From 15c766f0ecb97a484b2a1547577074adc60c8912 Mon Sep 17 00:00:00 2001
From: Raphael Southall <69826455+raphasouthall@users.noreply.github.com>
Date: Sun, 23 Aug 2026 18:57:08 +0100
Subject: [PATCH] fix(synthesize): bump SIBLING_THRESHOLD default to 0.75

Live runs 2026-08-23: 0.65 cross-linked unrelated topics and pulled a
credential-bearing memory into a cluster; 0.75 gave 10 tight clusters
whose 5 live learnings (1834-1838) read well. Calibration comment
updated with the evidence.

Part of #111
---
 src/neurostack/synthesize.py | 7 +++++--
 tests/test_synthesize.py     | 2 +-
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/src/neurostack/synthesize.py b/src/neurostack/synthesize.py
index ac91d14..0a26df9 100644
--- a/src/neurostack/synthesize.py
+++ b/src/neurostack/synthesize.py
@@ -38,10 +38,13 @@
 DEFAULT_MIN_SIBLINGS = 3
 # Live-calibrated on the prod embedder (embeddinggemma:300m) against the real
 # memory store: 0.35-0.55 chains dense early-harvest noise into 100+ member
-# mega-clusters; 0.65 yields coherent 4-15 member topic groups. Paraphrases
+# mega-clusters; 0.65 still cross-linked unrelated topics (one cluster mixed
+# neurostack-cloud, KPI-dashboard, AFD, firewall and Citrix observations) and
+# pulled credential-bearing rows along; 0.75 yields tight same-topic groups
+# whose live learnings read well (issue #111, run of 2026-08-23). Paraphrases
 # (~0.93) are handled earlier by harvest dedup — synthesis clusters are the
 # same-topic tier.
-SIBLING_THRESHOLD = 0.65
+SIBLING_THRESHOLD = 0.75
 # Hard per-cluster member ceiling: one learning cannot faithfully preserve the
 # facts of an unbounded heap, and the prompt must stay bounded. Oversized
 # clusters keep the anchor plus its most similar members; the rest stay
diff --git a/tests/test_synthesize.py b/tests/test_synthesize.py
index f1e513d..963c893 100644
--- a/tests/test_synthesize.py
+++ b/tests/test_synthesize.py
@@ -152,7 +152,7 @@ def test_missing_embeddings_counted_not_dropped(self, in_memory_db):
         assert len(clusters) == 1
 
     def test_threshold_default_live_calibrated(self):
-        assert SIBLING_THRESHOLD == 0.65
+        assert SIBLING_THRESHOLD == 0.75
 
     def test_oversized_cluster_capped_keeps_anchor_and_most_similar(
             self, in_memory_db):