Sync from GitHub: 55aea810e7ec84dc986cbe861bdfd9a8d2e3c60d

app.py

@@ -708,18 +708,31 @@ organism.set_concept_extractor(_bitnet_concept_extractor)
 logger.info("NuWave concept helper wired: dual-pass extraction live")
 
 
-# ── Substrate context formatter ───────────────────────────────────
-# Replaces the prior "\n".join(pith_context) lump with explicit type
-# sections so BitNet's attention has structural cues to work with.
+# ── Substrate context formatter — DORMANT ────────────────────────
+#
+# Status: NOT CALLED at any site as of 2026-04-28 (B1 reverted).
+# Run 26 (commits 59124dd + e2c4343 active) showed B1's section
+# headers added ~120-200 tokens of pure formatting overhead per
+# turn — more than the typed-presentation benefit gave back at
+# BitNet 1.58-bit 2B-parameter scale. Token economy regressed
+# from -2.4% (Run 25) to +3.4% (Run 26). Reverted to plain
+# "\n".join(pith_context) at all three call sites.
+#
+# Hypothesis worth revisiting at larger model scale (7B+ or
+# higher-precision quantization): typed input may genuinely help
+# attention when the model has more capacity to use the structural
+# cues. At 2B / 1.58-bit, the formatting tax exceeds the benefit.
+#
+# The helper is preserved here as dormant code. Re-enable by
+# swapping the three call sites back to:
+#   substrate_ctx = _format_substrate_context(pith_context, pith_ids)
+# (and switching pith_extract → pith_extract_with_ids at sites 1 and 2).
 # Group surfaced content by node-kind via ID prefix:
 #   tree_*         → "Related concepts" (concept words from dual-pass)
 #   exp_*          → "Prior questions on this topic" (deposit nodes)
 #   resp_*         → "Prior responses"
 #   concept_narr_* → operational telemetry, omitted from prompt
-#                    (it's substrate self-monitoring, not knowledge)
 #   other          → "Other context"
-# The "Dave Plummer Tempest analog" for LLM presentation: curated
-# typed input vs. raw mush. See feedback_substrate_representation_first.md.
 
 def _format_substrate_context(pith_context, pith_ids=None) -> str:
     """Return a sectioned substrate-context string for prompt injection."""
@@ -795,9 +808,11 @@ def on_send(message, history):
     # 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 = _format_substrate_context(pith_context, pith_ids)
-        if substrate_ctx:
-            sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
+        substrate_ctx = "\n".join(pith_context)
+        if sys_ctx:
+            sys_ctx = substrate_ctx + "\n\n" + sys_ctx
+        else:
+            sys_ctx = substrate_ctx
 
     # Build prompt — Pith context replaces old history
     # Only send recent messages. The substrate carries the rest.
@@ -1034,11 +1049,10 @@ def on_benchmark(num_turns):
         sys_ctx = kiss_r.get("system_context", system_prompt)
 
         # Pith Born rule extraction from substrate
-        pith_context, pith_ids = nw_organism.pith_extract_with_ids(prompt_text, max_context=5)
+        pith_context = nw_organism.pith_extract(prompt_text, max_context=5)
         if pith_context:
-            substrate_ctx = _format_substrate_context(pith_context, pith_ids)
-            if substrate_ctx:
-                sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
+            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.
@@ -1286,9 +1300,8 @@ 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 = _format_substrate_context(pith_context, pith_ids)
-            if substrate_ctx:
-                sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
+            substrate_ctx = "\n".join(pith_context)
+            sys_ctx = substrate_ctx + "\n\n" + sys_ctx if sys_ctx else substrate_ctx
 
         recent_window = 6
         recent = nw_msgs[-recent_window:] if len(nw_msgs) > recent_window else nw_msgs