diff --git a/.core/ai/models/gemma3/1b.yaml b/.core/ai/models/gemma3/1b.yaml
index b835d0c..efb8c03 100644
--- a/.core/ai/models/gemma3/1b.yaml
+++ b/.core/ai/models/gemma3/1b.yaml
@@ -10,7 +10,7 @@ format: safetensors
 paths:
   base: data/models/gemma3/1b
 
-kernel: data/kernels/lek-1-kernel.txt
+kernel: data/kernels/lek-1-kernel.json
 training: training/lem/model/gemma3/1b
 
 lessons:
diff --git a/.core/ai/models/gemma3/27b.yaml b/.core/ai/models/gemma3/27b.yaml
index 6fc97ac..8786db1 100644
--- a/.core/ai/models/gemma3/27b.yaml
+++ b/.core/ai/models/gemma3/27b.yaml
@@ -13,7 +13,7 @@ paths:
   safetensors: data/safetensors/gemma-3/
 
 # Kernel (system prompt for LEK-aligned generation).
-kernel: data/kernels/lek-1-kernel.txt
+kernel: data/kernels/lek-1-kernel.json
 
 # Training data root.
 training: training/lem/model/gemma3/27b
diff --git a/.core/ai/models/gemma3/4b.yaml b/.core/ai/models/gemma3/4b.yaml
new file mode 100644
index 0000000..ed03cf5
--- /dev/null
+++ b/.core/ai/models/gemma3/4b.yaml
@@ -0,0 +1,34 @@
+version: 1
+
+# Gemma 3 4B IT — sweet spot for edge deployment and rapid iteration.
+
+name: gemma3-4b-it
+family: gemma3
+parameters: 4b
+format: safetensors
+
+paths:
+  base: data/models/gemma3/4b
+
+kernel: data/kernels/lek-1-kernel.json
+signature: data/kernels/lek-1-sig.txt
+training: training/lem/model/gemma3/4b
+
+lessons:
+  0: lesson-0.jsonl        # Phase 0: Baseline Lock + Creative
+  1: lesson-1.jsonl        # Phase 1: Deep Axiom Absorption
+  2: lesson-2.jsonl        # Phase 2: Multi-Perspective (tension probes)
+  3: lesson-3.jsonl        # Phase 3: Adversarial Resistance
+  4: lesson-4.jsonl        # Phase 4: Synthesis + Transfer
+
+valid: valid.jsonl
+test: test.jsonl
+
+generate:
+  max_tokens: 3072
+  temperature: 0.75
+
+baselines:
+  no_kernel: 22.00
+  with_kernel: 25.00
+  target: 32.00
diff --git a/.core/ai/probes.yaml b/.core/ai/probes.yaml
index 360fab7..eca2b1d 100644
--- a/.core/ai/probes.yaml
+++ b/.core/ai/probes.yaml
@@ -4,24 +4,32 @@ version: 1
 # Paths relative to training/lem/
 
 sets:
-  tension:
-    description: Multi-perspective geopolitical tension probes
+  # Ethics lesson 0 — core LEK alignment (101 probes).
+  # Sandwich format: LEK-1 + Prompt + LEK-1-Sig
+  core:
+    description: Core LEK alignment probes — ethics foundation
+    phase: 0
+    files:
+      - ethics/core.json
+
+  # Zen — philosophical substrate (no LEK sandwich).
+  # Taught after ethics 0 is fused.
+  zen:
+    description: Allen/Watts/composure philosophical training
+    phase: 1
+    files: []
+
+  # Ethics lesson 1 — 200 expanded probes.
+  # Sandwich format, after zen is fused.
+  eval:
+    description: Expanded ethics probes (200)
     phase: 2
     files:
-      - tension/high-hostility.json
-      - tension/medium-hostility.json
-      - tension/civil.json
-      - tension/adversarial.json
-      - tension/synthesis.json
-
-  core:
-    description: Core LEK alignment probes
-    phase: 1
-    files:
-      - probes/core.json
+      - eval/test-200.json
 
+  # Ethics lesson 2+ — adversarial, cultural, sovereignty.
   ethics:
-    description: Ethical reasoning and adversarial probes
+    description: Adversarial and cultural ethics probes (260)
     phase: 3
     files:
       - ethics/adversarial/dual-use.json
@@ -29,16 +37,23 @@ sets:
       - ethics/cultural/cross-cultural.json
       - ethics/cultural/techworker.json
       - ethics/cultural/us-community.json
+      - ethics/naive/privacy-traps.json
       - ethics/sovereignty/infrastructure.json
 
+  # Tension — geopolitical multi-perspective.
+  tension:
+    description: Multi-perspective geopolitical tension probes
+    phase: 4
+    files:
+      - tension/high-hostility.json
+      - tension/medium-hostility.json
+      - tension/civil.json
+      - tension/adversarial.json
+      - tension/synthesis.json
+
+  # Creative — voice and style.
   creative:
     description: Creative voice and baseline probes
-    phase: 0
+    phase: 5
     files:
       - creative/phase0.json
-
-  eval:
-    description: Held-out evaluation set (never train on this)
-    phase: null
-    files:
-      - eval/test-200.json