Sync from GitHub: 9e469f21bece8d35a8d9e4e26e2d7eb6c8c55efe

app.py

@@ -1312,27 +1312,60 @@ def on_interleaved_benchmark(
         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})
 
-        # ── Per-turn correctness signal (Run 31+) ──────────────────────────
-        # Replaces the hardcoded success=True that fed the substrate uniform
-        # reward across every deposit. Run 30 diagnostic showed
-        # ignition_mean_cross_category (0.0942) > ignition_mean_same_category
-        # (0.0519) — three-factor learning was firing but had no
-        # discriminating signal: every co-firing pair got LTP'd equally,
-        # including cross-category contaminations.
+        # ── Per-turn correctness signal (Run 33+) ──────────────────────────
+        # Did this turn's pith pull predominantly USEFUL same-category
+        # content — excluding self-retrievals (pith ids whose embedding is
+        # near-identical to the query, i.e., the substrate handing the query
+        # back at us)?
         #
-        # Signal: did this turn's pith pull predominantly same-category
-        # content? Uses the same _categorize_node + centroid machinery the
-        # heatmap already trusts (cosine to per-category prompt centroids,
-        # 0.30 threshold). Untaggable nodes are ignored (don't dilute either
-        # side); fewer than 2 tagged ⇒ neutral True (cold-start protection).
+        # History:
+        #  - Run 30: hardcoded success=True → inverted ignition asymmetry
+        #    (cross-cat firing harder than same-cat).
+        #  - Run 31 (same-cat ratio threshold ≥ 0.5): ignition flipped sign
+        #    in one run; token regression collapsed +6.1% → +0.51%.
+        #  - Run 32: signal got gamed by question-repetition. Prior-run
+        #    deposits of the same query text are same-category-tagged, so
+        #    ratio = 1.0 on 5/8 turns. Substrate over-LTP'd at 5× normal
+        #    rate (~56K new synapses vs typical ~11K). Token regression
+        #    jumped to +12.1%, wall-clock +8.6% slower.
+        #
+        # Self-retrieval gate: for each pith id, cosine similarity between
+        # its embedding and the current query's embedding. If above
+        # _SELF_RETRIEVAL_THRESHOLD (0.92), node is a near-identical text
+        # repeat — counts toward tagged_total but NOT tagged_same. Drives
+        # ratio DOWN for question-repeat-heavy turns, so canonical STDP
+        # depresses self-retrieval synapses via LTD over multiple runs.
+        #
+        # This is a feedback-path correction (refines the reward signal we
+        # feed canonical inject_reward), NOT an extraction-path filter —
+        # pith still goes to the LLM unchanged. Substrate's STDP retrieves
+        # what it retrieves; we only refine our judgement of "did that
+        # help" so the canonical reward channel has accurate ground truth
+        # to learn against.
+        _SELF_RETRIEVAL_THRESHOLD = 0.92
+        _q_emb = np.asarray(nw_organism._embed_fn(prompt_text), dtype=np.float32)
+        _q_norm = float(np.linalg.norm(_q_emb)) + 1e-8
         _tagged_total = 0
         _tagged_same = 0
+        _self_retrievals = 0
         for _pid in pith_ids:
             _tag = _categorize_node(_pid)
-            if _tag is None:
+            _node_emb = nw_organism._embeddings.get(_pid)
+            _is_self = False
+            if _node_emb is not None:
+                _node_norm = float(np.linalg.norm(_node_emb)) + 1e-8
+                _cos_to_query = float(
+                    np.dot(_q_emb, _node_emb) / (_q_norm * _node_norm)
+                )
+                _is_self = _cos_to_query > _SELF_RETRIEVAL_THRESHOLD
+            if _is_self:
+                _self_retrievals += 1
+            # Skip only when BOTH untaggable AND not self-retrieval (no signal)
+            if _tag is None and not _is_self:
                 continue
             _tagged_total += 1
-            if _tag == category:
+            # Same-cat credit only if tag matches AND not a self-retrieval
+            if _tag == category and not _is_self:
                 _tagged_same += 1
         if _tagged_total >= 2:
             _same_cat_ratio = _tagged_same / _tagged_total
@@ -1439,9 +1472,12 @@ def on_interleaved_benchmark(
             # Run 31+ correctness-signal telemetry — what we fed the substrate
             # via record_outcome's success arg this turn, and the underlying
             # same-category proportion. ratio is None when fewer than 2 pith
-            # ids were taggable (cold-start neutral).
+            # ids were taggable (cold-start neutral). pith_self_retrievals
+            # added Run 33+: count of pith ids with cosine ≥ 0.92 to query
+            # (substrate handing the query back) — these count as misses.
             "success_signal": success_signal,
             "pith_same_cat_ratio": _same_cat_ratio,
+            "pith_self_retrievals": _self_retrievals,
         })
 
     # ── Heatmap A: ignition-set Jaccard overlap (symmetric) ──