Sync from GitHub: e776a7ae171da7f004d061e72643f07a30d0006c

app.py

@@ -5,6 +5,21 @@ The organism. NeuroGraph substrate + KISS bucket + Pith bucket +
 Splat-Lenia + BitNet model. On CPU. Gets smarter over time.
 
 # ---- Changelog ----
+# [2026-05-22] Claude Opus 4.7 (1M ctx) — D2: drop conversation history from prompt (3 sites)
+#   The 2026-05-16 pith→user-turn fix deployed cleanly (Run 49 confirmed:
+#   "My actual question:" label from the new template appearing IN BitNet
+#   output), but degeneracy persisted. Diagnosis: pollution feedback loop
+#   relocated from system slot to conversation history. BitNet's own prior
+#   degenerate responses were appended to nw_msgs/messages_nw as assistant
+#   turns; next turn's recent_window=6 pulled them back into context;
+#   BitNet kept echoing the pattern. Fix: drop the recent_window pull
+#   entirely. Prompt = [system, current user turn (with pith block)].
+#   No prior turns. nw_msgs/messages_nw still grows as a record (KISS
+#   still gets the full history for filtering) — the model only sees this
+#   turn. This is the architecturally-stated NuWave posture: "substrate
+#   carries continuity, not literal chat history" (per organism.py header).
+#   The recent_window=6 was always a pragmatic concession to model needs;
+#   removing it enforces the substrate-only-continuity design intent.
 # [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.
@@ -820,40 +835,33 @@ 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 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.
-
-    # 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
-    if len(messages_nw) > recent_window:
-        recent_msgs = messages_nw[-recent_window:]
-    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.
+    # D2 (2026-05-22): NO conversation history in the prompt.
+    # Substrate-only continuity per NuWave's design philosophy
+    # (organism.py header: "substrate carries continuity, not literal
+    # chat history"). The previous recent_window=6 was the source of
+    # the post-2026-05-16 degeneracy loop — BitNet's own prior degenerate
+    # outputs in messages_nw assistant turns were being pulled back into
+    # next turn's prompt, BitNet kept echoing the pattern. Closing it:
+    # prompt = [system instructions, current user turn (with pith block)].
+    # messages_nw still grows (KISS filter_context reads it for sys_ctx
+    # derivation, conversation record persists); the model just doesn't
+    # see prior turns directly.
     if pith_context:
         pith_block = "\n".join(f"  - {p}" for p in pith_context)
-        enriched_query = (
+        user_content = (
             "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}
-        ]
+    else:
+        user_content = message
 
     prompt_msgs = []
     if sys_ctx:
         prompt_msgs.append({"role": "system", "content": sys_ctx})
-    prompt_msgs.extend(recent_msgs)
+    prompt_msgs.append({"role": "user", "content": user_content})
 
     prompt = tokenizer.apply_chat_template(
         prompt_msgs, tokenize=False, add_generation_prompt=True,
@@ -1075,40 +1083,32 @@ 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 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 Born rule extraction from substrate.
         pith_context = nw_organism.pith_extract(prompt_text, max_context=5)
 
-        # Trim old messages — always, not gated on Pith.
-        # The substrate + KISS carry what the older messages contained.
-        # Even without Pith contexts, the recent window has what the model
-        # needs for immediate coherence.
-        recent_window = 6
-        if len(nw_msgs) > recent_window:
-            recent = nw_msgs[-recent_window:]
-        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.
+        # D2 (2026-05-22): NO conversation history in the prompt.
+        # Substrate-only continuity per NuWave's design. nw_msgs still
+        # grows (KISS reads it for sys_ctx); the model only sees this
+        # turn's user message with the labeled pith context block.
         if pith_context:
             pith_block = "\n".join(f"  - {p}" for p in pith_context)
-            enriched_query = (
+            user_content = (
                 "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}
-            ]
+        else:
+            user_content = prompt_text
+
+        prompt_msgs_nw = []
+        if sys_ctx:
+            prompt_msgs_nw.append({"role": "system", "content": sys_ctx})
+        prompt_msgs_nw.append({"role": "user", "content": user_content})
 
         prompt_nw = tokenizer.apply_chat_template(
-            [{"role": "system", "content": sys_ctx}] + recent if sys_ctx else recent,
-            tokenize=False, add_generation_prompt=True,
+            prompt_msgs_nw, tokenize=False, add_generation_prompt=True,
         )
         resp_nw, in_nw, out_nw, time_nw, tps_nw, lenia_r = do_generate(prompt_nw, max_new_tokens=128)
         nw_msgs.append({"role": "assistant", "content": resp_nw})
@@ -1469,30 +1469,31 @@ def on_interleaved_benchmark(
 
         kiss_r = nw_kiss_inst.filter_context(nw_msgs, system_prompt)
         sys_ctx = kiss_r.get("system_context", system_prompt)
-        # 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
 
+        # D2 (2026-05-22): NO conversation history in the prompt.
+        # Substrate-only continuity per NuWave's design philosophy.
+        # nw_msgs still grows as a record (KISS reads it for sys_ctx
+        # derivation just above); the model only sees this turn's user
+        # message with the labeled pith context block.
         if pith_context:
             pith_block = "\n".join(f"  - {p}" for p in pith_context)
-            enriched_query = (
+            user_content = (
                 "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}
-            ]
+        else:
+            user_content = prompt_text
+
+        prompt_msgs_iv = []
+        if sys_ctx:
+            prompt_msgs_iv.append({"role": "system", "content": sys_ctx})
+        prompt_msgs_iv.append({"role": "user", "content": user_content})
 
         prompt_nw = tokenizer.apply_chat_template(
-            [{"role": "system", "content": sys_ctx}] + recent if sys_ctx else recent,
-            tokenize=False, add_generation_prompt=True,
+            prompt_msgs_iv, tokenize=False, add_generation_prompt=True,
         )
         resp_nw, in_nw, out_nw, time_nw, tps_nw, _ = do_generate(prompt_nw, max_new_tokens=128)
         nw_msgs.append({"role": "assistant", "content": resp_nw})