Sync from GitHub: 87142a222fc1cc85c27aacef69860f9a8795ac97

app.py

@@ -5,6 +5,20 @@ The organism. NeuroGraph substrate + KISS bucket + Pith bucket +
 Splat-Lenia + BitNet model. On CPU. Gets smarter over time.
 
 # ---- Changelog ----
+# [2026-05-16] Claude Opus 4.7 (1M ctx) — Pith → user-turn labeled context block (3 sites)
+#   Root cause of Run 48's 24/24 degenerate BitNet output: pith was being
+#   injected into the system role slot via sys_ctx = "\n".join(pith) + sys.
+#   Chat-tuned models are trained with the system slot carrying instructions,
+#   not lists of prior user questions. BitNet was treating retrieved pith
+#   as questions to respond to, echoing them back and leaking chat-template
+#   fragments. Fix: pith content now goes in a clearly-labeled context block
+#   inside the LAST user turn for THIS turn's prompt only; bare user query
+#   persists in nw_msgs/messages_nw so next turn's recent_window isn't
+#   polluted. System slot stays canonical (instructions only). Plain-text
+#   labels, no delimiter tokens (per feedback_pith_presentation_layer memory).
+#   Three call sites updated identically: on_send live chat, first benchmark
+#   loop, interleaved benchmark loop. Universal RAG pattern — applies to
+#   any future LLM consumer of substrate-surfaced content.
 # [2026-04-06] Claude Code (Opus 4.6) — Full NeuroGraph organism integration
 # [2026-03-31] Claude Code (Opus 4.6) — Switch to BitNet 2B for CPU-native inference
 # [2026-03-29] Claude Code (Opus 4.6) — ZeroGPU compatible, model at startup
@@ -806,18 +820,12 @@ def on_send(message, history):
     # ── 4. Pith bucket — extract relevant context from the River ──
     pith_context, pith_ids = organism.pith_extract_with_ids(message, max_context=5)
 
-    # Pith context REPLACES old message history, not adds to it.
-    # The substrate carries what the older messages contained — the model
-    # doesn't need both. Recent messages pass verbatim (the model needs
-    # immediate context). Older messages are replaced by substrate context.
-    if pith_context:
-        substrate_ctx = "\n".join(pith_context)
-        if sys_ctx:
-            sys_ctx = substrate_ctx + "\n\n" + sys_ctx
-        else:
-            sys_ctx = substrate_ctx
+    # Pith does NOT go in the system slot. System slot is for the model's
+    # identity/instructions; chat-tuned models treat retrieved questions
+    # placed there as things to respond to (Run 48 root cause). Instead,
+    # pith goes in a labeled context block inside the LAST user turn for
+    # THIS turn's prompt only — see feedback_pith_presentation_layer memory.
 
-    # Build prompt — Pith context replaces old history
     # Only send recent messages. The substrate carries the rest.
     # Always trim to recent window. The substrate + KISS carry older context.
     recent_window = 6  # 3 turns of user+assistant
@@ -826,6 +834,22 @@ def on_send(message, history):
     else:
         recent_msgs = messages_nw
 
+    # Enrich the LAST user message for THIS turn's prompt only.
+    # messages_nw stays clean (bare query persists in history) so next
+    # turn's recent_window isn't polluted with this turn's pith.
+    if pith_context:
+        pith_block = "\n".join(f"  - {p}" for p in pith_context)
+        enriched_query = (
+            "Some context that may be relevant (recalled from earlier "
+            "related conversations; these are reference material, not "
+            "questions to answer):\n"
+            f"{pith_block}\n\n"
+            f"My actual question: {message}"
+        )
+        recent_msgs = list(recent_msgs[:-1]) + [
+            {"role": "user", "content": enriched_query}
+        ]
+
     prompt_msgs = []
     if sys_ctx:
         prompt_msgs.append({"role": "system", "content": sys_ctx})
@@ -1051,11 +1075,10 @@ def on_benchmark(num_turns):
         kiss_r = nw_kiss.filter_context(nw_msgs, system_prompt)
         sys_ctx = kiss_r.get("system_context", system_prompt)
 
-        # Pith Born rule extraction from substrate
+        # Pith Born rule extraction from substrate. Pith does NOT go in
+        # sys_ctx — it goes in a labeled context block in the last user
+        # turn below (see feedback_pith_presentation_layer memory).
         pith_context = nw_organism.pith_extract(prompt_text, max_context=5)
-        if pith_context:
-            substrate_ctx = "\n".join(pith_context)
-            sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
 
         # Trim old messages — always, not gated on Pith.
         # The substrate + KISS carry what the older messages contained.
@@ -1067,6 +1090,22 @@ def on_benchmark(num_turns):
         else:
             recent = nw_msgs
 
+        # Enrich the LAST user message for THIS turn's prompt only.
+        # nw_msgs stays clean (bare query persists in history) so next
+        # turn's recent_window isn't polluted with this turn's pith.
+        if pith_context:
+            pith_block = "\n".join(f"  - {p}" for p in pith_context)
+            enriched_query = (
+                "Some context that may be relevant (recalled from earlier "
+                "related conversations; these are reference material, not "
+                "questions to answer):\n"
+                f"{pith_block}\n\n"
+                f"My actual question: {prompt_text}"
+            )
+            recent = list(recent[:-1]) + [
+                {"role": "user", "content": enriched_query}
+            ]
+
         prompt_nw = tokenizer.apply_chat_template(
             [{"role": "system", "content": sys_ctx}] + recent if sys_ctx else recent,
             tokenize=False, add_generation_prompt=True,
@@ -1430,12 +1469,27 @@ def on_interleaved_benchmark(
 
         kiss_r = nw_kiss_inst.filter_context(nw_msgs, system_prompt)
         sys_ctx = kiss_r.get("system_context", system_prompt)
-        if pith_context:
-            substrate_ctx = "\n".join(pith_context)
-            sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
+        # Pith does NOT go in sys_ctx — see feedback_pith_presentation_layer
+        # memory. It goes in a labeled context block in the last user turn
+        # below for THIS turn's prompt only; nw_msgs stays clean so next
+        # turn's recent_window isn't polluted with this turn's pith.
 
         recent_window = 6
         recent = nw_msgs[-recent_window:] if len(nw_msgs) > recent_window else nw_msgs
+
+        if pith_context:
+            pith_block = "\n".join(f"  - {p}" for p in pith_context)
+            enriched_query = (
+                "Some context that may be relevant (recalled from earlier "
+                "related conversations; these are reference material, not "
+                "questions to answer):\n"
+                f"{pith_block}\n\n"
+                f"My actual question: {prompt_text}"
+            )
+            recent = list(recent[:-1]) + [
+                {"role": "user", "content": enriched_query}
+            ]
+
         prompt_nw = tokenizer.apply_chat_template(
             [{"role": "system", "content": sys_ctx}] + recent if sys_ctx else recent,
             tokenize=False, add_generation_prompt=True,