Add v7 (best-of-N, 13/14 strong) as default model

.claude/unitares-session.json

@@ -0,0 +1,6 @@
+{
+  "server_url": "http://localhost:8767",
+  "agent_name": "claude_dialectic-reasoning-space_20260403",
+  "agent_id": "Claude_Code_20260403",
+  "client_session_id": "agent-a2c5b67e-59c"
+}

app.py

@@ -7,7 +7,11 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
 
 MODELS = {
-    "Qwen3-4B v6 (falsifiable, best)": {
+    "Qwen3-4B v7 (best-of-N, latest)": {
+        "base": "Qwen/Qwen3-4B",
+        "adapter": "hikewa/dialectic-qwen3-4b-v7-lora",
+    },
+    "Qwen3-4B v6 (falsifiable)": {
         "base": "Qwen/Qwen3-4B",
         "adapter": "hikewa/dialectic-qwen3-4b-v6-lora",
     },
@@ -118,14 +122,14 @@ demo = gr.ChatInterface(
     additional_inputs=[
         gr.Dropdown(
             choices=list(MODELS.keys()),
-            value="Qwen3-4B v6 (falsifiable, best)",
+            value="Qwen3-4B v7 (best-of-N, latest)",
             label="Model",
         ),
     ],
     title="Dialectic Reasoning Models",
     description=(
         "Fine-tuned on dialectic reasoning traces with falsifiability-based quality filtering. "
-        "v6 (89 traces, 100% falsifiable) scores 8.4 on Mistral-strict rubric — up from v3's 6.8. "
+        "v7 (113 best-of-N traces from 6 providers) scores 8.3 Mistral-strict with 13/14 strong verdicts. "
         "Pick a model and ask a question involving competing perspectives."
     ),
     examples=[