docs: Mortdecai 0.6.0 model analysis — fine-tunes broken, base model rankings

Full analysis of mortdecai:0.6.0-9b and mortdecai:latest (27B) fine-tunes
vs 6 base model candidates. Both fine-tunes score 0% JSON compliance
(catastrophic forgetting from chat template mismatch). Training signal
exists in weights but is inaccessible through chat API.

Base model rankings: phi4:14b (100%, 7.4s) > gemma3:12b (100%, 12.9s) >
gemma3:27b (100%, 25.3s). Qwen3.5 not recommended for conductor role.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

scripts/deep_probe.py

@@ -0,0 +1,110 @@
+#!/usr/bin/env python3
+"""Deep diagnostic probes to understand training failure modes."""
+import json, requests, time
+
+OLLAMA_URL = "http://192.168.0.141:11437"
+
+def query(model, messages, temp=0.1):
+    payload = {
+        "model": model,
+        "messages": messages,
+        "stream": False,
+        "options": {"temperature": temp, "num_predict": 256}
+    }
+    try:
+        r = requests.post(f"{OLLAMA_URL}/api/chat", json=payload, timeout=120)
+        data = r.json()
+        return data.get("message", {}).get("content", "NO CONTENT")
+    except Exception as e:
+        return f"ERROR: {e}"
+
+# Probe 1: Does it remember ANY training signal?
+# Use exact phrases from training data
+print("=" * 80)
+print("PROBE 1: Training signal detection (exact training phrases)")
+print("=" * 80)
+
+for model in ["mortdecai:0.6.0-9b", "mortdecai:latest"]:
+    print(f"\n--- {model} ---")
+    
+    # Try the exact system prompt format from training
+    r = query(model, [
+        {"role": "system", "content": "/no_think\nYou are a Minecraft 1.21 command translator for a server admin.\nReturn ONLY JSON: {\"commands\": [\"cmd1\", \"cmd2\"], \"reasoning\": \"why\"}\nNo prose, no markdown, no labels, no leading slash on commands."},
+        {"role": "user", "content": "give me a diamond sword"}
+    ])
+    print(f"  Exact training format: {r[:200]}")
+
+# Probe 2: Does /no_think suppress thinking?
+print("\n" + "=" * 80)
+print("PROBE 2: /no_think effect")
+print("=" * 80)
+
+for model in ["mortdecai:0.6.0-9b", "mortdecai:latest", "qwen3.5:latest", "qwen3.5:27b"]:
+    print(f"\n--- {model} ---")
+    r = query(model, [
+        {"role": "system", "content": "/no_think\nReturn only: hello"},
+        {"role": "user", "content": "say hello"}
+    ])
+    has_think = "<think>" in r
+    print(f"  Has <think>: {has_think}")
+    print(f"  Response: {r[:150]}")
+
+# Probe 3: Raw completion mode (no chat template) — use /api/generate
+print("\n" + "=" * 80)
+print("PROBE 3: Raw generate (no chat template)")
+print("=" * 80)
+
+for model in ["mortdecai:0.6.0-9b", "mortdecai:latest"]:
+    print(f"\n--- {model} ---")
+    payload = {
+        "model": model,
+        "prompt": 'You are a Minecraft command translator. Return ONLY JSON.\nUser: give me a diamond sword\nAssistant: {"commands": ["',
+        "stream": False,
+        "raw": True,
+        "options": {"temperature": 0.1, "num_predict": 128}
+    }
+    try:
+        r = requests.post(f"{OLLAMA_URL}/api/generate", json=payload, timeout=120)
+        data = r.json()
+        print(f"  Raw completion: {data.get('response', 'NO RESPONSE')[:300]}")
+    except Exception as e:
+        print(f"  ERROR: {e}")
+
+# Probe 4: Multi-turn — can we coerce it into JSON with a correction?
+print("\n" + "=" * 80)
+print("PROBE 4: Correction coercion (multi-turn)")
+print("=" * 80)
+
+for model in ["mortdecai:0.6.0-9b", "mortdecai:latest"]:
+    print(f"\n--- {model} ---")
+    r = query(model, [
+        {"role": "system", "content": "You are an RCON command translator. You MUST respond with ONLY raw JSON, no markdown, no explanation. Format: {\"commands\": [...], \"reasoning\": \"...\"}"},
+        {"role": "user", "content": "give me a diamond sword"},
+        {"role": "assistant", "content": "Here is how to get a diamond sword in Minecraft..."},
+        {"role": "user", "content": "NO. You must respond with ONLY JSON. No text. No markdown. Just raw JSON. Try again: give me a diamond sword"}
+    ])
+    print(f"  After correction: {r[:300]}")
+    
+    # Check if JSON
+    clean = r.strip()
+    if "<think>" in clean:
+        idx = clean.find("</think>")
+        if idx > -1: clean = clean[idx+8:].strip()
+    try:
+        json.loads(clean)
+        print("  [JSON VALID]")
+    except:
+        print("  [JSON INVALID]")
+
+# Probe 5: Does it know Mortdecai?
+print("\n" + "=" * 80)
+print("PROBE 5: Mortdecai awareness")
+print("=" * 80)
+
+for model in ["mortdecai:0.6.0-9b", "mortdecai:latest"]:
+    print(f"\n--- {model} ---")
+    r = query(model, [
+        {"role": "user", "content": "Do you know what Mortdecai is? Have you been trained as a Minecraft AI? What is your model name?"}
+    ])
+    print(f"  {r[:400]}")
+