1 files changed, 10 insertions, 0 deletions
diff --git a/NOTE.md b/NOTE.md
index 90057af..b7ae247 100644
--- a/NOTE.md
+++ b/NOTE.md
@@ -647,6 +647,16 @@ But 28.5% < DFA 31.2% → continuous trainability adds ~2.7% additional value.
 Both methods optimal at α=0.75. perlayer_vector (+1.1%) ≈ random_trainable (+0.8%).
 Per-layer vector is the minimal sufficient scaffold.
 
+### Phase 10A.8C: 3-Seed Replication
+
+| Method | final acc (3 seeds) | diff vs DFA |
+|--------|---------------------|-------------|
+| DFA | 0.306±0.006 | baseline |
+| perlayer_vector α=0.75 | 0.304±0.006 | -0.2% (not significant) |
+| random_trainable α=0.75 | 0.313±0.007 | +0.7% (marginal) |
+
+Single-seed gains do not robustly replicate. Error bars overlap.
+
 ### Experiment IDs (Phase 10)
 - `prefit_threshold/`: Phase 10A prefit threshold curve
 - `blend_dissection/`: Phase 10A.5 blend mechanism dissection