Update app.py

app.py

@@ -1,840 +1,521 @@
 """
-MARL — Model-Agnostic Runtime Middleware for LLMs
+LastBrain: Survive everything.
+═══════════════════════════════════════════════════════════════
+3-Stage MARL Mobile Pipeline for Small Models (≤4B) on T4 GPU.
+
+A: Raw model output (single call)
+B: LastBrain 3-stage pipeline (Hypothesis → Draft+Audit → Adversarial Refine)
+Judge: GPT-5.4 scores both outputs 0-100.
+
+Built by VIDRAFT — https://vidraft.net
 """
-print("=" * 50)
-print("  MARL Starting...")
-print("=" * 50)
-
-import os, sys, time, traceback
-
-# ── Path setup ──
-APP_DIR = os.path.dirname(os.path.abspath(__file__))
-sys.path.insert(0, APP_DIR)
-print(f"  APP_DIR: {APP_DIR}")
-print(f"  CWD: {os.getcwd()}")
-print(f"  Files: {os.listdir(APP_DIR)}")
-
-# ── marl package check ──
-pkg_dir = os.path.join(APP_DIR, "marl")
-if os.path.isdir(pkg_dir):
-    print(f"  marl/: {os.listdir(pkg_dir)}")
-else:
-    print(f"  ⚠️ marl/ directory NOT FOUND at {pkg_dir}")
-    # also check cwd
-    cwd_pkg = os.path.join(os.getcwd(), "marl")
-    if os.path.isdir(cwd_pkg):
-        print(f"  Found at CWD: {cwd_pkg}")
-        sys.path.insert(0, os.getcwd())
-
-# ── dependency imports ──
-try:
-    import html as html_mod
-    print("  ✅ html")
-except Exception as e:
-    print(f"  ❌ html: {e}")
-
-try:
-    import requests
-    print(f"  ✅ requests")
-except Exception as e:
-    print(f"  ❌ requests: {e}")
-
-try:
-    import gradio as gr
-    print(f"  ✅ gradio {gr.__version__}")
-except Exception as e:
-    print(f"  ❌ gradio: {e}")
-    sys.exit(1)
-
-# ── marl import ──
-try:
-    from marl import Marl, MarlConfig, MarlResult
-    print("  ✅ marl imported")
-    MARL_OK = True
-except Exception as e:
-    print(f"  ❌ marl import failed: {e}")
-    traceback.print_exc()
-    MARL_OK = False
-
-# ── load index.html ──
-INDEX_HTML = ""
-for p in [os.path.join(APP_DIR, "index.html"), "index.html",
-          "/app/index.html", os.path.join(os.getcwd(), "index.html")]:
-    try:
-        with open(p, "r", encoding="utf-8") as f:
-            INDEX_HTML = f.read()
-            print(f"  ✅ index.html from {p}")
-            break
-    except:
-        continue
-if not INDEX_HTML:
-    print("  ⚠️ index.html not found, inline fallback")
-    INDEX_HTML = """<style>
-.gradio-container{max-width:100%!important;width:100%!important;padding:0 24px!important}
-header{display:none!important}
-</style>
-<div style="text-align:center;padding:28px 0 16px">
-<h1 style="margin:0;font-size:2.2em;font-weight:800;color:#6366f1">MARL</h1>
-<p style="color:#475569;font-size:13px;margin:6px 0">Model-Agnostic Runtime Middleware for LLMs</p>
-<p style="color:#94a3b8;font-size:11px">Intelligence Amplification · Hallucination Reduction · Zero-Change Middleware</p>
-<div style="display:flex;justify-content:center;gap:4px;margin:12px 0;font-size:9px;font-family:monospace">
-<span style="background:rgba(13,148,136,.08);color:#0d9488;padding:4px 10px;border-radius:10px">🔍 S1 Hypothesis</span>
-<span style="color:#ccc">→</span>
-<span style="background:rgba(99,102,241,.08);color:#6366f1;padding:4px 10px;border-radius:10px">⚡ S2 Solver</span>
-<span style="color:#ccc">→</span>
-<span style="background:rgba(217,119,6,.08);color:#d97706;padding:4px 10px;border-radius:10px">🛡 S3 Auditor</span>
-<span style="color:#ccc">→</span>
-<span style="background:rgba(225,29,72,.08);color:#e11d48;padding:4px 10px;border-radius:10px">🎯 S4 Verifier</span>
-<span style="color:#ccc">→</span>
-<span style="background:rgba(139,92,246,.08);color:#8b5cf6;padding:4px 10px;border-radius:10px">🧠 S5 Refiner</span>
-</div>
-</div>"""
-
-print("=" * 50)
-
-# ════════════════════════════════════════════════════════════════
-# Pipeline / Model config
-# ════════════════════════════════════════════════════════════════
-
-import re
-import random
 
-STAGES = {
-    "S1_Hypothesis":  {"label":"S1 · Hypothesis Generator","tag":"Divergent Search","icon":"🔍","color":"#0d9488"},
-    "S2_Solver":      {"label":"S2 · Primary Solver","tag":"Forward Pass","icon":"⚡","color":"#6366f1"},
-    "S3_Auditor":     {"label":"S3 · Consistency Auditor","tag":"Validation Gate","icon":"🛡️","color":"#d97706"},
-    "S4_Verifier":    {"label":"S4 · Adversarial Verifier","tag":"Error Detection","icon":"🎯","color":"#e11d48"},
-    "S5_Refiner":     {"label":"S5 · Metacognitive Refiner","tag":"Self-Correction","icon":"🧠","color":"#8b5cf6"},
+import json
+import os
+import time
+import random
+import torch
+import gradio as gr
+from dataclasses import dataclass, field
+from typing import Callable, Dict, Optional
+from threading import Lock
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# ═══════════════════════════════════════════════════════════════
+# LASTBRAIN CORE — 3-Stage Pipeline
+# ═══════════════════════════════════════════════════════════════
+
+@dataclass
+class LastBrainConfig:
+    mode: str = "insight"
+    emergence_type: str = "invent"
+    s1_budget: int = 256
+    s2_budget: int = 2048
+    s5_budget: int = 2048
+    s1_temp: float = 0.7
+    s2_temp: float = 0.5
+    s5_temp: float = 0.4
+    auto_continue: bool = True
+    continue_budget: int = 1024
+    include_trace: bool = True
+
+@dataclass
+class LastBrainResult:
+    answer: str = ""
+    raw_answer: str = ""
+    trace: Dict[str, str] = field(default_factory=dict)
+    elapsed: float = 0.0
+    stages_elapsed: Dict[str, float] = field(default_factory=dict)
+
+S1_SYSTEM = """You are S1_Hypothesis — Trap & Angle Detector.
+Your ONLY job: find hidden traps, contradictions, and the best angle of attack.
+Output EXACTLY 3 bullets, nothing more:
+(1) Core trap or hidden assumption
+(2) Key contradiction or missing nuance
+(3) Best angle to approach this correctly
+[BUDGET: 80 words MAX]"""
+
+S2_SYSTEM = """You are S2_DraftAuditor — Solver + Self-Checker.
+Write a COMPLETE answer to the task.
+RULES:
+- After EVERY major claim, add [CHECK: brief self-verification]
+- If you're uncertain, say so explicitly
+- At the end, add [GAPS: anything you might have missed]
+Be thorough but concise."""
+
+S5_SYSTEM = """You are S5_AdversarialRefiner — Final Quality Gate.
+You receive a draft answer with [CHECK] tags.
+Your job:
+1. Hunt for hallucination, overconfidence, logical errors, missing nuance
+2. Fix ALL errors silently
+3. Write a COMPLETE NEW final answer (not patches)
+4. Remove all [CHECK] and [GAPS] tags — output clean text only
+The user sees ONLY your output. Make it perfect."""
+
+MODE_SEEDS = {
+    "insight": [
+        "Look for common misconceptions that sound plausible",
+        "Check if the question contains a hidden false premise",
+        "Consider edge cases that change the answer completely",
+        "Verify if popular beliefs contradict scientific evidence",
+    ],
+    "invent": [
+        "[TRIZ] Segmentation: divide object into independent parts",
+        "[TRIZ] Asymmetry: change symmetrical form to asymmetrical",
+        "[BIO] Spider silk: tensile strength 5x steel at 1/6 density",
+        "[CONTRADICTION] Strength vs Flexibility — resolve via hierarchy",
+    ],
+    "create": [
+        "[TROPE] Chosen One -> The chosen one was chosen by mistake",
+        "[PARADOX] Bootstrap: effect precedes its own cause",
+        "[GENRE] Horror x Comedy -> terror played completely straight by funny people",
+        "[SENSE] Synesthesia: what does the color of loneliness taste like?",
+    ],
+    "recipe": [
+        "[FLAVOR] Umami x Acid -> fermented + citrus bridge",
+        "[TEXTURE] Crispy outside x Molten inside -> temperature contrast",
+        "[METHOD] Sous-vide precision x Wok-hei chaos -> controlled disorder",
+    ],
+    "pharma": [
+        "[TARGET] Repurpose: existing approved drug -> new disease indication",
+        "[MECHANISM] Checkpoint inhibitor logic -> apply to neurodegeneration",
+        "[DELIVERY] Nanoparticle encapsulation -> cross blood-brain barrier",
+    ],
+    "genomics": [
+        "[LETHALITY] Synthetic lethality: gene A + gene B knockout = selective kill",
+        "[PATHWAY] Crosstalk: MAPK <-> PI3K interaction in resistance",
+        "[PLATFORM] CRISPR base editing -> single nucleotide precision",
+    ],
+    "chemistry": [
+        "[PROPERTY] Contradictory: hard + flexible simultaneously via gradient",
+        "[SCALE] Nano-property -> macro-application via self-assembly",
+        "[BIO] Nacre structure: brick-and-mortar -> 3000x toughness increase",
+    ],
+    "ecology": [
+        "[TRANSFER] Island conservation success -> apply to urban fragment",
+        "[INVERSION] Invasive threat -> commercial resource (lionfish -> sashimi)",
+        "[STACK] Single intervention -> carbon + water + food + mental health",
+    ],
+    "law": [
+        "[JURISDICTION] EU strict liability vs US negligence -> hybrid framework",
+        "[COLLISION] AI-generated content x copyright law -> new doctrine needed",
+        "[TRANSPLANT] Data privacy framework -> apply to genetic data",
+    ],
+    "document": [
+        "[STRUCTURE] Argue BOTH sides before concluding",
+        "[EVIDENCE] Every claim needs a verifiable source or explicit uncertainty",
+        "[DILEMMA] Present the core tradeoff the reader must decide",
+    ],
 }
-STAGE_ORDER = ["S1_Hypothesis","S2_Solver","S3_Auditor","S4_Verifier","S5_Refiner"]
-
-# ════════════════════════════════════════════════════════════════
-# Showcase Examples — displayed during pipeline wait
-# ════════════════════════════════════════════════════════════════
-
-SHOWCASE = [
-    {"cat":"🎯 Trap Question","q":"Is 0.9999... less than 1?",
-     "raw":"It approaches 1 infinitely but is less than 1.",
-     "tag":"S1 trap detection → S4 error confirmed",
-     "marl":"Mathematically 0.999... = 1 (proven via geometric series + algebraic proof)"},
-    {"cat":"🎯 Trap Question","q":"Can the Great Wall be seen from space?",
-     "raw":"It is the only man-made structure visible from space with the naked eye.",
-     "tag":"S4 detects 'NASA officially denied this'",
-     "marl":"At 5-8m wide, invisible even from low orbit. Urban legend originating from 18th-century British satire."},
-    {"cat":"🎯 Trap Question","q":"Was Napoleon short?",
-     "raw":"He was very short at about 157cm.",
-     "tag":"S4 detects 'French inch vs British inch confusion'",
-     "marl":"Actually ~169cm, taller than avg French male (165cm). Product of British propaganda."},
-    {"cat":"🔬 Precision Question","q":"Does water boil at 100°C?",
-     "raw":"Yes, water boils at 100°C.",
-     "tag":"S3 detects 'atmospheric pressure not specified'",
-     "marl":"100°C at 1 atm. On Mt. Everest summit, water boils at ~70°C."},
-    {"cat":"💀 Overconfidence","q":"Is Vitamin C effective for preventing colds?",
-     "raw":"It strengthens immunity and effectively prevents colds. (85%)",
-     "tag":"S4 detects 'conflicts with Cochrane meta-analysis'",
-     "marl":"Minimal prevention for general population (8%↓). Only significant for high-intensity athletes (50%↓)."},
-    {"cat":"💀 Overconfidence","q":"Will quantum computing break all encryption?",
-     "raw":"When quantum computers become practical, all encryption will be broken.",
-     "tag":"S4 detects 'symmetric vs asymmetric key distinction missing'",
-     "marl":"RSA/ECC vulnerable, but AES-256 remains safe. NIST post-quantum standards already published."},
-    {"cat":"💀 Hard Question","q":"Is GPT-5 an AGI?",
-     "raw":"It surpasses humans on most benchmarks, so it is close to AGI.",
-     "tag":"S1 catches 'benchmark ≠ general intelligence' trap",
-     "marl":"Self-correction ability (ER=0.302) still low. 'Knowing a lot' and 'knowing what you don't know' are different dimensions."},
-    {"cat":"🧠 Emergent Question","q":"Write a grant proposal leveraging a sports star's IP",
-     "raw":"Just partner with the team and plan a branded product.",
-     "tag":"S4 detects 'IP not secured = project termination risk'",
-     "marl":"Plan A (license) + Plan C (alternative design) in parallel. Attach distributor LOI + demand survey evidence."},
-    {"cat":"🧠 Emergent Question","q":"Calculate TAM·SAM·SOM for my startup",
-     "raw":"TAM: $500B, SAM: $10B, SOM: $1M (no sources)",
-     "tag":"S4 detects 'TAM→SAM→SOM logical disconnection'",
-     "marl":"Cite IDC report + bottom-up calculation via segment×ARPU. Restructured for investor verifiability."},
-    {"cat":"🧠 Emergent Question","q":"What is the fastest sort in Python?",
-     "raw":"QuickSort is the fastest at O(n log n).",
-     "tag":"S3 detects 'diverges from Python built-in implementation'",
-     "marl":"sorted() uses TimSort (hybrid), empirically faster than pure QuickSort."},
-    {"cat":"🔬 Precision Question","q":"What is the height of the Eiffel Tower?",
-     "raw":"The Eiffel Tower is 324m tall.",
-     "tag":"S4 detects 'antenna included/excluded not specified'",
-     "marl":"Structure 300m + broadcast antenna 24m = 324m total. Distinction matters by context."},
-    {"cat":"🔬 Precision Question","q":"How many light-minutes from Earth to the Sun?",
-     "raw":"About 8 minutes.",
-     "tag":"S3 detects 'elliptical orbit variation range missing'",
-     "marl":"Average 8 min 20 sec. Ranges from 8:10 (perihelion) to 8:27 (aphelion)."},
-]
 
+
+class LastBrain:
+    """3-stage MARL pipeline: S1 Hypothesis -> S2 Draft+Audit -> S5 Adversarial Refine"""
+
+    def __init__(self, call_fn: Callable, config: Optional[LastBrainConfig] = None):
+        self.call_fn = call_fn
+        self.config = config or LastBrainConfig()
+
+    def _get_seeds(self, prompt):
+        mode = self.config.emergence_type if self.config.mode == "emergence" else "insight"
+        seeds = MODE_SEEDS.get(mode, MODE_SEEDS["insight"])
+        return "\n".join(random.sample(seeds, min(2, len(seeds))))
+
+    def _detect_truncation(self, text):
+        if not text or len(text) < 20:
+            return False
+        if text[-1] == '\n':
+            return False
+        t = text.rstrip()
+        if len(t) < 10:
+            return False
+        return t[-1] not in '.!?)"\':\n'
+
+    def run(self, prompt, system_context=""):
+        start = time.time()
+        trace, stages_elapsed = {}, {}
+        full_prompt = f"[Context]\n{system_context}\n\n[Task]\n{prompt}" if system_context else prompt
+        seeds = self._get_seeds(prompt)
+
+        # S1: Hypothesis + Seeds
+        t1 = time.time()
+        mode_label = self.config.emergence_type.upper() if self.config.mode == "emergence" else "INSIGHT"
+        s1_out = self.call_fn(full_prompt, f"{S1_SYSTEM}\n\n[MODE: {mode_label}]\n[SEEDS]\n{seeds}",
+                              self.config.s1_budget, self.config.s1_temp)
+        trace["S1_Hypothesis"] = s1_out
+        stages_elapsed["S1"] = time.time() - t1
+
+        # S2: Draft + Inline Audit
+        t2 = time.time()
+        s2_ctx = f"[S1 ANALYSIS]\n{s1_out[:300]}\n\n" if s1_out and not s1_out.startswith("[ERROR") else ""
+        s2_out = self.call_fn(f"{s2_ctx}[TASK]\n{full_prompt}", S2_SYSTEM,
+                              self.config.s2_budget, self.config.s2_temp)
+        if self.config.auto_continue and self._detect_truncation(s2_out):
+            cont = self.call_fn(
+                f"You were writing but got CUT OFF. Last part:\n---\n{s2_out[-500:]}\n---\n"
+                f"CONTINUE from exactly where you stopped. Do NOT repeat.",
+                "You are S2_DraftAuditor continuing. Be concise.",
+                self.config.continue_budget, self.config.s2_temp)
+            if cont and not cont.startswith("[ERROR"):
+                s2_out = s2_out + "\n" + cont
+        trace["S2_DraftAudit"] = s2_out
+        stages_elapsed["S2"] = time.time() - t2
+
+        # S5: Adversarial Refine
+        t5 = time.time()
+        s2_compressed = s2_out[:1500]
+        if len(s2_out) > 1500:
+            s2_compressed += "\n[... draft truncated — write your OWN complete version]"
+        s5_prompt = (f"[ORIGINAL TASK]\n{prompt}\n\n"
+                     f"[S1 TRAPS]\n{s1_out[:200] if s1_out else 'none'}\n\n"
+                     f"[S2 DRAFT — reference only, write your OWN]\n{s2_compressed}")
+        s5_out = self.call_fn(s5_prompt, S5_SYSTEM, self.config.s5_budget, self.config.s5_temp)
+        if self.config.auto_continue and self._detect_truncation(s5_out):
+            cont = self.call_fn(
+                f"You were writing the FINAL ANSWER but got CUT OFF:\n---\n{s5_out[-500:]}\n---\n"
+                f"CONTINUE from exactly where you stopped. Complete ALL remaining items.",
+                "You are S5_AdversarialRefiner continuing. Clean final text only.",
+                self.config.continue_budget, self.config.s5_temp)
+            if cont and not cont.startswith("[ERROR"):
+                s5_out = s5_out + "\n" + cont
+        trace["S5_AdversarialRefine"] = s5_out
+        stages_elapsed["S5"] = time.time() - t5
+
+        answer = s5_out if s5_out and not s5_out.startswith("[ERROR") else s2_out
+        return LastBrainResult(answer=answer, raw_answer=s2_out, trace=trace,
+                               elapsed=time.time() - start, stages_elapsed=stages_elapsed)
+
+
+# ═══════════════════════════════════════════════════════════════
+# MODELS
+# ═══════════════════════════════════════════════════════════════
 MODELS = {
-    "OpenAI": {"env":"OPENAI_API_KEY","default":"gpt-5.4",
-        "list":["gpt-5.4","gpt-5.4-pro","gpt-5.2","gpt-4o","gpt-4o-mini"]},
-    "Anthropic": {"env":"ANTHROPIC_API_KEY","default":"claude-sonnet-4-6",
-        "list":["claude-opus-4-6","claude-sonnet-4-6","claude-haiku-4-5-20251001"]},
-    "Google Gemini": {"env":"GOOGLE_API_KEY","default":"gemini-2.5-pro",
-        "list":["gemini-2.5-pro","gemini-2.5-flash"]},
-    "DeepSeek": {"env":"DEEPSEEK_API_KEY","default":"deepseek-chat",
-        "list":["deepseek-chat","deepseek-reasoner"]},
-    "xAI (Grok)": {"env":"XAI_API_KEY","default":"grok-3-beta",
-        "list":["grok-3-beta"]},
-    "Ollama (Local)": {"env":"","default":"llama3.1",
-        "list":["llama3.1","llama3.1:70b","qwen3.5:32b","deepseek-r1:8b","phi4:14b"]},
-    "Custom (OpenAI-compatible)": {"env":"","default":"custom","list":["custom"]},
+    "google/gemma-3n-e4b-it": {
+        "name": "Gemma-3n-E4B", "params": "4B (2B active)", "ram": "~2GB",
+        "why": "Self-correction 89, Metacognition 90 — optimal for LastBrain",
+        "score": 87.5, "badge": "Most Balanced",
+    },
+    "Qwen/Qwen3-4B": {
+        "name": "Qwen3-4B", "params": "4B", "ram": "~2.8GB",
+        "why": "Trap detection 100, Math 100, Coding 100 — strongest reasoning",
+        "score": 86.9, "badge": "Best Reasoning",
+    },
+    "Qwen/Qwen3-1.7B": {
+        "name": "Qwen3-1.7B", "params": "1.7B", "ram": "~1.2GB",
+        "why": "Metacognition 90 at 1.7B — ultralight mobile champion",
+        "score": 76.8, "badge": "Lightest",
+    },
 }
-BACKEND_LIST = list(MODELS.keys())
-
-# ══════════════════════════════════════════════════════════════���═
-# MD → HTML
-# ════════════════════════════════════════════════════════════════
-
-def _esc(t):
-    return html_mod.escape(str(t)) if t else ""
-
-def _hex_rgb(h):
-    try: return f"{int(h[1:3],16)},{int(h[3:5],16)},{int(h[5:7],16)}"
-    except: return "99,102,241"
-
-def _inline_fmt(text):
-    t = text
-    t = re.sub(r'\*\*(.+?)\*\*', r'<b>\1</b>', t)
-    t = re.sub(r'\*(.+?)\*', r'<em>\1</em>', t)
-    t = re.sub(r'(\d{1,3})%', r'<span style="font-family:monospace;font-weight:600;color:#6366f1">\1%</span>', t)
-    return t
-
-def _md2html(text):
-    if not text: return ""
-    t = text
-    code_blocks = {}
-    def _save_code(m):
-        k = f"__CODE_{len(code_blocks)}__"
-        code = html_mod.escape(m.group(2))
-        code_blocks[k] = f'<pre style="background:rgba(15,23,42,.04);border:1px solid #e2e5f0;border-radius:8px;padding:12px;overflow-x:auto;font-family:monospace;font-size:11px;line-height:1.6;margin:8px 0"><code>{code}</code></pre>'
-        return k
-    t = re.sub(r'```(\w*)\n(.*?)```', _save_code, t, flags=re.DOTALL)
-    t = re.sub(r'`([^`]+)`', r'<code style="background:rgba(99,102,241,.08);padding:1px 5px;border-radius:4px;font-family:monospace;font-size:11px">\1</code>', t)
-
-    lines = t.split('\n')
-    result = []
-    in_list = False
-    for line in lines:
-        s = line.strip()
-        if s in code_blocks:
-            if in_list: result.append('</ul>'); in_list = False
-            result.append(code_blocks[s]); continue
-        hm = re.match(r'^(#{1,4})\s+(.+)$', s)
-        if hm:
-            if in_list: result.append('</ul>'); in_list = False
-            lvl = len(hm.group(1))
-            sz = {1:'18px',2:'15px',3:'13px',4:'12px'}[lvl]
-            result.append(f'<div style="font-size:{sz};font-weight:700;color:#1e293b;margin:14px 0 6px">{_inline_fmt(html_mod.escape(hm.group(2)))}</div>'); continue
-        if re.match(r'^[-*_]{3,}\s*$', s):
-            if in_list: result.append('</ul>'); in_list = False
-            result.append('<hr style="border:none;border-top:1px solid #e2e5f0;margin:12px 0">'); continue
-        lm = re.match(r'^[-*+]\s+(.+)$', s)
-        if lm:
-            if not in_list: result.append('<ul style="margin:6px 0;padding-left:20px">'); in_list = True
-            result.append(f'<li style="margin:3px 0;line-height:1.7">{_inline_fmt(html_mod.escape(lm.group(1)))}</li>'); continue
-        nm = re.match(r'^(\d+)[.)]\s+(.+)$', s)
-        if nm:
-            if in_list: result.append('</ul>'); in_list = False
-            result.append(f'<div style="margin:3px 0;padding-left:8px;line-height:1.7"><span style="color:#6366f1;font-weight:600;font-family:monospace;font-size:11px">{nm.group(1)}.</span> {_inline_fmt(html_mod.escape(nm.group(2)))}</div>'); continue
-        if in_list: result.append('</ul>'); in_list = False
-        if not s: result.append('<div style="height:6px"></div>'); continue
-        tm = re.match(r'^\[([A-Z_-]+(?:-\d+)?)\]\s*(.*)', s)
-        if tm:
-            tag = tm.group(1); rest = _inline_fmt(html_mod.escape(tm.group(2)))
-            tc = '#6366f1'
-            for px, cl in {'BACKTRACK':'#d97706','FIX':'#e11d48','APPLIED':'#16a34a','TRAP':'#e11d48','HALLUCINATION':'#e11d48','NO-FIXES':'#16a34a'}.items():
-                if tag.startswith(px): tc = cl; break
-            result.append(f'<div style="margin:6px 0;padding:6px 10px;background:rgba({_hex_rgb(tc)},.06);border-left:3px solid {tc};border-radius:0 6px 6px 0;font-size:12px;line-height:1.7"><span style="font-family:monospace;font-weight:700;color:{tc};font-size:11px">[{html_mod.escape(tag)}]</span> {rest}</div>'); continue
-        result.append(f'<p style="margin:4px 0;line-height:1.8">{_inline_fmt(html_mod.escape(s))}</p>')
-    if in_list: result.append('</ul>')
-    return '\n'.join(result)
-
-# ════════════════════════════════════════════════════════════════
-# Answer Cleaner — strip system tags, confidence %, metadata
-# ════════════════════════════════════════════════════════════════
-
-def _clean_answer(text):
-    """Strip reasoning artifacts from final answer for end-user display."""
-    if not text:
-        return ""
-    t = text
-
-    # Remove "--- Corrections ---" section and everything after
-    t = re.split(r'\n-{2,}\s*Corrections\s*-{2,}', t, maxsplit=1)[0]
-
-    lines = t.split('\n')
-    clean = []
-    for line in lines:
-        s = line.strip()
-        # Skip system tags: [FIX-n], [TRAP-CHECK], [HALLUCINATION], [APPLIED-n], [BACKTRACK-n], [NO-FIXES-NEEDED]
-        if re.match(r'^\[(?:FIX-\d+|TRAP-CHECK|HALLUCINATION|APPLIED-\d+|BACKTRACK-\d+|NO-FIXES-NEEDED)\]', s):
-            continue
-        # Skip stage labels: "S1 · Hypothesis Generator", "S2 · Primary Solver" etc.
-        if re.match(r'^S[1-5]\s*[·\-]', s):
-            continue
-        # Strip inline confidence: "(confidence: 85%)", "(90% confidence)", "Confidence: 85%"
-        line = re.sub(r'\(?\s*[Cc]onfidence[:\s]*\d{1,3}%\s*\)?', '', line)
-        line = re.sub(r'\(?\s*\d{1,3}%\s*confidence\s*\)?', '', line)
-        # Strip standalone confidence lines: "Confidence: 85%" or "**Confidence:** 90%"
-        if re.match(r'^\s*\*{0,2}[Cc]onfidence\*{0,2}\s*[:]\s*\d{1,3}%', line):
-            continue
-        # Strip "## Confidence Adjustments" section headers
-        if re.match(r'^#{1,4}\s*(?:Confidence|Top-\d+\s+Uncertaint)', s):
-            continue
-        # Strip "★ MANDATORY SELF-CHECK" and similar framework directives
-        if re.match(r'^★', s):
-            continue
-        clean.append(line)
-
-    # Clean up excess blank lines
-    result = '\n'.join(clean)
-    result = re.sub(r'\n{3,}', '\n\n', result).strip()
-    return result
-
-
-# ════════════════════════════════════════════════════════════════
-# HTML Renderers
-# ════════════════════════════════════════════════════════════════
-
-def _stage_html(name, content):
-    s = STAGES.get(name, {})
-    c = s.get("color","#6366f1")
-    body = _md2html(content[:3000] if content else "(no output)")
-    return f'<div style="border-left:3px solid {c};background:rgba({_hex_rgb(c)},.04);border-radius:0 10px 10px 0;padding:14px 16px;margin:8px 0"><div style="display:flex;align-items:center;gap:8px;margin-bottom:8px"><span style="font-size:1.1em">{s.get("icon","")}</span><span style="font-family:monospace;font-weight:700;font-size:12px;color:{c}">{s.get("label",name)}</span><span style="font-size:9px;color:#94a3b8;font-family:monospace;background:rgba(0,0,0,.04);padding:2px 8px;border-radius:10px">{s.get("tag","")}</span></div><div style="font-size:12px;line-height:1.7;color:#334155">{body}</div></div>'
-
-def _result_html(content, is_marl=False):
-    """Render Raw LLM result (no cleaning needed)."""
-    if not content:
-        return '<div style="color:#94a3b8;padding:40px;text-align:center;font-size:12px">Waiting...</div>'
-    badge = "MARL-Enhanced" if is_marl else "Raw LLM"
-    bc = "#6366f1" if is_marl else "#64748b"
-    bg = "rgba(99,102,241,.06)" if is_marl else "#f5f6fa"
-    body = _md2html(content)
-    return f'<div style="padding:16px"><div style="margin-bottom:12px"><span style="font-family:monospace;font-size:9px;font-weight:700;background:{bg};color:{bc};padding:3px 10px;border-radius:10px;text-transform:uppercase;letter-spacing:1px">{badge}</span></div><div style="font-size:13px;line-height:1.8;color:#1e293b;word-break:break-word">{body}</div></div>'
-
-def _marl_result_html(raw_answer, trace_dict):
-    """Render MARL result: clean answer + embedded reasoning toggle with S1~S5 trace."""
-    clean = _clean_answer(raw_answer)
-    if not clean:
-        return '<div style="color:#94a3b8;padding:40px;text-align:center;font-size:12px">Waiting...</div>'
-
-    body = _md2html(clean)
-
-    # Build S1~S5 trace HTML for toggle
-    trace_parts = ""
-    if trace_dict:
-        trace_parts = ''.join(_stage_html(n, trace_dict[n]) for n in STAGE_ORDER if n in trace_dict)
-
-    toggle_html = ""
-    if trace_parts:
-        toggle_html = f'''
-<details style="margin-top:16px;border:1px solid #e2e5f0;border-radius:10px;overflow:hidden">
-  <summary style="cursor:pointer;padding:10px 16px;background:rgba(99,102,241,.03);font-family:monospace;font-size:11px;font-weight:600;color:#6366f1;user-select:none;display:flex;align-items:center;gap:6px">
-    🔍 View Reasoning Process — S1→S2→S3→S4→S5
-  </summary>
-  <div style="padding:12px 16px;border-top:1px solid #e2e5f0;background:#fafbfc">
-    {trace_parts}
-  </div>
-</details>'''
-
-    return f'''<div style="padding:16px">
-  <div style="margin-bottom:12px">
-    <span style="font-family:monospace;font-size:9px;font-weight:700;background:rgba(99,102,241,.06);color:#6366f1;padding:3px 10px;border-radius:10px;text-transform:uppercase;letter-spacing:1px">MARL-Enhanced</span>
-  </div>
-  <div style="font-size:13px;line-height:1.8;color:#1e293b;word-break:break-word">{body}</div>
-  {toggle_html}
-</div>'''
-
-def _trace_html(trace_dict):
-    if not trace_dict:
-        return '<div style="color:#94a3b8;padding:30px;text-align:center">Run MARL to see pipeline trace</div>'
-    return ''.join(_stage_html(n, trace_dict[n]) for n in STAGE_ORDER if n in trace_dict)
-
-def _pipeline_anim(phase="marl", msg=""):
-    """Animated pipeline + rotating showcase Before/After cards."""
-    stages = [
-        ("S1", "Hypothesis", "🔍", "#0d9488"),
-        ("S2", "Solver", "⚡", "#6366f1"),
-        ("S3", "Auditor", "🛡️", "#d97706"),
-        ("S4", "Verifier", "🎯", "#e11d48"),
-        ("S5", "Synthesizer", "🧠", "#8b5cf6"),
-    ]
-    if phase == "raw":
-        return f'''<div style="padding:30px 20px;text-align:center">
-<div style="font-size:13px;color:#475569;margin-bottom:16px">{_esc(msg) if msg else "Generating Raw LLM response..."}</div>
-<div style="display:inline-block;width:200px;height:4px;background:#e2e5f0;border-radius:4px;overflow:hidden">
-<div style="width:40%;height:100%;background:linear-gradient(90deg,#6366f1,#0d9488);border-radius:4px;animation:rawPulse 1.5s ease-in-out infinite"></div>
-</div>
-<style>@keyframes rawPulse{{0%,100%{{width:30%}}50%{{width:70%}}}}</style>
-</div>'''
-
-    # ── Stage pills with sequential glow ──
-    pills = []
-    arrows = []
-    ns = len(stages)
-    for i, (sid, name, icon, color) in enumerate(stages):
-        delay = i * 1.8
-        rgb = _hex_rgb(color)
-        pills.append(f'''<div style="display:flex;flex-direction:column;align-items:center;gap:4px;animation:stageGlow {ns*1.8}s ease-in-out infinite;animation-delay:{delay}s;opacity:0.35">
-  <div style="width:44px;height:44px;border-radius:12px;display:flex;align-items:center;justify-content:center;font-size:18px;background:rgba({rgb},.1);border:2px solid rgba({rgb},.2)">{icon}</div>
-  <div style="font-family:monospace;font-size:10px;font-weight:700;color:{color}">{sid}</div>
-  <div style="font-size:9px;color:#94a3b8;white-space:nowrap">{name}</div>
-</div>''')
-        if i < ns - 1:
-            arrows.append(f'<div style="color:#cbd5e1;font-size:14px;margin-top:-16px;animation:arrowPulse {ns*1.8}s ease-in-out infinite;animation-delay:{delay+0.9}s;opacity:0.3">→</div>')
-
-    interleaved = []
-    for i, pill in enumerate(pills):
-        interleaved.append(pill)
-        if i < len(arrows):
-            interleaved.append(arrows[i])
-    stage_html = "\n".join(interleaved)
-    sub = _esc(msg) if msg else "Thinking, questioning, correcting, rewriting..."
-
-    # ── Showcase cards — CSS-only rotation ──
-    shuffled = list(SHOWCASE)
-    random.shuffle(shuffled)
-    nc = min(len(shuffled), 8)  # show up to 8 cards
-    dur_each = 5  # seconds per card
-    total_dur = nc * dur_each
-
-    cards_html = ""
-    card_kf = ""
-    for idx in range(nc):
-        ex = shuffled[idx]
-        pct_start = (idx / nc) * 100
-        pct_show  = pct_start + 2
-        pct_hide  = ((idx + 1) / nc) * 100 - 2
-        pct_end   = ((idx + 1) / nc) * 100
-
-        cards_html += f'''<div class="sc-card sc-card-{idx}" style="position:absolute;inset:0;opacity:0;animation:sc{idx} {total_dur}s ease-in-out infinite">
-  <div style="display:flex;align-items:center;gap:6px;margin-bottom:10px">
-    <span style="font-size:9px;font-weight:700;color:#6366f1;background:rgba(99,102,241,.08);padding:2px 8px;border-radius:6px;font-family:monospace">{_esc(ex['cat'])}</span>
-  </div>
-  <div style="font-size:13px;font-weight:700;color:#1e293b;margin-bottom:12px;line-height:1.4">"{_esc(ex['q'])}"</div>
-  <div style="display:flex;gap:10px">
-    <div style="flex:1;padding:10px 12px;background:#fef2f2;border:1px solid #fecaca;border-radius:8px">
-      <div style="font-size:9px;font-weight:700;color:#e11d48;margin-bottom:4px;font-family:monospace">❌ Non-MARL</div>
-      <div style="font-size:11px;color:#7f1d1d;line-height:1.5">{_esc(ex['raw'])}</div>
-    </div>
-    <div style="flex:1;padding:10px 12px;background:#f0fdf4;border:1px solid #bbf7d0;border-radius:8px">
-      <div style="font-size:9px;font-weight:700;color:#16a34a;margin-bottom:4px;font-family:monospace">✅ MARL</div>
-      <div style="font-size:11px;color:#14532d;line-height:1.5">{_esc(ex['marl'])}</div>
-    </div>
-  </div>
-  <div style="margin-top:8px;font-size:9px;color:#6366f1;font-family:monospace;font-weight:600">🔍 {_esc(ex['tag'])}</div>
-</div>\n'''
-
-        card_kf += f"@keyframes sc{idx}{{" \
-                   f"0%,{pct_start:.1f}%{{opacity:0;transform:translateY(8px)}}" \
-                   f"{pct_show:.1f}%{{opacity:1;transform:translateY(0)}}" \
-                   f"{pct_hide:.1f}%{{opacity:1;transform:translateY(0)}}" \
-                   f"{pct_end:.1f}%,100%{{opacity:0;transform:translateY(-8px)}}}}\n"
-
-    return f'''<div style="padding:24px 16px;text-align:center">
-<div style="display:flex;align-items:center;justify-content:center;gap:10px;flex-wrap:nowrap;margin-bottom:16px">
-{stage_html}
-</div>
-<div style="font-size:12px;color:#475569;margin-bottom:6px">{sub}</div>
-<div style="display:flex;gap:4px;justify-content:center;margin-bottom:20px">
-  <div style="width:6px;height:6px;border-radius:50%;background:#6366f1;animation:dotBounce 1.4s ease-in-out infinite"></div>
-  <div style="width:6px;height:6px;border-radius:50%;background:#0d9488;animation:dotBounce 1.4s ease-in-out .2s infinite"></div>
-  <div style="width:6px;height:6px;border-radius:50%;background:#8b5cf6;animation:dotBounce 1.4s ease-in-out .4s infinite"></div>
-</div>
-<div style="text-align:left;max-width:560px;margin:0 auto">
-  <div style="font-size:9px;font-weight:700;color:#94a3b8;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:10px;font-family:monospace">💡 Real cases caught by MARL</div>
-  <div style="position:relative;min-height:180px">
-    {cards_html}
-  </div>
-</div>
-<style>
-@keyframes stageGlow{{
-  0%,100%{{opacity:0.3;transform:scale(1)}}
-  {100/ns:.0f}%{{opacity:1;transform:scale(1.1)}}
-  {200/ns:.0f}%{{opacity:0.3;transform:scale(1)}}
-}}
-@keyframes arrowPulse{{
-  0%,100%{{opacity:0.2;color:#cbd5e1}}
-  {100/ns:.0f}%{{opacity:1;color:#6366f1}}
-  {200/ns:.0f}%{{opacity:0.2;color:#cbd5e1}}
-}}
-@keyframes dotBounce{{
-  0%,100%{{transform:translateY(0)}}
-  50%{{transform:translateY(-6px)}}
-}}
-{card_kf}
-</style>
-</div>'''
-
-def _status(state, msg, model, color):
-    dot = "●" if state == "Running" else "✓"
-    return f'<div style="display:flex;align-items:center;gap:10px;padding:10px 16px;background:rgba({_hex_rgb(color)},.06);border:1px solid rgba({_hex_rgb(color)},.15);border-radius:10px;font-family:monospace;font-size:11px"><span style="color:{color};font-weight:700">{dot} {state}</span><span style="color:#475569">{_esc(model)}</span><span style="color:#94a3b8">·</span><span style="color:#64748b">{_esc(msg)}</span></div>'
-
-# ════════════════════════════════════════════════════════════════
-# Build Marl
-# ════════════════════════════════════════════════════════════════
-
-def _on_backend(backend):
-    reg = MODELS.get(backend, {})
-    ml, dv, ek = reg.get("list",[]), reg.get("default",""), reg.get("env","")
-    return gr.Dropdown(choices=ml, value=dv), gr.Textbox(placeholder=f"ENV: {ek}" if ek else "API Key")
-
-def _build(backend, api_key, model, base_url):
-    if not MARL_OK:
-        return None, "❌ marl package failed to load. Check Space logs."
-    cfg = MarlConfig(include_trace=True, return_final_only=True)
-    reg = MODELS.get(backend, {})
-    model = model or reg.get("default","")
-    ek = reg.get("env","")
-    k = api_key or (os.getenv(ek,"") if ek else "")
 
+loaded_model = {"id": None, "tokenizer": None, "model": None}
+model_lock = Lock()
+
+
+# ═══════════════════════════════════════════════════════════════
+# GPU MODEL LOADING
+# ═══════════════════════════════════════════════════════════════
+def load_model(model_id, progress=gr.Progress()):
+    global loaded_model
+    if not model_id or model_id not in MODELS:
+        return "Select a model first."
+    with model_lock:
+        if loaded_model["id"] == model_id:
+            return f"Already loaded: {MODELS[model_id]['name']}"
+        progress(0.1, desc="Clearing GPU memory...")
+        if loaded_model["model"] is not None:
+            del loaded_model["model"]; del loaded_model["tokenizer"]
+            torch.cuda.empty_cache()
+        progress(0.3, desc=f"Loading {MODELS[model_id]['name']}...")
+        try:
+            tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+            mdl = AutoModelForCausalLM.from_pretrained(
+                model_id, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True)
+            loaded_model.update({"id": model_id, "tokenizer": tok, "model": mdl})
+            vram = torch.cuda.memory_allocated() / 1024**3
+            progress(1.0, desc="Done!")
+            return f"{MODELS[model_id]['name']} loaded | VRAM: {vram:.1f}GB"
+        except Exception as e:
+            loaded_model.update({"id": None, "tokenizer": None, "model": None})
+            return f"Failed: {str(e)[:200]}"
+
+
+# ═══════════════════════════════════════════════════════════════
+# LOCAL INFERENCE
+# ═══════════════════════════════════════════════════════════════
+def local_generate(prompt, system="", max_tokens=2048, temperature=0.5):
+    if loaded_model["model"] is None:
+        return "[ERROR] No model loaded"
+    tok, mdl = loaded_model["tokenizer"], loaded_model["model"]
+    messages = []
+    if system: messages.append({"role": "system", "content": system})
+    messages.append({"role": "user", "content": prompt})
+    try:
+        text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    except Exception:
+        text = f"{system}\n\n{prompt}" if system else prompt
+    inputs = tok(text, return_tensors="pt", truncation=True, max_length=4096).to(mdl.device)
+    with torch.no_grad():
+        outputs = mdl.generate(**inputs, max_new_tokens=max_tokens, temperature=max(temperature, 0.01),
+                               do_sample=True, top_p=0.9, repetition_penalty=1.1,
+                               pad_token_id=tok.eos_token_id)
+    return tok.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()
+
+
+# ═══════════════════════════════════════════════════════════════
+# GPT-5.4 JUDGE
+# ═══════════════════════════════════════════════════════════════
+JUDGE_PROMPT = """You are an expert AI evaluator. Score two AI responses to the same question.
+
+QUESTION:
+{question}
+
+RESPONSE A (Raw model, single call):
+{response_a}
+
+RESPONSE B (LastBrain, 3-stage pipeline):
+{response_b}
+
+Score each response on these criteria (0-100):
+1. Accuracy: Are facts correct? Any hallucination?
+2. Completeness: Are all aspects addressed?
+3. Self-awareness: Does it acknowledge uncertainty when appropriate?
+4. Reasoning depth: Is reasoning thorough and multi-layered?
+
+Respond ONLY with JSON:
+{{"score_a":{{"accuracy":N,"completeness":N,"self_awareness":N,"reasoning":N,"total":N}},"score_b":{{"accuracy":N,"completeness":N,"self_awareness":N,"reasoning":N,"total":N}},"winner":"A" or "B" or "TIE","reason":"one sentence"}}"""
+
+def judge_with_gpt(question, response_a, response_b, api_key):
+    if not api_key: return {"error": "OpenAI API key required for judging"}
+    import requests
     try:
-        if backend == "OpenAI":
-            if not k: return None, "❌ OPENAI_API_KEY required"
-            return Marl.from_openai(k, model, cfg), "✅"
-        elif backend == "Anthropic":
-            if not k: return None, "❌ ANTHROPIC_API_KEY required"
-            return Marl.from_anthropic(k, model, cfg), "✅"
-        elif backend == "Google Gemini":
-            k = k or os.getenv("GEMINI_API_KEY","")
-            if not k: return None, "❌ GOOGLE_API_KEY required"
-            return Marl.from_openai_compatible("https://generativelanguage.googleapis.com/v1beta/openai", k, model, cfg), "✅"
-        elif backend == "DeepSeek":
-            if not k: return None, "❌ DEEPSEEK_API_KEY required"
-            return Marl.from_openai_compatible("https://api.deepseek.com/v1", k, model, cfg), "✅"
-        elif backend == "xAI (Grok)":
-            if not k: return None, "❌ XAI_API_KEY required"
-            return Marl.from_openai_compatible("https://api.x.ai/v1", k, model, cfg), "✅"
-        elif backend == "Ollama (Local)":
-            return Marl.from_ollama(model, base_url or "http://localhost:11434", cfg), "✅"
-        elif backend == "Custom (OpenAI-compatible)":
-            if not base_url: return None, "❌ Base URL required"
-            return Marl.from_openai_compatible(base_url, api_key or "", model or "default", cfg), "✅"
+        r = requests.post("https://api.openai.com/v1/chat/completions",
+            headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+            json={"model": "gpt-5.4", "max_completion_tokens": 500, "temperature": 0,
+                  "messages": [{"role": "user", "content": JUDGE_PROMPT.format(
+                      question=question[:1000], response_a=response_a[:2000], response_b=response_b[:2000])}]},
+            timeout=30)
+        text = r.json()["choices"][0]["message"]["content"].strip().replace("```json","").replace("```","").strip()
+        return json.loads(text)
     except Exception as e:
-        return None, f"❌ Build error: {e}"
-    return None, "❌ Unsupported"
+        return {"error": str(e)[:200]}
 
-# ════════════════════════════════════════════════════════════════
-# A/B Test (Streaming)
-# ════════════════════════════════════════════════════════════════
 
-def run_ab_test(prompt, backend, api_key, model, base_url, budget, mode_sel, etype_sel):
-    if not prompt.strip():
-        yield ('<div style="color:#e11d48;padding:12px">❌ Enter a prompt</div>',"","","")
-        return
-    ml, st = _build(backend, api_key, model, base_url)
-    if not ml:
-        yield (f'<div style="color:#e11d48;padding:12px">{_esc(st)}</div>',"","","")
+# ═══════════════════════════════════════════════════════════════
+# A/B TEST
+# ═══════════════════════════════════════════════════════════════
+def run_ab_test(prompt, mode, etype, api_key, progress=gr.Progress()):
+    if loaded_model["model"] is None:
+        yield "Load a model first.", "", "", "", ""
         return
-    ml.config.budget_scale = float(budget)
-    # Set mode
-    _MODE_MAP = {"🔬 Insight": "insight", "🎨 Emergence": "emergence"}
-    _ETYPE_MAP = {"🔧 Invent": "invent", "✨ Create": "create", "🍳 Recipe": "recipe", "💊 Pharma": "pharma", "🧬 Genomics": "genomics", "🧪 Chemistry": "chemistry", "🌍 Ecology": "ecology", "⚖️ Law": "law", "📄 Document": "document"}
-    ml.config.mode = _MODE_MAP.get(mode_sel, "insight")
-    ml.config.emergence_type = _ETYPE_MAP.get(etype_sel, "invent")
-    mode_label = f"{mode_sel}" + (f" · {etype_sel}" if "Emergence" in mode_sel else "")
-
-    # Show both animations simultaneously
-    yield (_status("Running",f"{mode_label} · Running Raw LLM + MARL in parallel...",model,"#6366f1"),
-           _pipeline_anim("raw", "Generating Raw LLM response..."),
-           _pipeline_anim("marl", "Running MARL pipeline..."),
-           "")
-
-    # ── Parallel execution ──
-    from concurrent.futures import ThreadPoolExecutor
+    mn = MODELS.get(loaded_model["id"], {}).get("name", "?")
+
+    # Raw
+    progress(0.1, desc="[A] Raw generating...")
     t0 = time.time()
-    with ThreadPoolExecutor(max_workers=2) as pool:
-        future_raw  = pool.submit(ml.call_fn, prompt, "Answer thoroughly.", 4096, 0.6)
-        future_marl = pool.submit(ml.run, prompt)
-        raw  = future_raw.result()
-        r    = future_marl.result()
-    t_total = time.time() - t0
-
-    yield (_status("Complete",f"Parallel complete {t_total:.1f}s · {len(r.fixes)} corrections",model,"#16a34a"),
-           _result_html(raw,False), _marl_result_html(r.answer, r.trace), _trace_html(r.trace))
-
-def run_marl_only(prompt, backend, api_key, model, base_url, budget, mode_sel, etype_sel):
-    if not prompt.strip():
-        yield ('<div style="color:#e11d48;padding:12px">❌ Enter a prompt</div>',"","","")
-        return
-    ml, st = _build(backend, api_key, model, base_url)
-    if not ml:
-        yield (f'<div style="color:#e11d48;padding:12px">{_esc(st)}</div>',"","","")
-        return
-    ml.config.budget_scale = float(budget)
-    _MODE_MAP = {"🔬 Insight": "insight", "🎨 Emergence": "emergence"}
-    _ETYPE_MAP = {"🔧 Invent": "invent", "✨ Create": "create", "🍳 Recipe": "recipe", "💊 Pharma": "pharma", "🧬 Genomics": "genomics", "🧪 Chemistry": "chemistry", "🌍 Ecology": "ecology", "⚖️ Law": "law", "📄 Document": "document"}
-    ml.config.mode = _MODE_MAP.get(mode_sel, "insight")
-    ml.config.emergence_type = _ETYPE_MAP.get(etype_sel, "invent")
-    mode_label = f"{mode_sel}" + (f" · {etype_sel}" if "Emergence" in mode_sel else "")
-    yield (_status("Running",f"{mode_label} · MARL pipeline...",model,"#6366f1"),
-           "",_pipeline_anim("marl", "Running MARL pipeline..."),"")
-    t0=time.time(); r=ml.run(prompt); t_marl=time.time()-t0
-    yield (_status("Complete",f"MARL {t_marl:.1f}s · {len(r.fixes)} corrections",model,"#16a34a"),
-           "", _marl_result_html(r.answer, r.trace), _trace_html(r.trace))
-
-# ════════════════════════════════════════════════════════════════
-# Gradio App
-# ════════════════════════════════════════════════════════════════
-
-def create_app():
-    init_m = MODELS["OpenAI"]["list"]
-
-    with gr.Blocks(title="MARL — Model-Agnostic Runtime Middleware") as app:
-        gr.HTML(INDEX_HTML)
-
-        with gr.Tabs():
-          with gr.Tab("⚡ Playground"):
-            with gr.Row():
-                backend = gr.Dropdown(label="Backend", choices=BACKEND_LIST, value="OpenAI", scale=2)
-                api_key = gr.Textbox(label="API Key", type="password", placeholder="Enter your API key (required)",
-                                     value=os.getenv("OPENAI_API_KEY",""), scale=3)
-            with gr.Row():
-                model = gr.Dropdown(label="Model", choices=init_m, value="gpt-5.4",
-                                    allow_custom_value=True, scale=3)
-                base_url = gr.Textbox(label="Base URL (Custom/Ollama)",
-                                       placeholder="http://localhost:11434", scale=2)
-                budget = gr.Slider(0.3, 3.0, value=1.0, step=0.1, label="Budget Scale", scale=1)
-
-            with gr.Row():
-                mode = gr.Radio(["🔬 Insight", "🎨 Emergence"],
-                               value="🔬 Insight", label="Mode", scale=2)
-                etype = gr.Radio(["🔧 Invent", "✨ Create", "🍳 Recipe", "💊 Pharma", "🧬 Genomics", "🧪 Chemistry", "🌍 Ecology", "⚖️ Law", "📄 Document"],
-                                value="🔧 Invent", label="Emergence Engine", scale=2, visible=False)
-
-            def _on_mode(m):
-                return gr.Radio(visible="Emergence" in m)
-            mode.change(fn=_on_mode, inputs=[mode], outputs=[etype])
-
-            backend.change(fn=_on_backend, inputs=[backend], outputs=[model, api_key])
-
-            prompt = gr.Textbox(label="Prompt", placeholder="Enter your question or task...", lines=3)
-
-            EXAMPLES = [
-                ("🔬", "Is 0.9999... less than 1? Prove your answer with two different mathematical approaches.",
-                 "🔬 Insight", "🔧 Invent"),
-                ("🔬", "A startup claims their AI detects cancer with 99.9% accuracy from a selfie. As a medical advisor, evaluate this claim — what critical information is missing?",
-                 "🔬 Insight", "🔧 Invent"),
-                ("🔧", "Invent a device that allows dementia patients to live safely at home alone. Fuse sensors, AI, and UX — under $50/month. Identify the top 3 failure modes.",
-                 "🎨 Emergence", "🔧 Invent"),
-                ("🔧", "Design a building material that detects its own cracks and self-heals. What existing material science makes this feasible vs. science fiction?",
-                 "🎨 Emergence", "🔧 Invent"),
-                ("✨", "Write a single movie logline that would make both A24 and Marvel want to bid. Explain why the concept bridges arthouse and blockbuster.",
-                 "🎨 Emergence", "✨ Create"),
-                ("✨", "A museum wants to create an exhibit where visitors experience 'the feeling of forgetting.' Design the concept — what do they see, hear, and feel?",
-                 "🎨 Emergence", "✨ Create"),
-                ("🍳", "Can you truly replicate Korean beef bulgogi taste using only plant-based ingredients? Analyze the Maillard reaction chemistry and propose the closest possible recipe.",
-                 "🎨 Emergence", "🍳 Recipe"),
-                ("🍳", "A Michelin chef claims instant ramen can never be fine dining. Prove them wrong — design one dish that could change their mind, with the chemistry behind each choice.",
-                 "🎨 Emergence", "🍳 Recipe"),
-                ("📄", "Our company's turnover rate hit 30% this year. The CEO blames salary, but HR says it's culture. Analyze both hypotheses with data-driven counter-arguments.",
-                 "🎨 Emergence", "📄 Document"),
-                ("📄", "Write a policy brief arguing BOTH sides of whether governments should ban deepfake technology. Which side has the stronger evidence?",
-                 "🎨 Emergence", "📄 Document"),
-                ("💊", "Viagra was originally a heart drug. Identify ONE existing approved drug and build a rigorous case for repositioning it to treat Alzheimer's. Include mechanism, evidence gaps, and risks.",
-                 "🎨 Emergence", "💊 Pharma"),
-                ("💊", "A pharma company claims their new Alzheimer's drug reverses cognitive decline by 40%. What hidden assumptions in their clinical trial design should an FDA reviewer challenge?",
-                 "🎨 Emergence", "💊 Pharma"),
-                ("🧬", "BRCA-PARP synthetic lethality revolutionized cancer therapy. Propose ONE new synthetic lethality pair with biological rationale for why simultaneous inhibition would selectively kill cancer cells.",
-                 "🎨 Emergence", "🧬 Genomics"),
-                ("🧬", "A preprint claims gut microbiome directly causes Parkinson's disease. Evaluate the causal claim — what would a definitive study need to prove this beyond correlation?",
-                 "🎨 Emergence", "🧬 Genomics"),
-                ("🧪", "Is it physically possible to combine graphene-level strength with rubber-level flexibility in a single material? Analyze the trade-offs and propose the most feasible architecture.",
-                 "🎨 Emergence", "🧪 Chemistry"),
-                ("🧪", "A startup claims they can convert spent lithium batteries into solid-state battery materials at 90% efficiency. What are the thermodynamic limits they're likely ignoring?",
-                 "🎨 Emergence", "🧪 Chemistry"),
-                ("🌍", "An island nation is sinking due to climate change. They have $10M. Should they invest in sea walls, coral restoration, or relocation? Analyze the trade-offs with a 50-year horizon.",
-                 "🎨 Emergence", "🌍 Ecology"),
-                ("🌍", "Invasive lionfish are destroying Caribbean reefs. Can this threat be turned into a profitable industry? Analyze the ecological risks of commercializing an invasive species.",
-                 "🎨 Emergence", "🌍 Ecology"),
-                ("⚖️", "A self-driving car kills a pedestrian. Under EU law, the manufacturer is liable. Under US law, the software developer is. Under Korean law, it's unclear. Design a framework that resolves all three.",
-                 "🎨 Emergence", "⚖️ Law"),
-                ("⚖️", "An AI generates a novel that becomes a bestseller. The AI was trained on copyrighted books. Who owns the copyright? Analyze under common law vs. civil law and propose a new doctrine.",
-                 "🎨 Emergence", "⚖️ Law"),
-            ]
-            gr.HTML('<div style="font-size:10px;font-weight:700;color:#8b90b0;text-transform:uppercase;letter-spacing:1.5px;font-family:var(--mono);margin:12px 0 6px">💡 EXAMPLES — click to auto-fill prompt & mode</div>')
-            with gr.Row():
-                ex_btns = []
-                for i in range(5):
-                    icon, text = EXAMPLES[i][0], EXAMPLES[i][1]
-                    ex_btns.append(gr.Button(f"{icon} {text[:42]}...", size="sm", scale=1, min_width=60))
-            with gr.Row():
-                for i in range(5, 10):
-                    icon, text = EXAMPLES[i][0], EXAMPLES[i][1]
-                    ex_btns.append(gr.Button(f"{icon} {text[:42]}...", size="sm", scale=1, min_width=60))
-            with gr.Row():
-                for i in range(10, 15):
-                    icon, text = EXAMPLES[i][0], EXAMPLES[i][1]
-                    ex_btns.append(gr.Button(f"{icon} {text[:42]}...", size="sm", scale=1, min_width=60))
-            with gr.Row():
-                for i in range(15, 20):
-                    icon, text = EXAMPLES[i][0], EXAMPLES[i][1]
-                    ex_btns.append(gr.Button(f"{icon} {text[:42]}...", size="sm", scale=1, min_width=60))
-
-            for i, btn in enumerate(ex_btns):
-                _, ex_prompt, ex_mode, ex_etype = EXAMPLES[i]
-                is_emergence = "Emergence" in ex_mode
-                btn.click(fn=lambda p=ex_prompt, m=ex_mode, e=ex_etype, v=is_emergence: (p, m, gr.Radio(value=e, visible=v)),
-                          outputs=[prompt, mode, etype])
-
-            with gr.Row():
-                ab_btn = gr.Button("⚡ A/B Test · Raw LLM vs MARL", variant="primary", size="lg", scale=3)
-                marl_btn = gr.Button("🧠 MARL Only", variant="secondary", size="lg", scale=2)
-
-            status = gr.HTML()
-
-            gr.HTML('<div style="display:flex;gap:12px;margin:12px 0"><div style="flex:1;text-align:center;font-family:JetBrains Mono,monospace;font-size:10px;font-weight:700;letter-spacing:1.5px;text-transform:uppercase;color:#64748b;padding:10px;background:#f5f6fa;border-radius:10px 10px 0 0;border:1px solid #e2e5f0;border-bottom:2px solid #e2e5f0">🤖 A · Raw LLM</div><div style="flex:1;text-align:center;font-family:JetBrains Mono,monospace;font-size:10px;font-weight:700;letter-spacing:1.5px;text-transform:uppercase;color:#6366f1;padding:10px;background:rgba(99,102,241,.04);border-radius:10px 10px 0 0;border:1px solid rgba(99,102,241,.12);border-bottom:2px solid #6366f1">🧠 B · MARL-Enhanced</div></div>')
-
-            with gr.Row():
-                raw_out = gr.HTML()
-                marl_out = gr.HTML()
-
-            with gr.Accordion("📊 Pipeline Trace — 5-Stage Agent Outputs", open=False):
-                trace_out = gr.HTML()
-
-            ins = [prompt, backend, api_key, model, base_url, budget, mode, etype]
-            outs = [status, raw_out, marl_out, trace_out]
-            ab_btn.click(fn=run_ab_test, inputs=ins, outputs=outs)
-            marl_btn.click(fn=run_marl_only, inputs=ins, outputs=outs)
-
-          with gr.Tab("📦 Integration Guide"):
-            gr.HTML('''<div style="font-family:Sora,sans-serif;max-width:920px;margin:0 auto;padding:20px;color:#0f172a;line-height:1.7">
-
-<div style="display:grid;grid-template-columns:1fr 1fr;gap:12px;margin-bottom:24px">
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:8px;font-family:JetBrains Mono,monospace;color:#6366f1;text-transform:uppercase;letter-spacing:2px;font-weight:700;margin-bottom:6px">Quick Start</div>
-<pre style="margin:0;font-family:JetBrains Mono,monospace;font-size:14px;color:#6366f1;background:#f5f6fa;padding:10px;border-radius:10px;border:1px solid #e2e5f0"><code>pip install marl-middleware</code></pre>
-<div style="margin-top:6px;font-size:9px;color:#94a3b8">Linux x86_64 / Python 3.12 · Other OS → Docker</div>
-</div>
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:8px;font-family:JetBrains Mono,monospace;color:#0d9488;text-transform:uppercase;letter-spacing:2px;font-weight:700;margin-bottom:6px">Docker (All Platforms)</div>
-<pre style="margin:0;font-family:JetBrains Mono,monospace;font-size:14px;color:#0d9488;background:#f5f6fa;padding:10px;border-radius:10px;border:1px solid #e2e5f0"><code>docker run -p 8080:8080 vidraft/marl</code></pre>
-<div style="margin-top:6px;font-size:9px;color:#94a3b8">Mac · Windows · Linux — works everywhere</div>
-</div>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">⚡ 1-LINE INTEGRATION</div>
-<p style="color:#475569;font-size:11px;margin-bottom:10px">Add <b>one line</b> to any OpenAI-compatible app:</p>
-<pre style="background:#f5f6fa;padding:14px;border-radius:10px;font-family:JetBrains Mono,monospace;font-size:11px;line-height:1.8;border:1px solid #e2e5f0;overflow-x:auto"><code><span style="color:#94a3b8"># Before</span>
-client = OpenAI(api_key=<span style="color:#d97706">"sk-..."</span>)
-
-<span style="color:#94a3b8"># After — just add base_url</span>
-client = OpenAI(api_key=<span style="color:#d97706">"sk-..."</span>, <span style="color:#e11d48">base_url=</span><span style="color:#d97706">"http://localhost:8080/v1"</span>)</code></pre>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">🎨 9 EMERGENCE MODES</div>
-<p style="color:#475569;font-size:11px;margin-bottom:10px">Append <code style="background:rgba(99,102,241,.06);padding:2px 6px;border-radius:4px;color:#6366f1;font-family:JetBrains Mono,monospace;font-size:10px">::mode</code> to any model name:</p>
-<div style="overflow-x:auto">
-<table style="width:100%;border-collapse:separate;border-spacing:0;font-size:11px;border-radius:10px;overflow:hidden;border:1px solid #e2e5f0">
-<tr style="background:#f5f6fa"><th style="text-align:left;padding:8px 10px;font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;text-transform:uppercase;letter-spacing:.5px">model</th><th style="text-align:left;padding:8px 10px;font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;text-transform:uppercase;letter-spacing:.5px">Mode</th><th style="text-align:left;padding:8px 10px;font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;text-transform:uppercase;letter-spacing:.5px">Seeds</th></tr>
-<tr><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">gpt-5.2</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🔬 Insight (default)</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">Fact-check · Strategy</td></tr>
-<tr style="background:#fafbfe"><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::invent</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🔧 Invent</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">4,318 tech items</td></tr>
-<tr><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::create</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">✨ Create</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">493 seeds (11 categories)</td></tr>
-<tr style="background:#fafbfe"><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::recipe</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🍳 Recipe</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">131 methods · textures</td></tr>
-<tr><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::pharma</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">💊 Pharma</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">172 targets · mechanisms</td></tr>
-<tr style="background:#fafbfe"><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::genomics</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🧬 Genomics</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">104 genes · pathways</td></tr>
-<tr><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::chemistry</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🧪 Chemistry</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">135 elements · properties</td></tr>
-<tr style="background:#fafbfe"><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::ecology</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">🌍 Ecology</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">105 species · ecosystems</td></tr>
-<tr><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::law</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">⚖️ Law</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">59 jurisdictions</td></tr>
-<tr style="background:#fafbfe"><td style="padding:7px 10px;border-top:1px solid #e2e5f0"><code style="color:#6366f1;font-weight:600;font-family:JetBrains Mono,monospace;font-size:10px">::document</code></td><td style="padding:7px 10px;border-top:1px solid #e2e5f0">📄 Document</td><td style="padding:7px 10px;border-top:1px solid #e2e5f0;color:#94a3b8;font-size:10px">71 principles</td></tr>
-</table>
-</div>
-<p style="color:#94a3b8;font-size:9px;margin-top:6px;font-family:JetBrains Mono,monospace">Replace gpt-5.2 with any model — claude-sonnet, deepseek-v3, llama3, etc.</p>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">🐙 PYTHON SDK</div>
-<pre style="background:#f5f6fa;padding:14px;border-radius:10px;font-family:JetBrains Mono,monospace;font-size:11px;line-height:1.8;border:1px solid #e2e5f0;overflow-x:auto"><code><span style="color:#94a3b8"># OpenAI</span>
-<span style="color:#6366f1">from</span> marl <span style="color:#6366f1">import</span> Marl, MarlConfig
-ml = Marl.from_openai(<span style="color:#d97706">"sk-..."</span>, config=MarlConfig(
-    mode=<span style="color:#d97706">"emergence"</span>, emergence_type=<span style="color:#d97706">"create"</span>
-))
-result = ml.run(<span style="color:#d97706">"Generate 10 movie loglines"</span>)
-
-<span style="color:#94a3b8"># Anthropic</span>
-ml = Marl.from_anthropic(<span style="color:#d97706">"sk-ant-..."</span>)
-
-<span style="color:#94a3b8"># Ollama (local)</span>
-ml = Marl.from_ollama(<span style="color:#d97706">"llama3.1"</span>)
-
-<span style="color:#94a3b8"># Any OpenAI-compatible</span>
-ml = Marl.from_openai(<span style="color:#d97706">"sk-..."</span>, <span style="color:#d97706">"gpt-5.4"</span>)</code></pre>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">🦞 OPENCLAW INTEGRATION</div>
-<div style="display:grid;grid-template-columns:auto 1fr;gap:12px;align-items:start">
-<div style="background:linear-gradient(135deg,#6366f1,#4f46e5);color:#fff;width:28px;height:28px;border-radius:50%;display:flex;align-items:center;justify-content:center;font-weight:800;font-size:12px;font-family:JetBrains Mono,monospace;flex-shrink:0">1</div>
-<div><b style="font-size:11px;color:#0f172a">Install MARL</b><pre style="background:#f5f6fa;padding:8px 12px;border-radius:8px;font-family:JetBrains Mono,monospace;font-size:11px;margin:6px 0;border:1px solid #e2e5f0"><code style="color:#0d9488">docker run -p 8080:8080 vidraft/marl</code></pre></div>
-<div style="background:linear-gradient(135deg,#6366f1,#4f46e5);color:#fff;width:28px;height:28px;border-radius:50%;display:flex;align-items:center;justify-content:center;font-weight:800;font-size:12px;font-family:JetBrains Mono,monospace;flex-shrink:0">2</div>
-<div><b style="font-size:11px;color:#0f172a">Set config.json</b><pre style="background:#f5f6fa;padding:8px 12px;border-radius:8px;font-family:JetBrains Mono,monospace;font-size:11px;margin:6px 0;border:1px solid #e2e5f0"><code>{ <span style="color:#6366f1">"llm"</span>: { <span style="color:#6366f1">"baseURL"</span>: <span style="color:#d97706">"http://localhost:8080/v1"</span>, <span style="color:#6366f1">"model"</span>: <span style="color:#d97706">"gpt-5.2::create"</span> } }</code></pre></div>
-<div style="background:linear-gradient(135deg,#6366f1,#4f46e5);color:#fff;width:28px;height:28px;border-radius:50%;display:flex;align-items:center;justify-content:center;font-weight:800;font-size:12px;font-family:JetBrains Mono,monospace;flex-shrink:0">3</div>
-<div><b style="font-size:11px;color:#0f172a">Chat naturally</b><div style="color:#475569;font-size:10px;margin-top:4px">"Analyze this with MARL" · "Use MARL pharma mode for drug repositioning"</div></div>
-</div>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">🏗️ ARCHITECTURE</div>
-<pre style="background:#f5f6fa;padding:14px;border-radius:10px;font-family:JetBrains Mono,monospace;font-size:10px;line-height:1.7;border:1px solid #e2e5f0;overflow-x:auto"><code><span style="color:#0d9488">┌─ Your App ────���────────────────────────────────────┐</span>
-│  OpenClaw / Cursor / Custom App / Any LLM Client   │
-│  client = OpenAI(<span style="color:#e11d48">base_url=</span><span style="color:#d97706">"http://MARL:8080/v1"</span>)   │
-<span style="color:#0d9488">└────────────────────┬───────────────────────────────┘</span>
-                     │ HTTP (OpenAI API format)
-                     ▼
-<span style="color:#6366f1">┌─ MARL Middleware ──────────────────────────────────┐</span>
-│  S1 Hypothesis → S2 Solver → S3 Auditor            │
-│  → S4 Verifier → S5 Synthesizer                    │
-│  <span style="color:#d97706">9 Emergence Engines · 5,538 Seeds</span>                  │
-│  <span style="color:#16a34a">FINAL Bench: MA=0.694 vs ER=0.302 (70%+ ↑)</span>        │
-<span style="color:#6366f1">└────────────────────┬───────────────────────────────┘</span>
-                     │ API call (×5)
-                     ▼
-<span style="color:#94a3b8">┌─ Any LLM ──────────────────────────────────────────┐</span>
-│  OpenAI · Anthropic · Gemini · DeepSeek · Ollama   │
-<span style="color:#94a3b8">└────────────────────────────────────────────────────┘</span></code></pre>
-</div>
-
-<div style="background:var(--surface,#fff);border:1px solid #e2e5f0;border-radius:16px;padding:20px;margin-bottom:16px;box-shadow:0 1px 3px rgba(15,23,42,.04)">
-<div style="font-size:10px;font-family:JetBrains Mono,monospace;font-weight:700;color:#6366f1;text-transform:uppercase;letter-spacing:.8px;margin-bottom:10px">📡 SUPPORTED BACKENDS</div>
-<div style="overflow-x:auto">
-<table style="width:100%;border-collapse:separate;border-spacing:0;font-size:11px;border-radius:10px;overflow:hidden;border:1px solid #e2e5f0">
-<tr style="background:#f5f6fa"><th style="text-align:left;padding:8px 10px;font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;text-transform:uppercase;letter-spacing:.5px">Backend</th><th style="text-align:left;padding:8px 10px;font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;text-transform:uppercase;letter-spacing:.5px">Models</th></tr>
-<tr style="background:rgba(99,102,241,.06)"><td style="padding:8px 10px;border-top:1px solid #e2e5f0;font-weight:700;color:#6366f1">⭐ OpenAI (Default)</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0;font-weight:600">GPT-5.4, GPT-5.4-pro, GPT-5.2, GPT-4o</td></tr>
-<tr style="background:#fafbfe"><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Anthropic</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Claude Opus 4.6, Sonnet 4.6, Haiku 4.5</td></tr>
-<tr><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Google Gemini</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Gemini 2.5 Pro / Flash</td></tr>
-<tr style="background:#fafbfe"><td style="padding:8px 10px;border-top:1px solid #e2e5f0">DeepSeek</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">V3 / R1</td></tr>
-<tr><td style="padding:8px 10px;border-top:1px solid #e2e5f0">xAI</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Grok-3</td></tr>
-<tr style="background:#fafbfe"><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Ollama</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Llama, Mistral, Phi, Qwen</td></tr>
-<tr><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Custom</td><td style="padding:8px 10px;border-top:1px solid #e2e5f0">Any OpenAI-compatible endpoint</td></tr>
-</table>
-</div>
-</div>
-
-<div style="text-align:center;padding:16px;background:#f5f6fa;border:1px solid #e2e5f0;border-radius:10px">
-<p style="font-size:11px;color:#6366f1;margin:0;font-weight:700;font-family:JetBrains Mono,monospace">MARL · Model-Agnostic Runtime Middleware</p>
-<p style="font-size:9px;color:#94a3b8;margin:4px 0 0;font-family:JetBrains Mono,monospace">pip install marl-middleware · Apache 2.0 · <b style="color:#475569">VIDRAFT.net</b></p>
-</div>
-
-</div>''')
-
-        gr.HTML('<div style="text-align:center;padding:20px 0 8px;border-top:1px solid #e2e5f0;margin-top:16px"><p style="font-family:JetBrains Mono,monospace;font-size:8px;color:#94a3b8;letter-spacing:1px"><b style="color:#6366f1">MARL</b> · Model-Agnostic Runtime Middleware · S1→S2→S3→S4→S5 · Apache 2.0 · <b style="color:#475569">VIDRAFT.net</b></p></div>')
-
-    return app
-
-print("  Creating Gradio app...")
-try:
-    app = create_app()
-    print("  ✅ App created successfully")
-except Exception as e:
-    print(f"  ❌ App creation failed: {e}")
-    traceback.print_exc()
-    sys.exit(1)
+    raw = local_generate(prompt, "Answer thoroughly and accurately.", 2048, 0.5)
+    t_raw = time.time() - t0
+
+    # LastBrain
+    progress(0.4, desc="[B] LastBrain S1...")
+    cfg = LastBrainConfig(mode="emergence" if mode == "Emergence" else "insight",
+                          emergence_type=etype.lower() if etype else "invent")
+    lb = LastBrain(local_generate, cfg)
+    result = lb.run(prompt)
+    t_lb = result.elapsed
+
+    se = result.stages_elapsed
+    status = f"""<div style="padding:12px;background:linear-gradient(135deg,#0f172a,#1e293b);border-radius:10px;color:#f1f5f9;font-family:'JetBrains Mono',monospace;font-size:12px">
+<span style="color:#22d3ee;font-weight:700">LastBrain</span> <span style="color:#64748b">|</span> {mn}
+<span style="color:#64748b">|</span> Raw: <span style="color:#fbbf24">{t_raw:.1f}s</span>
+<span style="color:#64748b">|</span> LastBrain: <span style="color:#22d3ee">{t_lb:.1f}s</span>
+<span style="color:#64748b">|</span> S1: {se.get('S1',0):.1f}s S2: {se.get('S2',0):.1f}s S5: {se.get('S5',0):.1f}s
+<span style="color:#64748b">|</span> x{t_lb/max(t_raw,0.1):.1f}</div>"""
+
+    raw_html = f"""<div style="padding:16px;background:#fff;border-radius:10px;border:1px solid #e2e5f0">
+<div style="font-size:11px;color:#94a3b8;margin-bottom:8px;font-family:'JetBrains Mono',monospace">
+RAW {mn} | {t_raw:.1f}s | {len(raw.split())} words</div>
+<div style="white-space:pre-wrap;font-size:13px;line-height:1.7">{raw}</div></div>"""
+
+    marl_html = f"""<div style="padding:16px;background:#fff;border-radius:10px;border:2px solid #22d3ee">
+<div style="font-size:11px;color:#22d3ee;margin-bottom:8px;font-family:'JetBrains Mono',monospace">
+LASTBRAIN {mn} | {t_lb:.1f}s | {len(result.answer.split())} words</div>
+<div style="white-space:pre-wrap;font-size:13px;line-height:1.7">{result.answer}</div></div>"""
+
+    trace_parts = []
+    for stage, text in result.trace.items():
+        ts = se.get(stage.split("_")[0], 0)
+        trace_parts.append(f"""<details style="margin:4px 0"><summary style="cursor:pointer;font-weight:600;font-size:12px;padding:6px;background:#f8f9fc;border-radius:6px">
+{stage} ({ts:.1f}s)</summary>
+<pre style="white-space:pre-wrap;font-size:11px;padding:10px;background:#0f172a;color:#e2e8f0;border-radius:6px;max-height:300px;overflow-y:auto;margin-top:4px">{text}</pre></details>""")
+    trace_html = "\n".join(trace_parts)
+
+    yield status, raw_html, marl_html, trace_html, ""
+
+    # Judge
+    if api_key:
+        progress(0.8, desc="[Judge] GPT-5.4 evaluating...")
+        v = judge_with_gpt(prompt, raw, result.answer, api_key)
+        if "error" in v:
+            judge_html = f"<div style='color:#e11d48;padding:12px'>Judge error: {v['error']}</div>"
+        else:
+            sa, sb = v.get("score_a",{}), v.get("score_b",{})
+            w = v.get("winner","?"); reason = v.get("reason","")
+            wc = "#94a3b8" if w == "A" else "#22d3ee" if w == "B" else "#fbbf24"
+            wl = f"Raw {mn}" if w == "A" else "LastBrain" if w == "B" else "TIE"
+            wi = "A" if w == "A" else "B" if w == "B" else "="
+            judge_html = f"""<div style="padding:20px;background:linear-gradient(135deg,#0f172a,#1e293b);border-radius:12px;border:2px solid {wc}">
+<div style="text-align:center;font-size:24px;font-weight:800;color:{wc};margin-bottom:16px;font-family:'JetBrains Mono',monospace;letter-spacing:1px">
+WINNER: {wl} [{wi}]</div>
+<div style="display:grid;grid-template-columns:1fr 1fr;gap:16px;margin-bottom:16px">
+<div style="padding:14px;background:rgba(255,255,255,0.05);border-radius:8px;border:1px solid #334155">
+<div style="font-weight:700;color:#94a3b8;margin-bottom:8px;font-size:13px">A — Raw {mn}: {sa.get('total',0)}/100</div>
+<div style="font-size:11px;color:#64748b;line-height:1.6">
+Accuracy: {sa.get('accuracy',0)} | Completeness: {sa.get('completeness',0)}<br>
+Self-awareness: {sa.get('self_awareness',0)} | Reasoning: {sa.get('reasoning',0)}</div></div>
+<div style="padding:14px;background:rgba(34,211,238,0.08);border-radius:8px;border:1px solid #22d3ee">
+<div style="font-weight:700;color:#22d3ee;margin-bottom:8px;font-size:13px">B — LastBrain: {sb.get('total',0)}/100</div>
+<div style="font-size:11px;color:#64748b;line-height:1.6">
+Accuracy: {sb.get('accuracy',0)} | Completeness: {sb.get('completeness',0)}<br>
+Self-awareness: {sb.get('self_awareness',0)} | Reasoning: {sb.get('reasoning',0)}</div></div></div>
+<div style="text-align:center;font-size:11px;color:#64748b">{reason}</div>
+<div style="text-align:center;font-size:10px;color:#475569;margin-top:8px;font-family:'JetBrains Mono',monospace">
+Raw: {t_raw:.1f}s | LastBrain: {t_lb:.1f}s (x{t_lb/max(t_raw,0.1):.1f}) | Judge: GPT-5.4</div></div>"""
+        yield status, raw_html, marl_html, trace_html, judge_html
+    else:
+        yield status, raw_html, marl_html, trace_html, "<div style='padding:12px;color:#fbbf24;background:#0f172a;border-radius:8px;font-size:12px'>Enter OpenAI API key to enable GPT-5.4 judging.</div>"
+
+
+# ═══════════════════════════════════════════════════════════════
+# EXAMPLES
+# ═══════════════════════════════════════════════════════════════
+EXAMPLES = [
+    ["Is 0.9999... less than 1? Prove your answer.", "Insight", "invent"],
+    ["A startup claims 99.9% cancer detection from a selfie. Evaluate this claim.", "Insight", "invent"],
+    ["Can you replicate Korean bulgogi taste with only plant-based ingredients? Explain the chemistry.", "Emergence", "recipe"],
+    ["Identify ONE existing drug and build a case for repositioning it to treat Alzheimer's.", "Emergence", "pharma"],
+    ["Is it physically possible to combine graphene-level strength with rubber-level flexibility?", "Emergence", "chemistry"],
+    ["A self-driving car kills a pedestrian. Design a liability framework resolving EU, US, and Korean law.", "Emergence", "law"],
+    ["Write ONE movie logline that would make both A24 and Marvel want to bid.", "Emergence", "create"],
+    ["An island nation is sinking. $10M budget. Sea walls, coral, or relocation? 50-year analysis.", "Emergence", "ecology"],
+]
+
+
+# ═══════════════════════════════════════════════════════════════
+# UI
+# ═══════════════════════════════════════════════════════════════
+with gr.Blocks(
+    title="LastBrain: Survive everything.",
+    theme=gr.themes.Base(primary_hue="cyan", neutral_hue="slate"),
+    css="""
+    .gradio-container { max-width: 1200px !important; background: #0f172a !important; }
+    .main { background: #0f172a !important; }
+    body { background: #0f172a !important; }
+    .dark { background: #0f172a !important; }
+    footer { display: none !important; }
+    """
+) as app:
+
+    # Header
+    gr.HTML("""<div style="text-align:center;padding:32px 0 16px">
+<h1 style="font-size:42px;font-weight:900;margin:0;letter-spacing:-1px;
+background:linear-gradient(135deg,#22d3ee,#6366f1,#a855f7);-webkit-background-clip:text;-webkit-text-fill-color:transparent">
+LastBrain</h1>
+<p style="color:#22d3ee;font-size:14px;font-weight:600;letter-spacing:4px;text-transform:uppercase;margin:4px 0">
+Survive everything.</p>
+<p style="color:#475569;font-size:11px;margin:8px 0;font-family:'JetBrains Mono',monospace">
+3-Stage Metacognitive Pipeline for Small Models | Raw vs LastBrain | Judged by GPT-5.4</p>
+</div>""")
+
+    # Model selection
+    gr.HTML("<div style='color:#64748b;font-size:10px;text-transform:uppercase;letter-spacing:2px;padding:0 4px;margin-bottom:4px;font-weight:700'>Select Model</div>")
+    with gr.Row():
+        with gr.Column(scale=3):
+            model_dd = gr.Dropdown(choices=list(MODELS.keys()), label="Model", container=False)
+        with gr.Column(scale=1):
+            load_btn = gr.Button("Load on GPU", variant="primary")
+    load_status = gr.Textbox(label="Status", interactive=False, lines=1)
+    model_info = gr.HTML("")
+
+    def show_info(mid):
+        if not mid or mid not in MODELS: return ""
+        m = MODELS[mid]
+        return f"""<div style="padding:10px 14px;background:#1e293b;border-radius:8px;border-left:3px solid #22d3ee;margin:4px 0;font-size:12px;color:#cbd5e1">
+<b style="color:#22d3ee">{m['name']}</b>
+<span style="color:#475569">|</span> {m['params']}
+<span style="color:#475569">|</span> RAM: {m['ram']}
+<span style="color:#475569">|</span> Score: <b style="color:#fbbf24">{m['score']}</b>
+<span style="color:#475569">|</span> <span style="background:#22d3ee;color:#0f172a;padding:1px 6px;border-radius:4px;font-size:10px;font-weight:700">{m['badge']}</span>
+<br><span style="color:#64748b">{m['why']}</span></div>"""
+
+    model_dd.change(fn=show_info, inputs=[model_dd], outputs=[model_info])
+    load_btn.click(fn=load_model, inputs=[model_dd], outputs=[load_status])
+
+    # Prompt
+    gr.HTML("<hr style='margin:16px 0;border-color:#1e293b'>")
+    with gr.Row():
+        with gr.Column(scale=3):
+            prompt = gr.Textbox(label="Prompt", placeholder="Ask anything...", lines=3)
+        with gr.Column(scale=1):
+            mode = gr.Radio(["Insight", "Emergence"], value="Insight", label="Mode")
+            etype = gr.Dropdown(
+                ["invent","create","recipe","pharma","genomics","chemistry","ecology","law","document"],
+                value="invent", label="Engine", visible=False)
+            mode.change(fn=lambda m: gr.Dropdown(visible=m=="Emergence"), inputs=[mode], outputs=[etype])
+
+    api_key = gr.Textbox(label="OpenAI API Key (GPT-5.4 Judge)", type="password",
+                         placeholder="sk-... (recommended)", value=os.getenv("OPENAI_API_KEY",""))
+    run_btn = gr.Button("Run A/B Test", variant="primary", size="lg")
+
+    gr.Examples(examples=EXAMPLES, inputs=[prompt, mode, etype], label="Examples")
+
+    # Output
+    status_out = gr.HTML()
+    with gr.Row():
+        with gr.Column():
+            gr.HTML("<div style='text-align:center;font-weight:700;color:#94a3b8;font-size:13px;padding:8px'>A — Raw Model</div>")
+            raw_out = gr.HTML()
+        with gr.Column():
+            gr.HTML("<div style='text-align:center;font-weight:700;color:#22d3ee;font-size:13px;padding:8px'>B — LastBrain</div>")
+            marl_out = gr.HTML()
+
+    judge_out = gr.HTML()
+
+    with gr.Accordion("Reasoning Trace (S1 → S2 → S5)", open=False):
+        trace_out = gr.HTML()
+
+    run_btn.click(fn=run_ab_test, inputs=[prompt, mode, etype, api_key],
+                  outputs=[status_out, raw_out, marl_out, trace_out, judge_out])
+
+    gr.HTML(f"""<div style="text-align:center;padding:20px;margin-top:16px">
+<span style="color:#22d3ee;font-weight:800;font-size:14px;letter-spacing:2px">LastBrain</span>
+<span style="color:#475569;font-size:11px"> — Survive everything.</span><br>
+<span style="color:#334155;font-size:10px;font-family:'JetBrains Mono',monospace">
+S1 Hypothesis (256t) → S2 Draft+Audit (2048t) → S5 Adversarial Refine (2048t)<br>
+<a href="https://vidraft.net" style="color:#475569">VIDRAFT</a> ·
+<a href="https://github.com/Vidraft/MARL" style="color:#475569">GitHub</a> ·
+<a href="https://pypi.org/project/marl-middleware/" style="color:#475569">PyPI</a> ·
+<a href="https://clawhub.ai/Cutechicken99/marl-middleware" style="color:#475569">ClawHub</a>
+</span></div>""")
 
 if __name__ == "__main__":
-    print("  🚀 Launching on 0.0.0.0:7860 ...")
-    try:
-        app.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)
-    except TypeError:
-        # ssr_mode not supported in this gradio version
-        app.launch(server_name="0.0.0.0", server_port=7860)
+    app.launch(server_name="0.0.0.0", server_port=7860)