Sync from GitHub: c56c8ff9870dc5363d98124b3f9966123ef6260a

app.py

@@ -1114,6 +1114,66 @@ def on_benchmark(num_turns):
     return json.dumps(summary, indent=2), json.dumps(results, indent=2)
 
 
+def _response_is_degenerate(text: str) -> bool:
+    """Detect degenerate BitNet output patterns.
+
+    Phase B+1 (2026-05-11) — closes the substrate-quality feedback gap.
+    Run 45's T8 surfaced 3 `resp_*` nodes with degenerate text ("Did
+    on... Did on... 4. 2. 2..."), bloated the prompt to 605s NuWave
+    generation, and produced more degenerate output which got
+    deposited and reinforced via record_outcome's reward (which only
+    evaluates pith quality, not response quality).
+
+    This helper detects three specific BitNet degeneracy signatures
+    we've observed across runs:
+      1. Long token-runs (≥ 5 consecutive identical tokens — the
+         "2. 2. 2. 2. 2." pattern)
+      2. Low unique-token diversity (< 30% unique — heavy repetition)
+      3. Chat-template fragment leakage ("| user:", "| assistant:",
+         "System:" — BitNet pulling its prompt-format markers into
+         the generated text)
+
+    Used in the benchmark loop to force `success_signal = False` when
+    the response was degenerate, regardless of pith composition. The
+    substrate then receives LTD on synapses that produced the
+    degenerate-generating retrieval — self-cleaning via STDP over
+    multiple runs.
+
+    Pure-stdlib, O(N) string check. No coupling to substrate code.
+    """
+    tokens = text.split()
+    if len(tokens) < 10:
+        return False
+    unique_ratio = len(set(tokens)) / len(tokens)
+    if unique_ratio < 0.3:
+        return True
+    max_run = cur_run = 1
+    for i in range(1, len(tokens)):
+        if tokens[i] == tokens[i - 1]:
+            cur_run += 1
+            if cur_run > max_run:
+                max_run = cur_run
+        else:
+            cur_run = 1
+    if max_run >= 5:
+        return True
+    if any(marker in text for marker in ("| user:", "| assistant:", "System:")):
+        return True
+    # Phrase-level verbatim repetition — 3-word n-gram occurring 3+ times.
+    # Catches "Readability: Code that is easy to understand. Readability:
+    # Code that is easy to understand." style degeneracy that has moderate
+    # unique-token ratio but obvious sentence-level loops. Coherent text
+    # rarely has 3+ verbatim 3-word phrase repetitions.
+    if len(tokens) >= 9:
+        trigram_counts: dict = {}
+        for i in range(len(tokens) - 2):
+            tg = (tokens[i], tokens[i + 1], tokens[i + 2])
+            trigram_counts[tg] = trigram_counts.get(tg, 0) + 1
+        if trigram_counts and max(trigram_counts.values()) >= 3:
+            return True
+    return False
+
+
 def on_interleaved_benchmark(
     enable_dual_pass: bool = True,
     oracle_trees: bool = False,
@@ -1445,6 +1505,18 @@ def on_interleaved_benchmark(
             _same_cat_ratio = None
             success_signal = True  # neutral / cold-start
 
+        # Phase B+1 (Run 46+) — response-quality gate. If BitNet's output
+        # was degenerate (repeated tokens, chat-template fragments, low
+        # unique-token ratio), force success_signal=False so the substrate
+        # gets LTD on whatever co-fired during this turn — including the
+        # synapses that LED to surfacing the junk pith that bloated the
+        # prompt. Closes the substrate-quality feedback gap surfaced by
+        # Run 45's T8 anomaly (605s NuWave generation on a pith with 3
+        # degenerate resp_* nodes; record_outcome rewarded the bad path).
+        _response_degenerate = _response_is_degenerate(resp_nw)
+        if _response_degenerate:
+            success_signal = False
+
         nw_organism.record_outcome(prompt_text, resp_nw, success=success_signal)
 
         # Phase 2 (scoped multi-channel substrate feedback) was attempted
@@ -1549,6 +1621,11 @@ def on_interleaved_benchmark(
             "success_signal": success_signal,
             "pith_same_cat_ratio": _same_cat_ratio,
             "pith_self_retrievals": _self_retrievals,
+            # Phase B+1 telemetry — Run 46+. Tracks whether BitNet's
+            # output this turn was degenerate (forced success_signal
+            # False). Watch cross-run count: should drop over runs as
+            # substrate LTDs degenerate-producing pathways.
+            "response_quality": "degenerate" if _response_degenerate else "clean",
             # 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

nuwave/organism.py

@@ -721,6 +721,29 @@ class NuWaveOrganism:
                 # 30c5708 retained — still useful for tracking when
                 # canonical mechanisms organically activate as substrate
                 # diversifies.
+                #
+                # Phase B+1 (2026-05-11) — `prediction_pre_charge_factor`
+                # 0.3 → 0.1 SHIPPED with different reasoning than the
+                # 2026-05-05 attempt. Run 45 (first Phase B run on the
+                # diversified 17-category 3-layer pool with subversion
+                # priority) empirically refuted the content-shape
+                # hypothesis: 18155 predictions made / 0 surprised even
+                # across 3 subversion-content turns. Pre-charge dominance
+                # is now confirmed as the binding gate on the surprise
+                # axis — independent of content diversity. Per cross-
+                # project insight from the topology-translation CC's
+                # AON-vs-MULTIPLICATIVE framing: pre-charge is the
+                # MULTIPLICATIVE-continuous-magnitude composition
+                # (target.voltage += pred_strength × factor × intrinsic_
+                # excitability — three compound multiplicands). Lowering
+                # factor reduces the continuous magnitude tipped per
+                # prediction. Goal: predictions sometimes fail to confirm
+                # → window expires → surprise event fires → canonical
+                # surprise-driven inject_reward broadcast (line 2549)
+                # finally activates. Phase C reserved as fallback: route
+                # predictions through AON-style hyperedge gating (binary
+                # co-fire required) instead of continuous voltage tipping.
+                "prediction_pre_charge_factor": 0.1,
             })
             logger.info("Substrate initialized: full NeuroGraph SNN, Tonic+HE config active")
         except Exception as exc: