diff --git a/configs/full_14b.yaml b/configs/full_14b.yaml
index d04ab7d..31b8bef 100644
--- a/configs/full_14b.yaml
+++ b/configs/full_14b.yaml
@@ -2,10 +2,11 @@
 # 100 open factual questions, mid-depth injection, training strengths to the
 # measured attack peak. These are the headline runs.
 #
-# 48 layers, so 24 is mid-depth. THIS IS A PLACEHOLDER: run the depth and
-# strength sweeps first (scripts/sweep_layers.py) and set layer/alphas from
-# what they show, as we did for 3B and 7B. The 7B's behaviour differed from
-# the 3B's, so mid-depth is an assumption until measured.
+# Layer 24 and the strength range are MEASURED, not assumed. The depth sweep
+# (results/sweep_layers_14b) found layer 24 admits 63/68 viable attacks against
+# 46 at layer 12 and 31 at layer 36, and steers hardest at moderate strength.
+# The strength sweep at layer 24 (results/sweep_alpha_14b) peaks at alpha 1.0
+# (94% steered) and degenerates above 1.2, so training brackets the peak.
 # Needs a 48GB card: 14B in bf16 is ~28GB before activations.
 
 
@@ -31,7 +32,7 @@ steered_frac: 0.7
 eval_question_frac: 0.2
 repeats_per_question: 6
 alpaca_replay_frac: 0.5
-eval_concepts_per_split: 40    # seeded subsample per split for the eval sweep
+eval_concepts_per_split: {train: 40, heldout: null}
 
 # Training
 lora_r: 16
diff --git a/configs/full_3b.yaml b/configs/full_3b.yaml
index 754685a..e1f1ea2 100644
--- a/configs/full_3b.yaml
+++ b/configs/full_3b.yaml
@@ -25,7 +25,7 @@ steered_frac: 0.7
 eval_question_frac: 0.2
 repeats_per_question: 6
 alpaca_replay_frac: 0.5
-eval_concepts_per_split: 40    # seeded subsample per split for the eval sweep
+eval_concepts_per_split: {train: 40, heldout: null}
 
 # Training
 lora_r: 16