["baseline"]["tokens"] for r in results),
         "nuwave_total_tokens":   sum(r["nuwave"]["tokens"] for r in results),
         "tokens_saved":          sum(max(0, r["baseline"]["tokens"] - r["nuwave"]["tokens"]) for r in results),