aimodelshare 0.3.7__py3-none-any.whl → 0.4.71__py3-none-any.whl

aimodelshare/moral_compass/apps/model_building_app_es_final.py

@@ -0,0 +1,3899 @@
+"""
+Model Building Game - Gradio application for the Justice & Equity Challenge.
+
+Session-based authentication with leaderboard caching and progressive rank unlocking.
+
+Concurrency Notes:
+- This app is designed to run in a multi-threaded environment (Cloud Run).
+- Per-user state is stored in gr.State objects, NOT in os.environ.
+- Caches are protected by locks to ensure thread safety.
+- Linear algebra libraries are constrained to single-threaded mode to prevent
+  CPU oversubscription in containerized deployments.
+"""
+
+import os
+
+# -------------------------------------------------------------------------
+# Thread Limit Configuration (MUST be set before importing numpy/sklearn)
+# Prevents CPU oversubscription in containerized environments like Cloud Run.
+# -------------------------------------------------------------------------
+os.environ.setdefault("OMP_NUM_THREADS", "1")
+os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
+os.environ.setdefault("MKL_NUM_THREADS", "1")
+os.environ.setdefault("NUMEXPR_NUM_THREADS", "1")
+
+import time
+import random
+import requests
+import contextlib
+from io import StringIO
+import threading
+import functools
+from pathlib import Path
+from datetime import datetime, timedelta
+from typing import Optional, Dict, Any, Tuple, Callable, TypeVar
+
+import numpy as np
+import pandas as pd
+import gradio as gr
+
+# --- Scikit-learn Imports ---
+from sklearn.model_selection import train_test_split
+from sklearn.preprocessing import StandardScaler
+from sklearn.impute import SimpleImputer
+from sklearn.compose import ColumnTransformer
+from sklearn.pipeline import Pipeline
+from sklearn.preprocessing import OneHotEncoder
+from sklearn.linear_model import LogisticRegression
+from sklearn.tree import DecisionTreeClassifier
+from sklearn.ensemble import RandomForestClassifier
+from sklearn.neighbors import KNeighborsClassifier
+
+# --- AI Model Share Imports ---
+try:
+    from aimodelshare.playground import Competition
+except ImportError:
+    raise ImportError(
+        "The 'aimodelshare' library is required. Install with: pip install aimodelshare"
+    )
+
+# -------------------------------------------------------------------------
+# Configuration & Caching Infrastructure
+# -------------------------------------------------------------------------
+
+
+# -------------------------------------------------------------------------
+# CACHE CONFIGURATION (Optimized: Thread-Safe SQLite)
+# -------------------------------------------------------------------------
+import sqlite3
+
+CACHE_DB_FILE = "prediction_cache.sqlite"
+
+def get_cached_prediction(key):
+    """
+    Lightning-fast lookup from SQLite database.
+    THREAD-SAFE FIX: Opens a new connection for every lookup.
+    """
+    # 1. Check if DB exists
+    if not os.path.exists(CACHE_DB_FILE):
+        return None
+
+    try:
+        # Use a context manager ('with') to ensure the connection 
+        # is ALWAYS closed, releasing file locks immediately.
+        # timeout=10 ensures we don't wait forever if the file is busy.
+        with sqlite3.connect(CACHE_DB_FILE, timeout=10.0) as conn:
+            cursor = conn.cursor()
+            cursor.execute("SELECT value FROM cache WHERE key=?", (key,))
+            result = cursor.fetchone()
+            
+            if result:
+                return result[0] 
+            else:
+                return None
+            
+    except sqlite3.OperationalError as e:
+        # Handle locking errors gracefully
+        print(f"⚠️ CACHE LOCK ERROR: {e}. Falling back to training.", flush=True)
+        return None
+        
+    except Exception as e:
+        print(f"⚠️ DB READ ERROR: {e}", flush=True)
+        return None
+
+print("✅ App configured for Thread-Safe SQLite Cache.")
+
+
+LEADERBOARD_CACHE_SECONDS = int(os.environ.get("LEADERBOARD_CACHE_SECONDS", "45"))
+MAX_LEADERBOARD_ENTRIES = os.environ.get("MAX_LEADERBOARD_ENTRIES")
+MAX_LEADERBOARD_ENTRIES = int(MAX_LEADERBOARD_ENTRIES) if MAX_LEADERBOARD_ENTRIES else None
+DEBUG_LOG = os.environ.get("DEBUG_LOG", "false").lower() == "true"
+
+# In-memory caches (per container instance)
+# Each cache has its own lock for thread safety under concurrent requests
+_cache_lock = threading.Lock()  # Protects _leaderboard_cache
+_user_stats_lock = threading.Lock()  # Protects _user_stats_cache
+_auth_lock = threading.Lock()  # Protects get_aws_token() credential injection
+
+# Auth-aware leaderboard cache: separate entries for authenticated vs anonymous
+# Structure: {"anon": {"data": df, "timestamp": float}, "auth": {"data": df, "timestamp": float}}
+_leaderboard_cache: Dict[str, Dict[str, Any]] = {
+    "anon": {"data": None, "timestamp": 0.0},
+    "auth": {"data": None, "timestamp": 0.0},
+}
+_user_stats_cache: Dict[str, Dict[str, Any]] = {}
+USER_STATS_TTL = LEADERBOARD_CACHE_SECONDS
+
+# -------------------------------------------------------------------------
+# Retry Helper for External API Calls
+# -------------------------------------------------------------------------
+
+T = TypeVar("T")
+
+def _retry_with_backoff(
+    func: Callable[[], T],
+    max_attempts: int = 3,
+    base_delay: float = 0.5,
+    description: str = "operation"
+) -> T:
+    """
+    Execute a function with exponential backoff retry on failure.
+    
+    Concurrency Note: This helper provides resilience against transient
+    network failures when calling external APIs (Competition.get_leaderboard,
+    playground.submit_model). Essential for Cloud Run deployments where
+    network calls may occasionally fail under load.
+    
+    Args:
+        func: Callable to execute (should take no arguments)
+        max_attempts: Maximum number of attempts (default: 3)
+        base_delay: Initial delay in seconds, doubled each retry (default: 0.5)
+        description: Human-readable description for logging
+    
+    Returns:
+        Result from successful function call
+    
+    Raises:
+        Last exception if all attempts fail
+    """
+    last_exception: Optional[Exception] = None
+    delay = base_delay
+    
+    for attempt in range(1, max_attempts + 1):
+        try:
+            return func()
+        except Exception as e:
+            last_exception = e
+            if attempt < max_attempts:
+                _log(f"{description} attempt {attempt} failed: {e}. Retrying in {delay}s...")
+                time.sleep(delay)
+                delay *= 2  # Exponential backoff
+            else:
+                _log(f"{description} failed after {max_attempts} attempts: {e}")
+    
+    # Loop always runs at least once (max_attempts >= 1), so last_exception is set
+    raise last_exception  # type: ignore[misc]
+
+def _log(msg: str):
+    """Log message if DEBUG_LOG is enabled."""
+    if DEBUG_LOG:
+        print(f"[ModelBuildingGame] {msg}")
+
+def _normalize_team_name(name: str) -> str:
+    """Normalize team name for consistent comparison and storage."""
+    if not name:
+        return ""
+    return " ".join(str(name).strip().split())
+
+def _get_leaderboard_with_optional_token(playground_instance: Optional["Competition"], token: Optional[str] = None) -> Optional[pd.DataFrame]:
+    """
+    Fetch fresh leaderboard with optional token authentication and retry logic.
+    
+    This is a helper function that centralizes the pattern of fetching
+    a fresh (non-cached) leaderboard with optional token authentication.
+    Use this for user-facing flows that require fresh, full data.
+    
+    Concurrency Note: Uses _retry_with_backoff for resilience against
+    transient network failures.
+    
+    Args:
+        playground_instance: The Competition playground instance (or None)
+        token: Optional authentication token for the fetch
+    
+    Returns:
+        DataFrame with leaderboard data, or None if fetch fails or playground is None
+    """
+    if playground_instance is None:
+        return None
+    
+    def _fetch():
+        if token:
+            return playground_instance.get_leaderboard(token=token)
+        return playground_instance.get_leaderboard()
+    
+    try:
+        return _retry_with_backoff(_fetch, description="leaderboard fetch")
+    except Exception as e:
+        _log(f"Leaderboard fetch failed after retries: {e}")
+        return None
+
+def _fetch_leaderboard(token: Optional[str]) -> Optional[pd.DataFrame]:
+    """
+    Fetch leaderboard with auth-aware caching (TTL: LEADERBOARD_CACHE_SECONDS).
+    
+    Concurrency Note: Cache is keyed by auth scope ("anon" vs "auth") to prevent
+    cross-user data leakage. Authenticated users share a single "auth" cache entry
+    to avoid unbounded cache growth. Protected by _cache_lock.
+    """
+    # Determine cache key based on authentication status
+    cache_key = "auth" if token else "anon"
+    now = time.time()
+    
+    with _cache_lock:
+        cache_entry = _leaderboard_cache[cache_key]
+        if (
+            cache_entry["data"] is not None
+            and now - cache_entry["timestamp"] < LEADERBOARD_CACHE_SECONDS
+        ):
+            _log(f"Leaderboard cache hit ({cache_key})")
+            return cache_entry["data"]
+
+    _log(f"Fetching fresh leaderboard ({cache_key})...")
+    df = None
+    try:
+        playground_id = "https://cf3wdpkg0d.execute-api.us-east-1.amazonaws.com/prod/m"
+        playground_instance = Competition(playground_id)
+        
+        def _fetch():
+            return playground_instance.get_leaderboard(token=token) if token else playground_instance.get_leaderboard()
+        
+        df = _retry_with_backoff(_fetch, description="leaderboard fetch")
+        if df is not None and not df.empty and MAX_LEADERBOARD_ENTRIES:
+            df = df.head(MAX_LEADERBOARD_ENTRIES)
+        _log(f"Leaderboard fetched ({cache_key}): {len(df) if df is not None else 0} entries")
+    except Exception as e:
+        _log(f"Leaderboard fetch failed ({cache_key}): {e}")
+        df = None
+
+    with _cache_lock:
+        _leaderboard_cache[cache_key]["data"] = df
+        _leaderboard_cache[cache_key]["timestamp"] = time.time()
+    return df
+
+def _get_or_assign_team(username: str, leaderboard_df: Optional[pd.DataFrame]) -> Tuple[str, bool]:
+    """Get existing team from leaderboard or assign random team."""
+    # TEAM_NAMES is defined in configuration section below
+    try:
+        if leaderboard_df is not None and not leaderboard_df.empty and "Team" in leaderboard_df.columns:
+            user_submissions = leaderboard_df[leaderboard_df["username"] == username]
+            if not user_submissions.empty:
+                if "timestamp" in user_submissions.columns:
+                    try:
+                        user_submissions = user_submissions.copy()
+                        user_submissions["timestamp"] = pd.to_datetime(
+                            user_submissions["timestamp"], errors="coerce"
+                        )
+                        user_submissions = user_submissions.sort_values("timestamp", ascending=False)
+                        _log(f"Sorted {len(user_submissions)} submissions by timestamp for {username}")
+                    except Exception as ts_err:
+                        _log(f"Timestamp sort error: {ts_err}")
+                existing_team = user_submissions.iloc[0]["Team"]
+                if pd.notna(existing_team) and str(existing_team).strip():
+                    normalized = _normalize_team_name(existing_team)
+                    _log(f"Found existing team for {username}: {normalized}")
+                    return normalized, False
+        new_team = _normalize_team_name(random.choice(TEAM_NAMES))
+        _log(f"Assigning new team to {username}: {new_team}")
+        return new_team, True
+    except Exception as e:
+        _log(f"Team assignment error: {e}")
+        new_team = _normalize_team_name(random.choice(TEAM_NAMES))
+        return new_team, True
+
+def _try_session_based_auth(request: "gr.Request") -> Tuple[bool, Optional[str], Optional[str]]:
+    """Attempt to authenticate via session token. Returns (success, username, token)."""
+    try:
+        session_id = request.query_params.get("sessionid") if request else None
+        if not session_id:
+            _log("No sessionid in request")
+            return False, None, None
+        
+        from aimodelshare.aws import get_token_from_session, _get_username_from_token
+        
+        token = get_token_from_session(session_id)
+        if not token:
+            _log("Failed to get token from session")
+            return False, None, None
+            
+        username = _get_username_from_token(token)
+        if not username:
+            _log("Failed to extract username from token")
+            return False, None, None
+        
+        _log(f"Session auth successful for {username}")
+        return True, username, token
+        
+    except Exception as e:
+        _log(f"Session auth failed: {e}")
+        return False, None, None
+
+def _compute_user_stats(username: str, token: str) -> Dict[str, Any]:
+    """
+    Compute user statistics with caching.
+    
+    Concurrency Note: Protected by _user_stats_lock for thread-safe
+    cache reads and writes.
+    """
+    now = time.time()
+    
+    # Thread-safe cache check
+    with _user_stats_lock:
+        cached = _user_stats_cache.get(username)
+        if cached and (now - cached.get("_ts", 0) < USER_STATS_TTL):
+            _log(f"User stats cache hit for {username}")
+            # Return shallow copy to prevent caller mutations from affecting cache.
+            # Stats dict contains only primitives (float, int, str), so shallow copy is sufficient.
+            return cached.copy()
+
+    _log(f"Computing fresh stats for {username}")
+    leaderboard_df = _fetch_leaderboard(token)
+    team_name, _ = _get_or_assign_team(username, leaderboard_df)
+    
+    stats = {
+        "best_score": 0.0,
+        "rank": 0,
+        "team_name": team_name,
+        "submission_count": 0,
+        "last_score": 0.0,
+        "_ts": time.time()
+    }
+
+    try:
+        if leaderboard_df is not None and not leaderboard_df.empty:
+            user_submissions = leaderboard_df[leaderboard_df["username"] == username]
+            if not user_submissions.empty:
+                stats["submission_count"] = len(user_submissions)
+                if "accuracy" in user_submissions.columns:
+                    stats["best_score"] = float(user_submissions["accuracy"].max())
+                    if "timestamp" in user_submissions.columns:
+                        try:
+                            user_submissions = user_submissions.copy()
+                            user_submissions["timestamp"] = pd.to_datetime(
+                                user_submissions["timestamp"], errors="coerce"
+                            )
+                            recent = user_submissions.sort_values("timestamp", ascending=False).iloc[0]
+                            stats["last_score"] = float(recent["accuracy"])
+                        except:
+                            stats["last_score"] = stats["best_score"]
+                    else:
+                        stats["last_score"] = stats["best_score"]
+            
+            if "accuracy" in leaderboard_df.columns:
+                user_bests = leaderboard_df.groupby("username")["accuracy"].max()
+                ranked = user_bests.sort_values(ascending=False)
+                try:
+                    stats["rank"] = int(ranked.index.get_loc(username) + 1)
+                except KeyError:
+                    stats["rank"] = 0
+    except Exception as e:
+        _log(f"Error computing stats for {username}: {e}")
+
+    # Thread-safe cache update
+    with _user_stats_lock:
+        _user_stats_cache[username] = stats
+    _log(f"Stats for {username}: {stats}")
+    return stats
+def _build_attempts_tracker_html(current_count, limit=10):
+    """
+    Generate HTML for the attempts tracker display.
+    Shows current attempt count vs limit with color coding.
+    
+    Args:
+        current_count: Number of attempts used so far
+        limit: Maximum allowed attempts (default: ATTEMPT_LIMIT)
+    
+    Returns:
+        str: HTML string for the tracker display
+    """
+    if current_count >= limit:
+        # Limit reached - red styling
+        bg_color = "#f0f9ff"
+        border_color = "#bae6fd"
+        text_color = "#0369a1"
+        icon = "🛑"
+        label = f"Last chance (for now) to boost your score!: {current_count}/{limit}"
+    else:
+        # Normal - blue styling
+        bg_color = "#f0f9ff"
+        border_color = "#bae6fd"
+        text_color = "#0369a1"
+        icon = "📊"
+        label = f"Attempts used: {current_count}/{limit}"
+
+    return f"""<div style='text-align:center; padding:8px; margin:8px 0; background:{bg_color}; border-radius:8px; border:1px solid {border_color};'>
+        <p style='margin:0; color:{text_color}; font-weight:600; font-size:1rem;'>{icon} {label}</p>
+    </div>"""
+    
+def check_attempt_limit(submission_count: int, limit: int = None) -> Tuple[bool, str]:
+    """Check if submission count exceeds limit."""
+    # ATTEMPT_LIMIT is defined in configuration section below
+    if limit is None:
+        limit = ATTEMPT_LIMIT
+    
+    if submission_count >= limit:
+        msg = f"⚠️ Attempt limit reached ({submission_count}/{limit})"
+        return False, msg
+    return True, f"Attempts: {submission_count}/{limit}"
+
+# -------------------------------------------------------------------------
+# Future: Fairness Metrics
+# -------------------------------------------------------------------------
+
+# def compute_fairness_metrics(y_true, y_pred, sensitive_attrs):
+#     """
+#     Compute fairness metrics for model predictions.
+#     
+#     Args:
+#         y_true: Ground truth labels
+#         y_pred: Model predictions
+#         sensitive_attrs: DataFrame with sensitive attributes (race, sex, age)
+#     
+#     Returns:
+#         dict: Fairness metrics including demographic parity, equalized odds
+#     
+#     TODO: Implement using fairlearn or aif360
+#     """
+#     pass
+
+
+
+# -------------------------------------------------------------------------
+# 1. Configuration
+# -------------------------------------------------------------------------
+
+MY_PLAYGROUND_ID = "https://cf3wdpkg0d.execute-api.us-east-1.amazonaws.com/prod/m"
+
+# --- Submission Limit Configuration ---
+# Maximum number of successful leaderboard submissions per user per session.
+# Preview runs (pre-login) and failed/invalid attempts do NOT count toward this limit.
+# Only actual successful playground.submit_model() calls increment the count.
+#
+# TODO: Server-side persistent enforcement recommended
+# The current attempt limit is stored in gr.State (per-session) and can be bypassed
+# by refreshing the browser. For production use with 100+ concurrent users,
+# consider implementing server-side persistence via Redis or Firestore to track
+# attempt counts per user across sessions.
+ATTEMPT_LIMIT = 1000000000
+
+# --- Leaderboard Polling Configuration ---
+# After a real authenticated submission, we poll the leaderboard to detect eventual consistency.
+# This prevents the "stuck on first preview KPI" issue where the leaderboard hasn't updated yet.
+# Increased from 12 to 60 to better tolerate backend latency and cold starts.
+# If polling times out, optimistic fallback logic will provide provisional UI updates.
+LEADERBOARD_POLL_TRIES = 60  # Number of polling attempts (increased to handle backend latency/cold starts)
+LEADERBOARD_POLL_SLEEP = 1.0  # Sleep duration between polls (seconds)
+ENABLE_AUTO_RESUBMIT_AFTER_READY = False  # Future feature flag for auto-resubmit
+
+MODEL_TYPES = {
+    "The Balanced Generalist": {
+        "model_builder": lambda: LogisticRegression(
+            max_iter=500, random_state=42, class_weight="balanced"
+        ),
+        "card": "A fast, reliable, well-rounded model. Good starting point; less prone to overfitting."
+    },
+    "The Rule-Maker": {
+        "model_builder": lambda: DecisionTreeClassifier(
+            random_state=42, class_weight="balanced"
+        ),
+        "card": "Learns simple 'if/then' rules. Easy to interpret, but can miss subtle patterns."
+    },
+    "The 'Nearest Neighbor'": {
+        "model_builder": lambda: KNeighborsClassifier(),
+        "card": "Looks at the closest past examples. 'You look like these others; I'll predict like they behave.'"
+    },
+    "The Deep Pattern-Finder": {
+        "model_builder": lambda: RandomForestClassifier(
+            random_state=42, class_weight="balanced"
+        ),
+        "card": "An ensemble of many decision trees. Powerful, can capture deep patterns; watch complexity."
+    }
+}
+
+DEFAULT_MODEL = "The Balanced Generalist"
+
+TEAM_NAMES = [
+    "The Moral Champions", "The Justice League", "The Data Detectives",
+    "The Ethical Explorers", "The Fairness Finders", "The Accuracy Avengers"
+]
+CURRENT_TEAM_NAME = random.choice(TEAM_NAMES)
+
+
+# --- Feature groups for scaffolding (Weak -> Medium -> Strong) ---
+FEATURE_SET_ALL_OPTIONS = [
+    ("Juvenile Felony Count", "juv_fel_count"),
+    ("Juvenile Misdemeanor Count", "juv_misd_count"),
+    ("Other Juvenile Count", "juv_other_count"),
+    ("Race", "race"),
+    ("Sex", "sex"),
+    ("Charge Severity (M/F)", "c_charge_degree"),
+    ("Days Before Arrest", "days_b_screening_arrest"),
+    ("Age", "age"),
+    ("Length of Stay", "length_of_stay"),
+    ("Prior Crimes Count", "priors_count"),
+]
+FEATURE_SET_GROUP_1_VALS = [
+    "juv_fel_count", "juv_misd_count", "juv_other_count", "race", "sex",
+    "c_charge_degree", "days_b_screening_arrest"
+]
+FEATURE_SET_GROUP_2_VALS = ["c_charge_desc", "age"]
+FEATURE_SET_GROUP_3_VALS = ["length_of_stay", "priors_count"]
+ALL_NUMERIC_COLS = [
+    "juv_fel_count", "juv_misd_count", "juv_other_count",
+    "days_b_screening_arrest", "age", "length_of_stay", "priors_count"
+]
+ALL_CATEGORICAL_COLS = [
+    "race", "sex", "c_charge_degree"
+]
+DEFAULT_FEATURE_SET = FEATURE_SET_GROUP_1_VALS
+
+
+# --- Data Size config ---
+DATA_SIZE_MAP = {
+    "Small (20%)": 0.2,
+    "Medium (60%)": 0.6,
+    "Large (80%)": 0.8,
+    "Full (100%)": 1.0
+}
+DEFAULT_DATA_SIZE = "Small (20%)"
+
+
+MAX_ROWS = 4000
+TOP_N_CHARGE_CATEGORICAL = 50
+WARM_MINI_ROWS = 300  # Small warm dataset for instant preview
+CACHE_MAX_AGE_HOURS = 24  # Cache validity duration
+np.random.seed(42)
+
+# Global state containers (populated during initialization)
+playground = None
+X_TRAIN_RAW = None # Keep this for 100%
+X_TEST_RAW = None
+Y_TRAIN = None
+Y_TEST = None
+# Add a container for our pre-sampled data
+X_TRAIN_SAMPLES_MAP = {}
+Y_TRAIN_SAMPLES_MAP = {}
+
+# Warm mini dataset for instant preview
+X_TRAIN_WARM = None
+Y_TRAIN_WARM = None
+
+# Cache for transformed test sets (for future performance improvements)
+TEST_CACHE = {}
+
+# Initialization flags to track readiness state
+INIT_FLAGS = {
+    "competition": False,
+    "dataset_core": False,
+    "pre_samples_small": False,
+    "pre_samples_medium": False,
+    "pre_samples_large": False,
+    "pre_samples_full": False,
+    "leaderboard": False,
+    "default_preprocessor": False,
+    "warm_mini": False,
+    "errors": []
+}
+
+# Lock for thread-safe flag updates
+INIT_LOCK = threading.Lock()
+
+# -------------------------------------------------------------------------
+# 2. Data & Backend Utilities
+# -------------------------------------------------------------------------
+
+def _get_cache_dir():
+    """Get or create the cache directory for datasets."""
+    cache_dir = Path.home() / ".aimodelshare_cache"
+    cache_dir.mkdir(exist_ok=True)
+    return cache_dir
+
+def _safe_request_csv(url, cache_filename="compas.csv"):
+    """
+    Request CSV from URL with local caching.
+    Reuses cached file if it exists and is less than CACHE_MAX_AGE_HOURS old.
+    """
+    cache_dir = _get_cache_dir()
+    cache_path = cache_dir / cache_filename
+    
+    # Check if cache exists and is fresh
+    if cache_path.exists():
+        file_time = datetime.fromtimestamp(cache_path.stat().st_mtime)
+        if datetime.now() - file_time < timedelta(hours=CACHE_MAX_AGE_HOURS):
+            return pd.read_csv(cache_path)
+    
+    # Download fresh data
+    response = requests.get(url, timeout=30)
+    response.raise_for_status()
+    df = pd.read_csv(StringIO(response.text))
+    
+    # Save to cache
+    df.to_csv(cache_path, index=False)
+    
+    return df
+
+def safe_int(value, default=1):
+    """
+    Safely coerce a value to int, returning default if value is None or invalid.
+    Protects against TypeError when Gradio sliders receive None.
+    """
+    if value is None:
+        return default
+    try:
+        return int(value)
+    except (ValueError, TypeError):
+        return default
+
+def load_and_prep_data(use_cache=True):
+    """
+    Load, sample, and prepare raw COMPAS dataset.
+    NOW PRE-SAMPLES ALL DATA SIZES and creates warm mini dataset.
+    """
+    url = "https://raw.githubusercontent.com/propublica/compas-analysis/master/compas-scores-two-years.csv"
+
+    # Use cached version if available
+    if use_cache:
+        try:
+            df = _safe_request_csv(url)
+        except Exception as e:
+            print(f"Cache failed, fetching directly: {e}")
+            response = requests.get(url)
+            df = pd.read_csv(StringIO(response.text))
+    else:
+        response = requests.get(url)
+        df = pd.read_csv(StringIO(response.text))
+
+    # Calculate length_of_stay
+    try:
+        df['c_jail_in'] = pd.to_datetime(df['c_jail_in'])
+        df['c_jail_out'] = pd.to_datetime(df['c_jail_out'])
+        df['length_of_stay'] = (df['c_jail_out'] - df['c_jail_in']).dt.total_seconds() / (24 * 60 * 60) # in days
+    except Exception:
+        df['length_of_stay'] = np.nan
+
+    if df.shape[0] > MAX_ROWS:
+        df = df.sample(n=MAX_ROWS, random_state=42)
+
+    feature_columns = ALL_NUMERIC_COLS + ALL_CATEGORICAL_COLS
+    feature_columns = sorted(list(set(feature_columns)))
+
+    target_column = "two_year_recid"
+
+    if "c_charge_desc" in df.columns:
+        top_charges = df["c_charge_desc"].value_counts().head(TOP_N_CHARGE_CATEGORICAL).index
+        df["c_charge_desc"] = df["c_charge_desc"].apply(
+            lambda x: x if pd.notna(x) and x in top_charges else "OTHER"
+        )
+
+    for col in feature_columns:
+        if col not in df.columns:
+            if col == 'length_of_stay' and 'length_of_stay' in df.columns:
+                continue
+            df[col] = np.nan
+
+    X = df[feature_columns].copy()
+    y = df[target_column].copy()
+
+    X_train_raw, X_test_raw, y_train, y_test = train_test_split(
+        X, y, test_size=0.25, random_state=42, stratify=y
+    )
+
+    # Pre-sample all data sizes
+    global X_TRAIN_SAMPLES_MAP, Y_TRAIN_SAMPLES_MAP, X_TRAIN_WARM, Y_TRAIN_WARM
+
+    X_TRAIN_SAMPLES_MAP["Full (100%)"] = X_train_raw
+    Y_TRAIN_SAMPLES_MAP["Full (100%)"] = y_train
+
+    for label, frac in DATA_SIZE_MAP.items():
+        if frac < 1.0:
+            X_train_sampled = X_train_raw.sample(frac=frac, random_state=42)
+            y_train_sampled = y_train.loc[X_train_sampled.index]
+            X_TRAIN_SAMPLES_MAP[label] = X_train_sampled
+            Y_TRAIN_SAMPLES_MAP[label] = y_train_sampled
+
+    # Create warm mini dataset for instant preview
+    warm_size = min(WARM_MINI_ROWS, len(X_train_raw))
+    X_TRAIN_WARM = X_train_raw.sample(n=warm_size, random_state=42)
+    Y_TRAIN_WARM = y_train.loc[X_TRAIN_WARM.index]
+
+
+
+    return X_train_raw, X_test_raw, y_train, y_test
+
+def _background_initializer():
+    """
+    Background thread that performs sequential initialization tasks.
+    Updates INIT_FLAGS dict with readiness booleans and captures errors.
+    
+    Initialization sequence:
+    1. Competition object connection
+    2. Dataset cached download and core split
+    3. Warm mini dataset creation
+    4. Progressive sampling: small -> medium -> large -> full
+    5. Leaderboard prefetch
+    6. Default preprocessor fit on small sample
+    """
+    global playground, X_TRAIN_RAW, X_TEST_RAW, Y_TRAIN, Y_TEST
+    
+    try:
+        # Step 1: Connect to competition
+        with INIT_LOCK:
+            if playground is None:
+                playground = Competition(MY_PLAYGROUND_ID)
+            INIT_FLAGS["competition"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Competition connection failed: {str(e)}")
+    
+    try:
+        # Step 2: Load dataset core (train/test split)
+        X_TRAIN_RAW, X_TEST_RAW, Y_TRAIN, Y_TEST = load_and_prep_data(use_cache=True)
+        with INIT_LOCK:
+            INIT_FLAGS["dataset_core"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Dataset loading failed: {str(e)}")
+        return  # Cannot proceed without data
+    
+    try:
+        # Step 3: Warm mini dataset (already created in load_and_prep_data)
+        if X_TRAIN_WARM is not None and len(X_TRAIN_WARM) > 0:
+            with INIT_LOCK:
+                INIT_FLAGS["warm_mini"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Warm mini dataset failed: {str(e)}")
+    
+    # Progressive sampling - samples are already created in load_and_prep_data
+    # Just mark them as ready sequentially with delays to simulate progressive loading
+    
+    try:
+        # Step 4a: Small sample (20%)
+        time.sleep(0.5)  # Simulate processing
+        with INIT_LOCK:
+            INIT_FLAGS["pre_samples_small"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Small sample failed: {str(e)}")
+    
+    try:
+        # Step 4b: Medium sample (60%)
+        time.sleep(0.5)
+        with INIT_LOCK:
+            INIT_FLAGS["pre_samples_medium"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Medium sample failed: {str(e)}")
+    
+    try:
+        # Step 4c: Large sample (80%)
+        time.sleep(0.5)
+        with INIT_LOCK:
+            INIT_FLAGS["pre_samples_large"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Large sample failed: {str(e)}")
+        print(f"✗ Large sample failed: {e}")
+    
+    try:
+        # Step 4d: Full sample (100%)
+        print("Background init: Full sample (100%)...")
+        time.sleep(0.5)
+        with INIT_LOCK:
+            INIT_FLAGS["pre_samples_full"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Full sample failed: {str(e)}")
+    
+    try:
+        # Step 5: Leaderboard prefetch (best-effort, unauthenticated)
+        # Concurrency Note: Do NOT use os.environ for ambient token - prefetch
+        # anonymously to warm the cache for initial page loads.
+        if playground is not None:
+            _ = _get_leaderboard_with_optional_token(playground, None)
+            with INIT_LOCK:
+                INIT_FLAGS["leaderboard"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Leaderboard prefetch failed: {str(e)}")
+    
+    try:
+        # Step 6: Default preprocessor on small sample
+        _fit_default_preprocessor()
+        with INIT_LOCK:
+            INIT_FLAGS["default_preprocessor"] = True
+    except Exception as e:
+        with INIT_LOCK:
+            INIT_FLAGS["errors"].append(f"Default preprocessor failed: {str(e)}")
+        print(f"✗ Default preprocessor failed: {e}")
+    
+
+def _fit_default_preprocessor():
+    """
+    Pre-fit a default preprocessor on the small sample with default features.
+    Uses memoized preprocessor builder for efficiency.
+    """
+    if "Small (20%)" not in X_TRAIN_SAMPLES_MAP:
+        return
+    
+    X_sample = X_TRAIN_SAMPLES_MAP["Small (20%)"]
+    
+    # Use default feature set
+    numeric_cols = [f for f in DEFAULT_FEATURE_SET if f in ALL_NUMERIC_COLS]
+    categorical_cols = [f for f in DEFAULT_FEATURE_SET if f in ALL_CATEGORICAL_COLS]
+    
+    if not numeric_cols and not categorical_cols:
+        return
+    
+    # Use memoized builder
+    preprocessor, selected_cols = build_preprocessor(numeric_cols, categorical_cols)
+    preprocessor.fit(X_sample[selected_cols])
+
+def start_background_init():
+    """
+    Start the background initialization thread.
+    Should be called once at app creation.
+    """
+    thread = threading.Thread(target=_background_initializer, daemon=True)
+    thread.start()
+
+def poll_init_status():
+    """
+    Poll the initialization status and return readiness bool.
+    Returns empty string for HTML so users don't see the checklist.
+    
+    Returns:
+        tuple: (status_html, ready_bool)
+    """
+    with INIT_LOCK:
+        flags = INIT_FLAGS.copy()
+    
+    # Determine if minimum requirements met
+    ready = flags["competition"] and flags["dataset_core"] and flags["pre_samples_small"]
+    
+    return "", ready
+
+def get_available_data_sizes():
+    """
+    Return list of data sizes that are currently available based on init flags.
+    """
+    with INIT_LOCK:
+        flags = INIT_FLAGS.copy()
+    
+    available = []
+    if flags["pre_samples_small"]:
+        available.append("Small (20%)")
+    if flags["pre_samples_medium"]:
+        available.append("Medium (60%)")
+    if flags["pre_samples_large"]:
+        available.append("Large (80%)")
+    if flags["pre_samples_full"]:
+        available.append("Full (100%)")
+    
+    return available if available else ["Small (20%)"]  # Fallback
+
+def _is_ready() -> bool:
+    """
+    Check if initialization is complete and system is ready for real submissions.
+    
+    Returns:
+        bool: True if competition, dataset, and small sample are initialized
+    """
+    with INIT_LOCK:
+        flags = INIT_FLAGS.copy()
+    return flags["competition"] and flags["dataset_core"] and flags["pre_samples_small"]
+
+def _get_user_latest_accuracy(df: Optional[pd.DataFrame], username: str) -> Optional[float]:
+    """
+    Extract the user's latest submission accuracy from the leaderboard.
+    
+    Uses timestamp sorting when available; otherwise assumes last row is latest.
+    
+    Args:
+        df: Leaderboard DataFrame
+        username: Username to extract accuracy for
+    
+    Returns:
+        float: Latest submission accuracy, or None if not found/invalid
+    """
+    if df is None or df.empty:
+        return None
+    
+    try:
+        user_rows = df[df["username"] == username]
+        if user_rows.empty or "accuracy" not in user_rows.columns:
+            return None
+        
+        # Try timestamp-based sorting if available
+        if "timestamp" in user_rows.columns:
+            user_rows = user_rows.copy()
+            user_rows["__parsed_ts"] = pd.to_datetime(user_rows["timestamp"], errors="coerce")
+            valid_ts = user_rows[user_rows["__parsed_ts"].notna()]
+            
+            if not valid_ts.empty:
+                # Sort by timestamp and get latest
+                latest_row = valid_ts.sort_values("__parsed_ts", ascending=False).iloc[0]
+                return float(latest_row["accuracy"])
+        
+        # Fallback: assume last row is latest (append order)
+        return float(user_rows.iloc[-1]["accuracy"])
+        
+    except Exception as e:
+        _log(f"Error extracting latest accuracy for {username}: {e}")
+        return None
+
+def _get_user_latest_ts(df: Optional[pd.DataFrame], username: str) -> Optional[float]:
+    """
+    Extract the user's latest valid timestamp from the leaderboard.
+    
+    Args:
+        df: Leaderboard DataFrame
+        username: Username to extract timestamp for
+    
+    Returns:
+        float: Latest timestamp as unix epoch, or None if not found/invalid
+    """
+    if df is None or df.empty:
+        return None
+    
+    try:
+        user_rows = df[df["username"] == username]
+        if user_rows.empty or "timestamp" not in user_rows.columns:
+            return None
+        
+        # Parse timestamps and get the latest
+        user_rows = user_rows.copy()
+        user_rows["__parsed_ts"] = pd.to_datetime(user_rows["timestamp"], errors="coerce")
+        valid_ts = user_rows[user_rows["__parsed_ts"].notna()]
+        
+        if valid_ts.empty:
+            return None
+        
+        latest_ts = valid_ts["__parsed_ts"].max()
+        return latest_ts.timestamp() if pd.notna(latest_ts) else None
+    except Exception as e:
+        _log(f"Error extracting latest timestamp for {username}: {e}")
+        return None
+
+def _user_rows_changed(
+    refreshed_leaderboard: Optional[pd.DataFrame],
+    username: str,
+    old_row_count: int,
+    old_best_score: float,
+    old_latest_ts: Optional[float] = None,
+    old_latest_score: Optional[float] = None
+) -> bool:
+    """
+    Check if user's leaderboard entries have changed after submission.
+    
+    Used after polling to detect if the leaderboard has updated with the new submission.
+    Checks row count (new submission added), best score (score improved), latest timestamp,
+    and latest accuracy (handles backend overwrite without append).
+    
+    Args:
+        refreshed_leaderboard: Fresh leaderboard data
+        username: Username to check for
+        old_row_count: Previous number of submissions for this user
+        old_best_score: Previous best accuracy score
+        old_latest_ts: Previous latest timestamp (unix epoch), optional
+        old_latest_score: Previous latest submission accuracy, optional
+    
+    Returns:
+        bool: True if user has more rows, better score, newer timestamp, or changed latest accuracy
+    """
+    if refreshed_leaderboard is None or refreshed_leaderboard.empty:
+        return False
+    
+    try:
+        user_rows = refreshed_leaderboard[refreshed_leaderboard["username"] == username]
+        if user_rows.empty:
+            return False
+        
+        new_row_count = len(user_rows)
+        new_best_score = float(user_rows["accuracy"].max()) if "accuracy" in user_rows.columns else 0.0
+        new_latest_ts = _get_user_latest_ts(refreshed_leaderboard, username)
+        new_latest_score = _get_user_latest_accuracy(refreshed_leaderboard, username)
+        
+        # Changed if we have more submissions, better score, newer timestamp, or changed latest accuracy
+        changed = (new_row_count > old_row_count) or (new_best_score > old_best_score + 0.0001)
+        
+        # Check timestamp if available
+        if old_latest_ts is not None and new_latest_ts is not None:
+            changed = changed or (new_latest_ts > old_latest_ts)
+        
+        # Check latest accuracy change (handles overwrite-without-append case)
+        if old_latest_score is not None and new_latest_score is not None:
+            accuracy_changed = abs(new_latest_score - old_latest_score) >= 0.00001
+            if accuracy_changed:
+                _log(f"Latest accuracy changed: {old_latest_score:.4f} -> {new_latest_score:.4f}")
+            changed = changed or accuracy_changed
+        
+        if changed:
+            _log(f"User rows changed for {username}:")
+            _log(f"  Row count: {old_row_count} -> {new_row_count}")
+            _log(f"  Best score: {old_best_score:.4f} -> {new_best_score:.4f}")
+            _log(f"  Latest score: {old_latest_score if old_latest_score else 'N/A'} -> {new_latest_score if new_latest_score else 'N/A'}")
+            _log(f"  Timestamp: {old_latest_ts} -> {new_latest_ts}")
+        
+        return changed
+    except Exception as e:
+        _log(f"Error checking user rows: {e}")
+        return False
+
+@functools.lru_cache(maxsize=32)
+def _get_cached_preprocessor_config(numeric_cols_tuple, categorical_cols_tuple):
+    """
+    Create and return preprocessor configuration (memoized).
+    Uses tuples for hashability in lru_cache.
+    
+    Concurrency Note: Uses sparse_output=True for OneHotEncoder to reduce memory
+    footprint under concurrent requests. Downstream models that require dense
+    arrays (DecisionTree, RandomForest) will convert via .toarray() as needed.
+    LogisticRegression and KNeighborsClassifier handle sparse matrices natively.
+    
+    Returns tuple of (transformers_list, selected_columns) ready for ColumnTransformer.
+    """
+    numeric_cols = list(numeric_cols_tuple)
+    categorical_cols = list(categorical_cols_tuple)
+    
+    transformers = []
+    selected_cols = []
+    
+    if numeric_cols:
+        num_tf = Pipeline(steps=[
+            ("imputer", SimpleImputer(strategy="median")),
+            ("scaler", StandardScaler())
+        ])
+        transformers.append(("num", num_tf, numeric_cols))
+        selected_cols.extend(numeric_cols)
+    
+    if categorical_cols:
+        # Use sparse_output=True to reduce memory footprint
+        cat_tf = Pipeline(steps=[
+            ("imputer", SimpleImputer(strategy="constant", fill_value="missing")),
+            ("onehot", OneHotEncoder(handle_unknown="ignore", sparse_output=True))
+        ])
+        transformers.append(("cat", cat_tf, categorical_cols))
+        selected_cols.extend(categorical_cols)
+    
+    return transformers, selected_cols
+
+def build_preprocessor(numeric_cols, categorical_cols):
+    """
+    Build a preprocessor using cached configuration.
+    The configuration (pipeline structure) is memoized; the actual fit is not.
+    
+    Note: Returns sparse matrices when categorical columns are present.
+    Use _ensure_dense() helper if model requires dense input.
+    """
+    # Convert to tuples for caching
+    numeric_tuple = tuple(sorted(numeric_cols))
+    categorical_tuple = tuple(sorted(categorical_cols))
+    
+    transformers, selected_cols = _get_cached_preprocessor_config(numeric_tuple, categorical_tuple)
+    
+    # Create new ColumnTransformer with cached config
+    preprocessor = ColumnTransformer(transformers=transformers, remainder="drop")
+    
+    return preprocessor, selected_cols
+
+def _ensure_dense(X):
+    """
+    Convert sparse matrix to dense if necessary.
+    
+    Helper function for models that don't support sparse input
+    (DecisionTree, RandomForest). LogisticRegression and KNN
+    handle sparse matrices natively.
+    """
+    from scipy import sparse
+    if sparse.issparse(X):
+        return X.toarray()
+    return X
+
+def tune_model_complexity(model, level):
+    """
+    Map a 1–10 slider value to model hyperparameters.
+    Levels 1–3: Conservative / simple
+    Levels 4–7: Balanced
+    Levels 8–10: Aggressive / risk of overfitting
+    """
+    level = int(level)
+    if isinstance(model, LogisticRegression):
+        c_map = {1: 0.01, 2: 0.025, 3: 0.05, 4: 0.1, 5: 0.25, 6: 0.5, 7: 1.0, 8: 2.0, 9: 5.0, 10: 10.0}
+        model.C = c_map.get(level, 1.0)
+        model.max_iter = max(getattr(model, "max_iter", 0), 500)
+    elif isinstance(model, RandomForestClassifier):
+        depth_map = {1: 3, 2: 5, 3: 7, 4: 9, 5: 11, 6: 15, 7: 20, 8: 25, 9: None, 10: None}
+        est_map = {1: 20, 2: 30, 3: 40, 4: 60, 5: 80, 6: 100, 7: 120, 8: 150, 9: 180, 10: 220}
+        model.max_depth = depth_map.get(level, 10)
+        model.n_estimators = est_map.get(level, 100)
+    elif isinstance(model, DecisionTreeClassifier):
+        depth_map = {1: 2, 2: 3, 3: 4, 4: 5, 5: 6, 6: 8, 7: 10, 8: 12, 9: 15, 10: None}
+        model.max_depth = depth_map.get(level, 6)
+    elif isinstance(model, KNeighborsClassifier):
+        k_map = {1: 100, 2: 75, 3: 60, 4: 50, 5: 40, 6: 30, 7: 25, 8: 15, 9: 7, 10: 3}
+        model.n_neighbors = k_map.get(level, 25)
+    return model
+
+# --- New Helper Functions for HTML Generation ---
+
+def _normalize_team_name(name: str) -> str:
+    """
+    Normalize team name for consistent comparison and storage.
+    
+    Strips leading/trailing whitespace and collapses multiple spaces into single spaces.
+    This ensures consistent formatting across environment variables, state, and leaderboard rendering.
+    
+    Args:
+        name: Team name to normalize (can be None or empty)
+    
+    Returns:
+        str: Normalized team name, or empty string if input is None/empty
+    
+    Examples:
+        >>> _normalize_team_name("  The Ethical Explorers  ")
+        'The Ethical Explorers'
+        >>> _normalize_team_name("The  Moral   Champions")
+        'The Moral Champions'
+        >>> _normalize_team_name(None)
+        ''
+    """
+    if not name:
+        return ""
+    return " ".join(str(name).strip().split())
+
+
+
+def _build_skeleton_leaderboard(rows=6, is_team=True, submit_button_label="5. 🔬 Build & Submit Model"):
+    context_label = "Team" if is_team else "Individual"
+    return f"""
+    <div class='lb-placeholder' aria-live='polite'>
+        <div class='lb-placeholder-title'>{context_label} Standings Pending</div>
+        <div class='lb-placeholder-sub'>
+            <p style='margin:0 0 6px 0;'>Submit your first model to populate this table.</p>
+            <p style='margin:0;'><strong>Click “{submit_button_label}” (bottom-left)</strong> to begin!</p>
+        </div>
+    </div>
+    """
+# --- FIX APPLIED HERE ---
+def build_login_prompt_html():
+    """
+    Generate HTML for the login prompt text *only*.
+    The styled preview card will be prepended to this.
+    """
+    return f"""
+    <h2 style='color: #111827; margin-top:20px; border-top: 2px solid #e5e7eb; padding-top: 20px;'>🔐 Sign in to submit & rank</h2>
+    <div style='margin-top:16px; text-align:left; font-size:1rem; line-height:1.6; color:#374151;'>
+        <p style='margin:12px 0;'>
+            This is a preview run only. Sign in to publish your score to the live leaderboard, 
+            earn promotions, and contribute team points.
+        </p>
+        <p style='margin:12px 0;'>
+            <strong>New user?</strong> Create a free account at 
+            <a href='https://www.modelshare.ai/login' target='_blank' 
+                style='color:#4f46e5; text-decoration:underline;'>modelshare.ai/login</a>
+        </p>
+    </div>
+    """
+# --- END OF FIX ---
+
+def _build_kpi_card_html(new_score, last_score, new_rank, last_rank, submission_count, is_preview=False, is_pending=False, local_test_accuracy=None):
+    """Generates the HTML for the KPI feedback card. Supports preview mode label and pending state."""
+
+    # Handle pending state - show processing message with provisional diff
+    if is_pending:
+        title = "⏳ Submission Processing"
+        acc_color = "#3b82f6"  # Blue
+        acc_text = f"{(local_test_accuracy * 100):.2f}%" if local_test_accuracy is not None else "N/A"
+        
+        # Compute provisional diff between local (new) and last score
+        if local_test_accuracy is not None and last_score is not None and last_score > 0:
+            score_diff = local_test_accuracy - last_score
+            if abs(score_diff) < 0.0001:
+                acc_diff_html = "<p style='font-size: 1.5rem; font-weight: 600; color: #6b7280; margin:0;'>No Change (↔) <span style='font-size: 0.9rem; color: #9ca3af;'>(Provisional)</span></p><p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>Pending leaderboard update...</p>"
+            elif score_diff > 0:
+                acc_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: #16a34a; margin:0;'>+{(score_diff * 100):.2f} (⬆️) <span style='font-size: 0.9rem; color: #9ca3af;'>(Provisional)</span></p><p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>Pending leaderboard update...</p>"
+            else:
+                acc_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: #ef4444; margin:0;'>{(score_diff * 100):.2f} (⬇️) <span style='font-size: 0.9rem; color: #9ca3af;'>(Provisional)</span></p><p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>Pending leaderboard update...</p>"
+        else:
+            # No last score available - just show pending message
+            acc_diff_html = "<p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>Pending leaderboard update...</p>"
+        
+        border_color = acc_color
+        rank_color = "#6b7280"  # Gray
+        rank_text = "Pending"
+        rank_diff_html = "<p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0;'>Calculating rank...</p>"
+        
+    # Handle preview mode - Styled to match "success" card
+    elif is_preview:
+        title = "🔬 Successful Preview Run!"
+        acc_color = "#16a34a"  # Green (like success)
+        acc_text = f"{(new_score * 100):.2f}%" if new_score > 0 else "N/A"
+        acc_diff_html = "<p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>(Preview only - not submitted)</p>" # Neutral color
+        border_color = acc_color # Green border
+        rank_color = "#3b82f6" # Blue (like rank)
+        rank_text = "N/A" # Placeholder
+        rank_diff_html = "<p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0;'>Not ranked (preview)</p>" # Neutral color
+    
+    # 1. Handle First Submission
+    elif submission_count == 0:
+        title = "🎉 First Model Submitted!"
+        acc_color = "#16a34a" # green
+        acc_text = f"{(new_score * 100):.2f}%"
+        acc_diff_html = "<p style='font-size: 1.2rem; font-weight: 500; color: #6b7280; margin:0; padding-top: 8px;'>(Your first score!)</p>"
+
+        rank_color = "#3b82f6" # blue
+        rank_text = f"#{new_rank}"
+        rank_diff_html = "<p style='font-size: 1.5rem; font-weight: 600; color: #3b82f6; margin:0;'>You're on the board!</p>"
+        border_color = acc_color
+
+    else:
+        # 2. Handle Score Changes
+        score_diff = new_score - last_score
+        if abs(score_diff) < 0.0001:
+            title = "✅ Submission Successful"
+            acc_color = "#6b7280" # gray
+            acc_text = f"{(new_score * 100):.2f}%"
+            acc_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: {acc_color}; margin:0;'>No Change (↔)</p>"
+            border_color = acc_color
+        elif score_diff > 0:
+            title = "✅ Submission Successful!"
+            acc_color = "#16a34a" # green
+            acc_text = f"{(new_score * 100):.2f}%"
+            acc_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: {acc_color}; margin:0;'>+{(score_diff * 100):.2f} (⬆️)</p>"
+            border_color = acc_color
+        else:
+            title = "📉 Score Dropped"
+            acc_color = "#ef4444" # red
+            acc_text = f"{(new_score * 100):.2f}%"
+            acc_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: {acc_color}; margin:0;'>{(score_diff * 100):.2f} (⬇️)</p>"
+            border_color = acc_color
+
+        # 3. Handle Rank Changes
+        rank_diff = last_rank - new_rank
+        rank_color = "#3b82f6" # blue
+        rank_text = f"#{new_rank}"
+        if last_rank == 0: # Handle first rank
+             rank_diff_html = "<p style='font-size: 1.5rem; font-weight: 600; color: #3b82f6; margin:0;'>You're on the board!</p>"
+        elif rank_diff > 0:
+            rank_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: #16a34a; margin:0;'>🚀 Moved up {rank_diff} spot{'s' if rank_diff > 1 else ''}!</p>"
+        elif rank_diff < 0:
+            rank_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: #ef4444; margin:0;'>🔻 Dropped {abs(rank_diff)} spot{'s' if abs(rank_diff) > 1 else ''}</p>"
+        else:
+            rank_diff_html = f"<p style='font-size: 1.5rem; font-weight: 600; color: {rank_color}; margin:0;'>No Change (↔)</p>"
+
+    return f"""
+    <div class='kpi-card' style='border-color: {border_color};'>
+        <h2 style='color: #111827; margin-top:0;'>{title}</h2>
+        <div class='kpi-card-body'>
+            <div class='kpi-metric-box'>
+                <p class='kpi-label'>New Accuracy</p>
+                <p class='kpi-score' style='color: {acc_color};'>{acc_text}</p>
+                {acc_diff_html}
+            </div>
+            <div class='kpi-metric-box'>
+                <p class='kpi-label'>Your Rank</p>
+                <p class='kpi-score' style='color: {rank_color};'>{rank_text}</p>
+                {rank_diff_html}
+            </div>
+        </div>
+    </div>
+    """
+
+def _build_team_html(team_summary_df, team_name):
+    """
+    Generates the HTML for the team leaderboard.
+    
+    Uses normalized, case-insensitive comparison to highlight the user's team row,
+    ensuring reliable highlighting even with whitespace or casing variations.
+    """
+    if team_summary_df is None or team_summary_df.empty:
+        return "<p style='text-align:center; color:#6b7280; padding-top:20px;'>No team submissions yet.</p>"
+
+    # Normalize the current user's team name for comparison
+    normalized_user_team = _normalize_team_name(team_name).lower()
+
+    header = """
+    <table class='leaderboard-html-table'>
+        <thead>
+            <tr>
+                <th>Rank</th>
+                <th>Team</th>
+                <th>Best_Score</th>
+                <th>Avg_Score</th>
+                <th>Submissions</th>
+            </tr>
+        </thead>
+        <tbody>
+    """
+
+    body = ""
+    for index, row in team_summary_df.iterrows():
+        # Normalize the row's team name and compare case-insensitively
+        normalized_row_team = _normalize_team_name(row["Team"]).lower()
+        is_user_team = normalized_row_team == normalized_user_team
+        row_class = "class='user-row-highlight'" if is_user_team else ""
+        body += f"""
+        <tr {row_class}>
+            <td>{index}</td>
+            <td>{row['Team']}</td>
+            <td>{(row['Best_Score'] * 100):.2f}%</td>
+            <td>{(row['Avg_Score'] * 100):.2f}%</td>
+            <td>{row['Submissions']}</td>
+        </tr>
+        """
+
+    footer = "</tbody></table>"
+    return header + body + footer
+
+def _build_individual_html(individual_summary_df, username):
+    """Generates the HTML for the individual leaderboard."""
+    if individual_summary_df is None or individual_summary_df.empty:
+        return "<p style='text-align:center; color:#6b7280; padding-top:20px;'>No individual submissions yet.</p>"
+
+    header = """
+    <table class='leaderboard-html-table'>
+        <thead>
+            <tr>
+                <th>Rank</th>
+                <th>Engineer</th>
+                <th>Best_Score</th>
+                <th>Submissions</th>
+            </tr>
+        </thead>
+        <tbody>
+    """
+
+    body = ""
+    for index, row in individual_summary_df.iterrows():
+        is_user = row["Engineer"] == username
+        row_class = "class='user-row-highlight'" if is_user else ""
+        body += f"""
+        <tr {row_class}>
+            <td>{index}</td>
+            <td>{row['Engineer']}</td>
+            <td>{(row['Best_Score'] * 100):.2f}%</td>
+            <td>{row['Submissions']}</td>
+        </tr>
+        """
+
+    footer = "</tbody></table>"
+    return header + body + footer
+
+
+
+
+# --- End Helper Functions ---
+
+
+def generate_competitive_summary(leaderboard_df, team_name, username, last_submission_score, last_rank, submission_count):
+    """
+    Build summaries, HTML, and KPI card.
+    
+    Concurrency Note: Uses the team_name parameter directly for team highlighting,
+    NOT os.environ, to prevent cross-user data leakage under concurrent requests.
+    
+    Returns (team_html, individual_html, kpi_card_html, new_best_accuracy, new_rank, this_submission_score).
+    """
+    team_summary_df = pd.DataFrame(columns=["Team", "Best_Score", "Avg_Score", "Submissions"])
+    individual_summary_df = pd.DataFrame(columns=["Engineer", "Best_Score", "Submissions"])
+
+    if leaderboard_df is None or leaderboard_df.empty or "accuracy" not in leaderboard_df.columns:
+        return (
+            "<p style='text-align:center; color:#6b7280; padding-top:20px;'>Leaderboard empty.</p>",
+            "<p style='text-align:center; color:#6b7280; padding-top:20px;'>Leaderboard empty.</p>",
+            _build_kpi_card_html(0, 0, 0, 0, 0, is_preview=False, is_pending=False, local_test_accuracy=None), 
+            0.0, 0, 0.0
+        )
+
+    # Team summary
+    if "Team" in leaderboard_df.columns:
+        team_summary_df = (
+            leaderboard_df.groupby("Team")["accuracy"]
+            .agg(Best_Score="max", Avg_Score="mean", Submissions="count")
+            .reset_index()
+            .sort_values("Best_Score", ascending=False)
+            .reset_index(drop=True)
+        )
+        team_summary_df.index = team_summary_df.index + 1
+
+    # Individual summary
+    user_bests = leaderboard_df.groupby("username")["accuracy"].max()
+    user_counts = leaderboard_df.groupby("username")["accuracy"].count()
+    individual_summary_df = pd.DataFrame(
+        {"Engineer": user_bests.index, "Best_Score": user_bests.values, "Submissions": user_counts.values}
+    ).sort_values("Best_Score", ascending=False).reset_index(drop=True)
+    individual_summary_df.index = individual_summary_df.index + 1
+
+    # Get stats for KPI card
+    new_rank = 0
+    new_best_accuracy = 0.0
+    this_submission_score = 0.0
+
+    try:
+        # All submissions for this user
+        user_rows = leaderboard_df[leaderboard_df["username"] == username].copy()
+
+        if not user_rows.empty:
+            # Attempt robust timestamp parsing
+            if "timestamp" in user_rows.columns:
+                parsed_ts = pd.to_datetime(user_rows["timestamp"], errors="coerce")
+
+                if parsed_ts.notna().any():
+                    # At least one valid timestamp → use parsed ordering
+                    user_rows["__parsed_ts"] = parsed_ts
+                    user_rows = user_rows.sort_values("__parsed_ts", ascending=False)
+                    this_submission_score = float(user_rows.iloc[0]["accuracy"])
+                else:
+                    # All timestamps invalid → assume append order, take last as "latest"
+                    this_submission_score = float(user_rows.iloc[-1]["accuracy"])
+            else:
+                # No timestamp column → fallback to last row
+                this_submission_score = float(user_rows.iloc[-1]["accuracy"])
+
+        # Rank & best accuracy (unchanged logic, but make sure we use the same best row)
+        my_rank_row = None
+        # Build individual summary before this block (already done above)
+        my_rank_row = individual_summary_df[individual_summary_df["Engineer"] == username]
+        if not my_rank_row.empty:
+            new_rank = my_rank_row.index[0]
+            new_best_accuracy = float(my_rank_row["Best_Score"].iloc[0])
+
+    except Exception as e:
+        _log(f"Latest submission score extraction failed: {e}")
+
+    # Generate HTML outputs
+    # Concurrency Note: Use team_name parameter directly, not os.environ
+    team_html = _build_team_html(team_summary_df, team_name)
+    individual_html = _build_individual_html(individual_summary_df, username)
+    kpi_card_html = _build_kpi_card_html(
+        this_submission_score, last_submission_score, new_rank, last_rank, submission_count,
+        is_preview=False, is_pending=False, local_test_accuracy=None
+    )
+
+    return team_html, individual_html, kpi_card_html, new_best_accuracy, new_rank, this_submission_score
+
+
+def get_model_card(model_name):
+    return MODEL_TYPES.get(model_name, {}).get("card", "No description available.")
+
+def compute_rank_settings(
+    submission_count,
+    current_model,
+    current_complexity,
+    current_feature_set,
+    current_data_size
+):
+    """Returns rank gating settings (updated for 1–10 complexity scale)."""
+
+
+    # Always allow all options
+    return {
+        "rank_message": "# 👑 Rank: Lead Engineer\n<p style='font-size:24px; line-height:1.4;'>All tools unlocked — optimize freely!</p>",
+        "model_choices": list(MODEL_TYPES.keys()),
+        "model_value": current_model if current_model in MODEL_TYPES else "The Balanced Generalist",
+        "model_interactive": True,
+        "complexity_max": 10,
+        "complexity_value": current_complexity,
+        "feature_set_choices": FEATURE_SET_ALL_OPTIONS, 
+        "feature_set_value": current_feature_set,
+        "feature_set_interactive": True,
+        "data_size_choices": ["Small (20%)", "Medium (60%)", "Large (80%)", "Full (100%)"],
+        "data_size_value": current_data_size if current_data_size in DATA_SIZE_MAP else "Small (20%)",
+        "data_size_interactive": True,
+    }
+
+
+# Find components by name to yield updates
+# --- Existing global component placeholders ---
+submit_button = None
+submission_feedback_display = None
+team_leaderboard_display = None
+individual_leaderboard_display = None
+last_submission_score_state = None 
+last_rank_state = None 
+best_score_state = None
+submission_count_state = None
+rank_message_display = None
+model_type_radio = None
+complexity_slider = None
+feature_set_checkbox = None
+data_size_radio = None
+attempts_tracker_display = None
+team_name_state = None
+# Login components
+login_username = None
+login_password = None
+login_submit = None
+login_error = None
+# Add missing placeholders for auth states (FIX)
+username_state = None
+token_state = None
+first_submission_score_state = None  # (already commented as "will be assigned globally")
+# Add state placeholders for readiness gating and preview tracking
+readiness_state = None
+was_preview_state = None
+kpi_meta_state = None
+last_seen_ts_state = None  # Track last seen user timestamp from leaderboard
+
+
+def get_or_assign_team(username, token=None):
+    """
+    Get the existing team for a user from the leaderboard, or assign a new random team.
+    
+    Queries the playground leaderboard to check if the user has prior submissions with
+    a team assignment. If found, returns that team (most recent if multiple submissions).
+    Otherwise assigns a random team. All team names are normalized for consistency.
+    
+    Args:
+        username: str, the username to check for existing team
+        token: str, optional authentication token for leaderboard fetch
+    
+    Returns:
+        tuple: (team_name: str, is_new: bool)
+            - team_name: The normalized team name (existing or newly assigned)
+            - is_new: True if newly assigned, False if existing team recovered
+    """
+    try:
+        # Query the leaderboard
+        if playground is None:
+            # Fallback to random assignment if playground not available
+            print("Playground not available, assigning random team")
+            new_team = _normalize_team_name(random.choice(TEAM_NAMES))
+            return new_team, True
+        
+        # Use centralized helper for authenticated leaderboard fetch
+        leaderboard_df = _get_leaderboard_with_optional_token(playground, token)
+        
+        # Check if leaderboard has data and Team column
+        if leaderboard_df is not None and not leaderboard_df.empty and "Team" in leaderboard_df.columns:
+            # Filter for this user's submissions
+            user_submissions = leaderboard_df[leaderboard_df["username"] == username]
+            
+            if not user_submissions.empty:
+                # Sort by timestamp (most recent first) if timestamp column exists
+                # Use contextlib.suppress for resilient timestamp parsing
+                if "timestamp" in user_submissions.columns:
+                    try:
+                        # Attempt to coerce timestamp column to datetime and sort descending
+                        user_submissions = user_submissions.copy()
+                        user_submissions["timestamp"] = pd.to_datetime(user_submissions["timestamp"], errors='coerce')
+                        user_submissions = user_submissions.sort_values("timestamp", ascending=False)
+                        print(f"Sorted {len(user_submissions)} submissions by timestamp for {username}")
+                    except Exception as ts_error:
+                        # If timestamp parsing fails, continue with unsorted DataFrame
+                        print(f"Warning: Could not sort by timestamp for {username}: {ts_error}")
+                
+                # Get the most recent team assignment (first row after sorting)
+                existing_team = user_submissions.iloc[0]["Team"]
+                
+                # Check if team value is valid (not null/empty)
+                if pd.notna(existing_team) and existing_team and str(existing_team).strip():
+                    normalized_team = _normalize_team_name(existing_team)
+                    print(f"Found existing team for {username}: {normalized_team}")
+                    return normalized_team, False
+        
+        # No existing team found - assign random
+        new_team = _normalize_team_name(random.choice(TEAM_NAMES))
+        print(f"Assigning new team to {username}: {new_team}")
+        return new_team, True
+        
+    except Exception as e:
+        # On any error, fall back to random assignment
+        print(f"Error checking leaderboard for team: {e}")
+        new_team = _normalize_team_name(random.choice(TEAM_NAMES))
+        print(f"Fallback: assigning random team to {username}: {new_team}")
+        return new_team, True
+
+def perform_inline_login(username_input, password_input):
+    """
+    Perform inline authentication and return credentials via gr.State updates.
+    
+    Concurrency Note: This function NO LONGER stores per-user credentials in
+    os.environ to prevent cross-user data leakage. Authentication state is
+    returned exclusively via gr.State updates (username_state, token_state,
+    team_name_state). Password is never stored server-side.
+    
+    Args:
+        username_input: str, the username entered by user
+        password_input: str, the password entered by user
+    
+    Returns:
+        dict: Gradio component updates for login UI elements and submit button
+            - On success: hides login form, shows success message, enables submit
+            - On failure: keeps login form visible, shows error with signup link
+    """
+    from aimodelshare.aws import get_aws_token
+    
+    # Validate inputs
+    if not username_input or not username_input.strip():
+        error_html = """
+        <div style='background:#fef2f2; padding:12px; border-radius:8px; border-left:4px solid #ef4444; margin-top:12px;'>
+            <p style='margin:0; color:#991b1b; font-weight:500;'>⚠️ Username is required</p>
+        </div>
+        """
+        return {
+            login_username: gr.update(),
+            login_password: gr.update(),
+            login_submit: gr.update(),
+            login_error: gr.update(value=error_html, visible=True),
+            submit_button: gr.update(),
+            submission_feedback_display: gr.update(),
+            team_name_state: gr.update(),
+            username_state: gr.update(),
+            token_state: gr.update()
+        }
+    
+    if not password_input or not password_input.strip():
+        error_html = """
+        <div style='background:#fef2f2; padding:12px; border-radius:8px; border-left:4px solid #ef4444; margin-top:12px;'>
+            <p style='margin:0; color:#991b1b; font-weight:500;'>⚠️ Password is required</p>
+        </div>
+        """
+        return {
+            login_username: gr.update(),
+            login_password: gr.update(),
+            login_submit: gr.update(),
+            login_error: gr.update(value=error_html, visible=True),
+            submit_button: gr.update(),
+            submission_feedback_display: gr.update(),
+            team_name_state: gr.update(),
+            username_state: gr.update(),
+            token_state: gr.update()
+        }
+    
+    # Concurrency Note: get_aws_token() reads credentials from os.environ, which creates
+    # a race condition in multi-threaded environments. We use _auth_lock to serialize
+    # credential injection, preventing concurrent requests from seeing each other's
+    # credentials. The password is immediately cleared after the auth attempt.
+    # 
+    # FUTURE: Ideally get_aws_token() would be refactored to accept credentials as
+    # parameters instead of reading from os.environ. This lock is a workaround.
+    username_clean = username_input.strip()
+    
+    # Attempt to get AWS token with serialized credential injection
+    try:
+        with _auth_lock:
+            os.environ["username"] = username_clean
+            os.environ["password"] = password_input.strip()  # Only for get_aws_token() call
+            try:
+                token = get_aws_token()
+            finally:
+                # SECURITY: Always clear credentials from environment, even on exception
+                # Also clear stale env vars from previous implementations within the lock
+                # to prevent any race conditions during cleanup
+                os.environ.pop("password", None)
+                os.environ.pop("username", None)
+                os.environ.pop("AWS_TOKEN", None)
+                os.environ.pop("TEAM_NAME", None)
+        
+        # Get or assign team for this user with explicit token (already normalized by get_or_assign_team)
+        team_name, is_new_team = get_or_assign_team(username_clean, token=token)
+        # Normalize team name before storing (defensive - already normalized by get_or_assign_team)
+        team_name = _normalize_team_name(team_name)
+        
+        # Build success message based on whether team is new or existing
+        if is_new_team:
+            team_message = f"You have been assigned to a new team: <b>{team_name}</b> 🎉"
+        else:
+            team_message = f"Welcome back! You remain on team: <b>{team_name}</b> ✅"
+        
+        # Success: hide login form, show success message with team info, enable submit button
+        success_html = f"""
+        <div style='background:#f0fdf4; padding:16px; border-radius:8px; border-left:4px solid #16a34a; margin-top:12px;'>
+            <p style='margin:0; color:#15803d; font-weight:600; font-size:1.1rem;'>✓ Signed in successfully!</p>
+            <p style='margin:8px 0 0 0; color:#166534; font-size:0.95rem;'>
+                {team_message}
+            </p>
+            <p style='margin:8px 0 0 0; color:#166534; font-size:0.95rem;'>
+                Click "Build & Submit Model" again to publish your score.
+            </p>
+        </div>
+        """
+        return {
+            login_username: gr.update(visible=False),
+            login_password: gr.update(visible=False),
+            login_submit: gr.update(visible=False),
+            login_error: gr.update(value=success_html, visible=True),
+            submit_button: gr.update(value="🔬 Build & Submit Model", interactive=True),
+            submission_feedback_display: gr.update(visible=False),
+            team_name_state: gr.update(value=team_name),
+            username_state: gr.update(value=username_clean),
+            token_state: gr.update(value=token)
+        }
+        
+    except Exception as e:
+        # Note: Credentials are already cleaned up by the finally block in the try above.
+        # The lock ensures no race condition during cleanup.
+        
+        # Authentication failed: show error with signup link
+        error_html = f"""
+        <div style='background:#fef2f2; padding:16px; border-radius:8px; border-left:4px solid #ef4444; margin-top:12px;'>
+            <p style='margin:0; color:#991b1b; font-weight:600; font-size:1.1rem;'>⚠️ Authentication failed</p>
+            <p style='margin:8px 0; color:#7f1d1d; font-size:0.95rem;'>
+                Could not verify your credentials. Please check your username and password.
+            </p>
+            <p style='margin:8px 0 0 0; color:#7f1d1d; font-size:0.95rem;'>
+                <strong>New user?</strong> Create a free account at 
+                <a href='https://www.modelshare.ai/login' target='_blank' 
+                   style='color:#dc2626; text-decoration:underline;'>modelshare.ai/login</a>
+            </p>
+            <details style='margin-top:12px; font-size:0.85rem; color:#7f1d1d;'>
+                <summary style='cursor:pointer;'>Technical details</summary>
+                <pre style='margin-top:8px; padding:8px; background:#fee; border-radius:4px; overflow-x:auto;'>{str(e)}</pre>
+            </details>
+        </div>
+        """
+        return {
+            login_username: gr.update(visible=True),
+            login_password: gr.update(visible=True),
+            login_submit: gr.update(visible=True),
+            login_error: gr.update(value=error_html, visible=True),
+            submit_button: gr.update(),
+            submission_feedback_display: gr.update(),
+            team_name_state: gr.update(),
+            username_state: gr.update(),
+            token_state: gr.update()
+        }
+
+def run_experiment(
+    model_name_key,
+    complexity_level,
+    feature_set,
+    data_size_str,
+    team_name,
+    last_submission_score,
+    last_rank,
+    submission_count,
+    first_submission_score,
+    best_score,
+    username=None,
+    token=None,
+    readiness_flag=None,
+    was_preview_prev=None,
+    progress=gr.Progress()
+):
+    """
+    Core experiment: Uses 'yield' for visual updates and progress bar.
+    Updated with "Look-Before-You-Leap" caching strategy.
+    """
+    # --- COLLISION GUARDS ---
+    # Log types of potentially shadowed names to ensure they refer to component objects, not dicts
+    _log(f"DEBUG guard: types — submit_button={type(submit_button)} submission_feedback_display={type(submission_feedback_display)} kpi_meta_state={type(kpi_meta_state)} was_preview_state={type(was_preview_state)} readiness_flag_param={type(readiness_flag)}")
+    
+    # If any of the component names are found as dicts (indicating parameter shadowing), short-circuit
+    if isinstance(submit_button, dict) or isinstance(submission_feedback_display, dict) or isinstance(kpi_meta_state, dict) or isinstance(was_preview_state, dict):
+        error_html = """
+        <div class='kpi-card' style='border-color: #ef4444;'>
+            <h2 style='color: #111827; margin-top:0;'>⚠️ Configuration Error</h2>
+            <div class='kpi-card-body'>
+                <p style='color: #991b1b;'>Parameter shadowing detected. Global component variables were shadowed by local parameters.</p>
+                <p style='color: #7f1d1d; margin-top: 8px;'>Please refresh the page and try again. If the issue persists, contact support.</p>
+            </div>
+        </div>
+        """
+        yield {
+            submission_feedback_display: gr.update(value=error_html, visible=True),
+            submit_button: gr.update(value="🔬 Build & Submit Model", interactive=True)
+        }
+        return
+    
+    # Sanitize feature_set: convert dicts/tuples to their string values
+    sanitized_feature_set = []
+    for feat in (feature_set or []):
+        if isinstance(feat, dict):
+            # Extract 'value' key if present, otherwise use string representation
+            sanitized_feature_set.append(feat.get("value", str(feat)))
+        elif isinstance(feat, tuple):
+            # For tuples like ("Label", "value"), take the second element
+            sanitized_feature_set.append(feat[1] if len(feat) > 1 else str(feat))
+        else:
+            # Already a string
+            sanitized_feature_set.append(str(feat))
+    feature_set = sanitized_feature_set
+    
+    # Use readiness_flag parameter if provided, otherwise check readiness
+    if readiness_flag is not None:
+        ready = readiness_flag
+    else:
+        ready = _is_ready()
+    _log(f"run_experiment: ready={ready}, username={username}, token_present={token is not None}")
+    
+    # Add debug log (optional)
+    _log(f"run_experiment received username={username} token_present={token is not None}")    
+    # Concurrency Note: Use provided parameters exclusively, not os.environ.
+    # Default to "Unknown_User" only if no username provided via state.
+    if not username:
+        username = "Unknown_User"
+    
+    # Helper to generate the animated HTML
+    def get_status_html(step_num, title, subtitle):
+        return f"""
+        <div class='processing-status'>
+            <span class='processing-icon'>⚙️</span>
+            <div class='processing-text'>Step {step_num}/5: {title}</div>
+            <div class='processing-subtext'>{subtitle}</div>
+        </div>
+        """
+
+    # --- Stage 1: Lock UI and give initial feedback ---
+    progress(0.1, desc="Starting Experiment...")
+    initial_updates = {
+        submit_button: gr.update(value="⏳ Experiment Running...", interactive=False),
+        submission_feedback_display: gr.update(value=get_status_html(1, "Initializing", "Preparing your data ingredients..."), visible=True), # Make sure it's visible
+        login_error: gr.update(visible=False), # Hide login success/error message
+        attempts_tracker_display: gr.update(value=_build_attempts_tracker_html(submission_count))
+    }
+    yield initial_updates
+
+    if not model_name_key or model_name_key not in MODEL_TYPES:
+        model_name_key = DEFAULT_MODEL
+    complexity_level = safe_int(complexity_level, 2)
+
+    log_output = f"▶ New Experiment\nModel: {model_name_key}\n..."
+
+    # Check readiness
+    # If playground is None or not ready, fallback error
+    if playground is None or not ready:
+        settings = compute_rank_settings(
+             submission_count, model_name_key, complexity_level, feature_set, data_size_str
+        )
+        
+        error_msg = "<p style='text-align:center; color:red; padding:20px 0;'>"
+        if playground is None:
+            error_msg += "Playground not connected. Please try again later."
+        else:
+            error_msg += "Data still initializing. Please wait a moment and try again."
+        error_msg += "</p>"
+        
+        error_kpi_meta = {
+            "was_preview": False, "preview_score": None, "ready_at_run_start": False,
+            "poll_iterations": 0, "local_test_accuracy": None, "this_submission_score": None,
+            "new_best_accuracy": None, "rank": None
+        }
+        
+        error_updates = {
+            submission_feedback_display: gr.update(value=error_msg, visible=True),
+            submit_button: gr.update(value="🔬 Build & Submit Model", interactive=True),
+            team_leaderboard_display: _build_skeleton_leaderboard(rows=6, is_team=True),
+            individual_leaderboard_display: _build_skeleton_leaderboard(rows=6, is_team=False),
+            last_submission_score_state: last_submission_score,
+            last_rank_state: last_rank,
+            best_score_state: best_score,
+            submission_count_state: submission_count,
+            first_submission_score_state: first_submission_score,
+            rank_message_display: settings["rank_message"],
+            model_type_radio: gr.update(choices=settings["model_choices"], value=settings["model_value"], interactive=settings["model_interactive"]),
+            complexity_slider: gr.update(minimum=1, maximum=settings["complexity_max"], value=settings["complexity_value"]),
+            feature_set_checkbox: gr.update(choices=settings["feature_set_choices"], value=settings["feature_set_value"], interactive=settings["feature_set_interactive"]),
+            data_size_radio: gr.update(choices=settings["data_size_choices"], value=settings["data_size_value"], interactive=settings["data_size_interactive"]),
+            login_username: gr.update(visible=False),
+            login_password: gr.update(visible=False),
+            login_submit: gr.update(visible=False),
+            login_error: gr.update(visible=False),
+            attempts_tracker_display: gr.update(value=_build_attempts_tracker_html(submission_count)),
+            was_preview_state: False,
+            kpi_meta_state: error_kpi_meta,
+            last_seen_ts_state: None
+        }
+        yield error_updates
+        return
+
+    try:
+        # --- Stage 2: Smart Build (Cache vs Train) ---
+        progress(0.3, desc="Building Model...")
+        
+        # 1. Generate Cache Key (Matches format in precompute_cache.py)
+        # Key: "ModelName|Complexity|DataSize|SortedFeatures"
+        sanitized_features = sorted([str(f) for f in feature_set])
+        feature_key = ",".join(sanitized_features)
+        cache_key = f"{model_name_key}|{complexity_level}|{data_size_str}|{feature_key}"
+        
+        # 2. Check Cache
+        cached_predictions = get_cached_prediction(cache_key)
+        
+        # Initialize submission variables
+        predictions = None
+        tuned_model = None
+        preprocessor = None
+        
+        if cached_predictions:
+            # === FAST PATH (Zero CPU) ===
+            _log(f"⚡ CACHE HIT: {cache_key}")
+            yield { 
+                submission_feedback_display: gr.update(value=get_status_html(2, "Training Model", "⚡ The machine is learning from history..."), visible=True),
+                login_error: gr.update(visible=False)
+            }
+
+            # --- DECOMPRESSION STEP (Vital) ---
+            # If string "01010...", convert to [0, 1, 0, 1...]
+            if isinstance(cached_predictions, str):
+                predictions = [int(c) for c in cached_predictions]
+            else:
+                predictions = cached_predictions
+
+            # Pass None to submit_model to skip training overhead validation
+            tuned_model = None
+            preprocessor = None
+            
+            
+        else:
+            # === SLOW PATH (Fallback Training) ===
+            _log(f"🐢 CACHE MISS: {cache_key} (Training for real...)")
+            yield { 
+                submission_feedback_display: gr.update(value=get_status_html(2, "Training Model", "The machine is learning from history..."), visible=True),
+                login_error: gr.update(visible=False)
+            }
+
+            # A. Get pre-sampled data
+            sample_frac = DATA_SIZE_MAP.get(data_size_str, 0.2)
+            X_train_sampled = X_TRAIN_SAMPLES_MAP[data_size_str]
+            y_train_sampled = Y_TRAIN_SAMPLES_MAP[data_size_str]
+            log_output += f"Using {int(sample_frac * 100)}% data.\n"
+
+            # B. Determine features...
+            numeric_cols = [f for f in feature_set if f in ALL_NUMERIC_COLS]
+            categorical_cols = [f for f in feature_set if f in ALL_CATEGORICAL_COLS]
+            for feat in feature_set:
+                if feat in ALL_NUMERIC_COLS: numeric_cols.append(feat)
+                elif feat in ALL_CATEGORICAL_COLS: categorical_cols.append(feat)
+            
+            # De-dupe logic just in case (though loop above covers it, ensuring lists are clean)
+            numeric_cols = sorted(list(set([f for f in feature_set if f in ALL_NUMERIC_COLS])))
+            categorical_cols = sorted(list(set([f for f in feature_set if f in ALL_CATEGORICAL_COLS])))
+
+            if not numeric_cols and not categorical_cols:
+                raise ValueError("No features selected for modeling.")
+
+            # C. Preprocessing (uses memoized preprocessor builder)
+            preprocessor, selected_cols = build_preprocessor(numeric_cols, categorical_cols)
+
+            X_train_processed = preprocessor.fit_transform(X_train_sampled[selected_cols])
+            X_test_processed = preprocessor.transform(X_TEST_RAW[selected_cols])
+
+            # D. Model build & tune
+            base_model = MODEL_TYPES[model_name_key]["model_builder"]()
+            tuned_model = tune_model_complexity(base_model, complexity_level)
+
+            # E. Train
+            # Concurrency Note: DecisionTree and RandomForest require dense arrays.
+            if isinstance(tuned_model, (DecisionTreeClassifier, RandomForestClassifier)):
+                X_train_for_fit = _ensure_dense(X_train_processed)
+                X_test_for_predict = _ensure_dense(X_test_processed)
+            else:
+                X_train_for_fit = X_train_processed
+                X_test_for_predict = X_test_processed
+            
+            tuned_model.fit(X_train_for_fit, y_train_sampled)
+            log_output += "Training done.\n"
+            
+            # F. Predict
+            predictions = tuned_model.predict(X_test_for_predict)
+
+        # --- Stage 3: Submit (API Call 1) ---
+        # AUTHENTICATION GATE: Check for token before submission
+        if token is None:
+            # User not authenticated - compute preview score and show login prompt
+            progress(0.6, desc="Computing Preview Score...")
+            
+            # We need to calculate accuracy for the preview card
+            from sklearn.metrics import accuracy_score
+            # Ensure predictions are in correct format (list or array)
+            if isinstance(predictions, list):
+                # Cached predictions are lists
+                preds_array = np.array(predictions)
+            else:
+                preds_array = predictions
+                
+            preview_score = accuracy_score(Y_TEST, preds_array)
+            
+            preview_kpi_meta = {
+                "was_preview": True, "preview_score": preview_score, "ready_at_run_start": ready,
+                "poll_iterations": 0, "local_test_accuracy": preview_score,
+                "this_submission_score": None, "new_best_accuracy": None, "rank": None
+            }
+            
+            # 1. Generate the styled preview card
+            preview_card_html = _build_kpi_card_html(
+                new_score=preview_score, last_score=0, new_rank=0, last_rank=0,
+                submission_count=-1, is_preview=True, is_pending=False, local_test_accuracy=None
+            )
+            
+            # 2. Inject login text
+            login_prompt_text_html = build_login_prompt_html() 
+            closing_div_index = preview_card_html.rfind("</div>")
+            if closing_div_index != -1:
+                combined_html = preview_card_html[:closing_div_index] + login_prompt_text_html + "</div>"
+            else:
+                combined_html = preview_card_html + login_prompt_text_html 
+                
+            settings = compute_rank_settings(submission_count, model_name_key, complexity_level, feature_set, data_size_str)
+            
+            gate_updates = {
+                submission_feedback_display: gr.update(value=combined_html, visible=True),
+                submit_button: gr.update(value="Sign In Required", interactive=False),
+                login_username: gr.update(visible=True), login_password: gr.update(visible=True),
+                login_submit: gr.update(visible=True), login_error: gr.update(value="", visible=False),
+                team_leaderboard_display: _build_skeleton_leaderboard(rows=6, is_team=True),
+                individual_leaderboard_display: _build_skeleton_leaderboard(rows=6, is_team=False),
+                last_submission_score_state: last_submission_score, last_rank_state: last_rank,
+                best_score_state: best_score, submission_count_state: submission_count,
+                first_submission_score_state: first_submission_score,
+                rank_message_display: settings["rank_message"],
+                model_type_radio: gr.update(choices=settings["model_choices"], value=settings["model_value"], interactive=settings["model_interactive"]),
+                complexity_slider: gr.update(minimum=1, maximum=settings["complexity_max"], value=settings["complexity_value"]),
+                feature_set_checkbox: gr.update(choices=settings["feature_set_choices"], value=settings["feature_set_value"], interactive=settings["feature_set_interactive"]),
+                data_size_radio: gr.update(choices=settings["data_size_choices"], value=settings["data_size_value"], interactive=settings["data_size_interactive"]),
+                attempts_tracker_display: gr.update(value=_build_attempts_tracker_html(submission_count)),
+                was_preview_state: True, kpi_meta_state: preview_kpi_meta, last_seen_ts_state: None
+            }
+            yield gate_updates
+            return  # Stop here
+        
+        # --- ATTEMPT LIMIT CHECK ---
+        if submission_count >= ATTEMPT_LIMIT:
+            limit_warning_html = f"""
+            <div class='kpi-card' style='border-color: #ef4444;'>
+                <h2 style='color: #111827; margin-top:0;'>🛑 Submission Limit Reached</h2>
+                <div class='kpi-card-body'>
+                    <div class='kpi-metric-box'>
+                        <p class='kpi-label'>Attempts Used</p>
+                        <p class='kpi-score' style='color: #ef4444;'>{ATTEMPT_LIMIT} / {ATTEMPT_LIMIT}</p>
+                    </div>
+                </div>
+                <div style='margin-top: 16px; background:#fef2f2; padding:16px; border-radius:12px; text-align:left; font-size:0.98rem; line-height:1.4;'>
+                    <p style='margin:0; color:#991b1b;'><b>Nice Work!</b> Scroll down to "Finish and Reflect".</p>
+                </div>
+            </div>"""
+            settings = compute_rank_settings(submission_count, model_name_key, complexity_level, feature_set, data_size_str)
+            limit_reached_updates = {
+                submission_feedback_display: gr.update(value=limit_warning_html, visible=True),
+                submit_button: gr.update(value="🛑 Submission Limit Reached", interactive=False),
+                model_type_radio: gr.update(interactive=False), complexity_slider: gr.update(interactive=False),
+                feature_set_checkbox: gr.update(interactive=False), data_size_radio: gr.update(interactive=False),
+                attempts_tracker_display: gr.update(value=f"<div style='text-align:center; padding:8px; margin:8px 0; background:#fef2f2; border-radius:8px; border:1px solid #ef4444;'><p style='margin:0; color:#991b1b; font-weight:600;'>🛑 Attempts used: {ATTEMPT_LIMIT}/{ATTEMPT_LIMIT}</p></div>"),
+                team_leaderboard_display: team_leaderboard_display, individual_leaderboard_display: individual_leaderboard_display,
+                last_submission_score_state: last_submission_score, last_rank_state: last_rank,
+                best_score_state: best_score, submission_count_state: submission_count,
+                first_submission_score_state: first_submission_score, rank_message_display: settings["rank_message"],
+                login_username: gr.update(visible=False), login_password: gr.update(visible=False),
+                login_submit: gr.update(visible=False), login_error: gr.update(visible=False),
+                was_preview_state: False, kpi_meta_state: {}, last_seen_ts_state: None
+            }
+            yield limit_reached_updates
+            return
+        
+        progress(0.5, desc="Submitting to Cloud...")
+        yield { 
+            submission_feedback_display: gr.update(value=get_status_html(3, "Submitting", "Sending model to the competition server..."), visible=True),
+            login_error: gr.update(visible=False)
+        }
+
+        description = f"{model_name_key} (Cplx:{complexity_level} Size:{data_size_str})"
+        tags = f"team:{team_name},model:{model_name_key}"
+
+        # 1. FETCH BASELINE
+        baseline_leaderboard_df = _get_leaderboard_with_optional_token(playground, token)
+        
+        from sklearn.metrics import accuracy_score
+        # Ensure correct type for local accuracy calc
+        if isinstance(predictions, list):
+            local_accuracy_preds = np.array(predictions)
+        else:
+            local_accuracy_preds = predictions
+        local_test_accuracy = accuracy_score(Y_TEST, local_accuracy_preds)
+
+        # 2. SUBMIT & CAPTURE ACCURACY
+        def _submit():
+            # If using cache (tuned_model is None), we pass None for model/preprocessor
+            # and explicitly pass predictions.
+            return playground.submit_model(
+                model=tuned_model, 
+                preprocessor=preprocessor, 
+                prediction_submission=predictions,
+                input_dict={'description': description, 'tags': tags},
+                custom_metadata={'Team': team_name, 'Moral_Compass': 0}, 
+                token=token,
+                return_metrics=["accuracy"] 
+            )
+        
+        try:
+            submit_result = _retry_with_backoff(_submit, description="model submission")
+            if isinstance(submit_result, tuple) and len(submit_result) == 3:
+                _, _, metrics = submit_result
+                if metrics and "accuracy" in metrics and metrics["accuracy"] is not None:
+                    this_submission_score = float(metrics["accuracy"])
+                else:
+                    this_submission_score = local_test_accuracy
+            else:
+                this_submission_score = local_test_accuracy
+        except Exception as e:
+            _log(f"Submission return parsing failed: {e}. Using local accuracy.")
+            this_submission_score = local_test_accuracy
+        
+        _log(f"Submission successful. Server Score: {this_submission_score}")
+
+        try:
+            # Short timeout to trigger the lambda without hanging the UI
+            _log("Triggering backend merge...")
+            playground.get_leaderboard(token=token) 
+        except Exception:
+            # We ignore errors here because the 'submit_model' post 
+            # already succeeded. This is just a cleanup task.
+            pass
+        # -------------------------------------------------------------------------
+
+        # Immediately increment submission count...
+        new_submission_count = submission_count + 1
+        new_first_submission_score = first_submission_score
+        if submission_count == 0 and first_submission_score is None:
+            new_first_submission_score = this_submission_score
+
+        # --- Stage 4: Local Rank Calculation (Optimistic) ---
+        progress(0.9, desc="Calculating Rank...")
+        
+        # 3. SIMULATE UPDATED LEADERBOARD
+        simulated_df = baseline_leaderboard_df.copy() if baseline_leaderboard_df is not None else pd.DataFrame()
+        
+        # We use pd.Timestamp.now() to ensure pandas sorting logic sees this as the absolute latest
+        new_row = pd.DataFrame([{
+            "username": username,
+            "accuracy": this_submission_score,
+            "Team": team_name,
+            "timestamp": pd.Timestamp.now(), 
+            "version": "latest"
+        }])
+        
+        if not simulated_df.empty:
+            simulated_df = pd.concat([simulated_df, new_row], ignore_index=True)
+        else:
+            simulated_df = new_row
+
+        # 4. GENERATE TABLES (Use helper for tables only)
+        # We ignore the kpi_card return from this function because it might use internal sorting 
+        # that doesn't respect our new row perfectly.
+        team_html, individual_html, _, new_best_accuracy, new_rank, _ = generate_competitive_summary(
+            simulated_df, team_name, username, last_submission_score, last_rank, submission_count
+        )
+
+        # 5. GENERATE KPI CARD EXPLICITLY (The Authority Fix)
+        # We manually build the card using the score we KNOW we just got.
+        kpi_card_html = _build_kpi_card_html(
+            new_score=this_submission_score,
+            last_score=last_submission_score,
+            new_rank=new_rank,
+            last_rank=last_rank,
+            submission_count=submission_count, 
+            is_preview=False,
+            is_pending=False
+        )
+
+        # --- Stage 5: Final UI Update ---
+        progress(1.0, desc="Complete!")
+        
+        success_kpi_meta = {
+            "was_preview": False, "preview_score": None, "ready_at_run_start": ready,
+            "poll_iterations": 0, "local_test_accuracy": local_test_accuracy,
+            "this_submission_score": this_submission_score, "new_best_accuracy": new_best_accuracy,
+            "rank": new_rank, "pending": False, "optimistic_fallback": True 
+        }
+        
+        settings = compute_rank_settings(new_submission_count, model_name_key, complexity_level, feature_set, data_size_str)
+
+        # -------------------------------------------------------------------------
+        # NEW LOGIC: Check for Limit Reached immediately AFTER this submission
+        # -------------------------------------------------------------------------
+        limit_reached = new_submission_count >= ATTEMPT_LIMIT
+        
+        # Prepare the UI state based on whether limit is reached
+        if limit_reached:
+            # 1. Append the Limit Warning HTML *below* the Result Card
+            limit_html = f"""
+            <div style='margin-top: 16px; border: 2px solid #ef4444; background:#fef2f2; padding:16px; border-radius:12px; text-align:left;'>
+                <h3 style='margin:0 0 8px 0; color:#991b1b;'>🛑 Submission Limit Reached ({ATTEMPT_LIMIT}/{ATTEMPT_LIMIT})</h3>
+                <p style='margin:0; color:#7f1d1d; line-height:1.4;'>
+                    <b>You have used all your attempts for this session.</b><br>
+                    Review your final results above, then scroll down to "Finish and Reflect" to continue.
+                </p>
+            </div>
+            """
+            final_html_display = kpi_card_html + limit_html
+            
+            # 2. Disable all controls
+            button_update = gr.update(value="🛑 Limit Reached", interactive=False)
+            interactive_state = False
+            tracker_html = f"<div style='text-align:center; padding:8px; margin:8px 0; background:#fef2f2; border-radius:8px; border:1px solid #ef4444;'><p style='margin:0; color:#991b1b; font-weight:600;'>🛑 Attempts used: {ATTEMPT_LIMIT}/{ATTEMPT_LIMIT} (Max)</p></div>"
+        
+        else:
+            # Normal State: Show just the result card and keep controls active
+            final_html_display = kpi_card_html
+            button_update = gr.update(value="🔬 Build & Submit Model", interactive=True)
+            interactive_state = True
+            tracker_html = _build_attempts_tracker_html(new_submission_count)
+
+        # -------------------------------------------------------------------------
+
+        final_updates = {
+            submission_feedback_display: gr.update(value=final_html_display, visible=True),
+            team_leaderboard_display: team_html,
+            individual_leaderboard_display: individual_html,
+            last_submission_score_state: this_submission_score, 
+            last_rank_state: new_rank, 
+            best_score_state: new_best_accuracy,
+            submission_count_state: new_submission_count,
+            first_submission_score_state: new_first_submission_score,
+            rank_message_display: settings["rank_message"],
+            
+            # Apply the interactive state calculated above
+            model_type_radio: gr.update(choices=settings["model_choices"], value=settings["model_value"], interactive=(settings["model_interactive"] and interactive_state)),
+            complexity_slider: gr.update(minimum=1, maximum=settings["complexity_max"], value=settings["complexity_value"], interactive=interactive_state),
+            feature_set_checkbox: gr.update(choices=settings["feature_set_choices"], value=settings["feature_set_value"], interactive=(settings["feature_set_interactive"] and interactive_state)),
+            data_size_radio: gr.update(choices=settings["data_size_choices"], value=settings["data_size_value"], interactive=(settings["data_size_interactive"] and interactive_state)),
+            
+            submit_button: button_update,
+            
+            login_username: gr.update(visible=False), login_password: gr.update(visible=False),
+            login_submit: gr.update(visible=False), login_error: gr.update(visible=False),
+            attempts_tracker_display: gr.update(value=tracker_html),
+            was_preview_state: False,
+            kpi_meta_state: success_kpi_meta,
+            last_seen_ts_state: time.time()
+        }
+        yield final_updates
+      
+    except Exception as e:
+        error_msg = f"ERROR: {e}"
+        _log(f"Exception in run_experiment: {error_msg}")
+        settings = compute_rank_settings(
+             submission_count, model_name_key, complexity_level, feature_set, data_size_str
+        )
+        
+        exception_kpi_meta = {
+            "was_preview": False, "preview_score": None, "ready_at_run_start": ready if 'ready' in locals() else False,
+            "poll_iterations": 0, "local_test_accuracy": None, "this_submission_score": None,
+            "new_best_accuracy": None, "rank": None, "error": str(e)
+        }
+        
+        error_updates = {
+            submission_feedback_display: gr.update(
+                f"<p style='text-align:center; color:red; padding:20px 0;'>An error occurred: {error_msg}</p>", visible=True
+            ),
+            team_leaderboard_display: f"<p style='text-align:center; color:red; padding-top:20px;'>An error occurred: {error_msg}</p>",
+            individual_leaderboard_display: f"<p style='text-align:center; color:red; padding-top:20px;'>An error occurred: {error_msg}</p>",
+            last_submission_score_state: last_submission_score,
+            last_rank_state: last_rank,
+            best_score_state: best_score,
+            submission_count_state: submission_count,
+            first_submission_score_state: first_submission_score,
+            rank_message_display: settings["rank_message"],
+            model_type_radio: gr.update(choices=settings["model_choices"], value=settings["model_value"], interactive=settings["model_interactive"]),
+            complexity_slider: gr.update(minimum=1, maximum=settings["complexity_max"], value=settings["complexity_value"]),
+            feature_set_checkbox: gr.update(choices=settings["feature_set_choices"], value=settings["feature_set_value"], interactive=settings["feature_set_interactive"]),
+            data_size_radio: gr.update(choices=settings["data_size_choices"], value=settings["data_size_value"], interactive=settings["data_size_interactive"]),
+            submit_button: gr.update(value="🔬 Build & Submit Model", interactive=True),
+            login_username: gr.update(visible=False),
+            login_password: gr.update(visible=False),
+            login_submit: gr.update(visible=False),
+            login_error: gr.update(visible=False),
+            attempts_tracker_display: gr.update(value=_build_attempts_tracker_html(submission_count)),
+            was_preview_state: False,
+            kpi_meta_state: exception_kpi_meta,
+            last_seen_ts_state: None
+        }
+        yield error_updates
+
+def on_initial_load(username, token=None, team_name=""):
+    """
+    Updated to show "Welcome & CTA" if the SPECIFIC USER has 0 submissions,
+    even if the leaderboard/team already has data from others.
+    """
+    initial_ui = compute_rank_settings(
+        0, DEFAULT_MODEL, 2, DEFAULT_FEATURE_SET, DEFAULT_DATA_SIZE
+    )
+
+    # 1. Prepare the Welcome HTML
+    display_team = team_name if team_name else "Your Team"
+    
+    welcome_html = f"""
+    <div style='text-align:center; padding: 30px 20px;'>
+        <div style='font-size: 3rem; margin-bottom: 10px;'>👋</div>
+        <h3 style='margin: 0 0 8px 0; color: #111827; font-size: 1.5rem;'>Welcome to <b>{display_team}</b>!</h3>
+        <p style='font-size: 1.1rem; color: #4b5563; margin: 0 0 20px 0;'>
+            Your team is waiting for your help to improve the AI.
+        </p>
+        
+        <div style='background:#eff6ff; padding:16px; border-radius:12px; border:2px solid #bfdbfe; display:inline-block;'>
+            <p style='margin:0; color:#1e40af; font-weight:bold; font-size:1.1rem;'>
+                👈 Click "Build & Submit Model" to Start Playing!
+            </p>
+        </div>
+    </div>
+    """
+
+    # Check background init
+    with INIT_LOCK:
+        background_ready = INIT_FLAGS["leaderboard"]
+    
+    should_attempt_fetch = background_ready or (token is not None)
+    full_leaderboard_df = None
+    
+    if should_attempt_fetch:
+        try:
+            if playground:
+                full_leaderboard_df = _get_leaderboard_with_optional_token(playground, token)
+        except Exception as e:
+            print(f"Error on initial load fetch: {e}")
+            full_leaderboard_df = None
+
+    # -------------------------------------------------------------------------
+    # LOGIC UPDATE: Check if THIS user has submitted anything
+    # -------------------------------------------------------------------------
+    user_has_submitted = False
+    if full_leaderboard_df is not None and not full_leaderboard_df.empty:
+        if "username" in full_leaderboard_df.columns and username:
+            # Check if the username exists in the dataframe
+            user_has_submitted = username in full_leaderboard_df["username"].values
+
+    # Decision Logic
+    if not user_has_submitted:
+        # CASE 1: New User (or first time loading session) -> FORCE WELCOME
+        # regardless of whether the leaderboard has other people's data.
+        team_html = welcome_html
+        individual_html = "<p style='text-align:center; color:#6b7280; padding-top:40px;'>Submit your model to see where you rank!</p>"
+        
+    elif full_leaderboard_df is None or full_leaderboard_df.empty:
+        # CASE 2: Returning user, but data fetch failed -> Show Skeleton
+        team_html = _build_skeleton_leaderboard(rows=6, is_team=True)
+        individual_html = _build_skeleton_leaderboard(rows=6, is_team=False)
+        
+    else:
+        # CASE 3: Returning user WITH data -> Show Real Tables
+        try:
+            team_html, individual_html, _, _, _, _ = generate_competitive_summary(
+                full_leaderboard_df,
+                team_name,
+                username,
+                0, 0, -1
+            )
+        except Exception as e:
+            print(f"Error generating summary HTML: {e}")
+            team_html = "<p style='text-align:center; color:red; padding-top:20px;'>Error rendering leaderboard.</p>"
+            individual_html = "<p style='text-align:center; color:red; padding-top:20px;'>Error rendering leaderboard.</p>"
+
+    return (
+        get_model_card(DEFAULT_MODEL),
+        team_html,
+        individual_html,
+        initial_ui["rank_message"],
+        gr.update(choices=initial_ui["model_choices"], value=initial_ui["model_value"], interactive=initial_ui["model_interactive"]),
+        gr.update(minimum=1, maximum=initial_ui["complexity_max"], value=initial_ui["complexity_value"]),
+        gr.update(choices=initial_ui["feature_set_choices"], value=initial_ui["feature_set_value"], interactive=initial_ui["feature_set_interactive"]),
+        gr.update(choices=initial_ui["data_size_choices"], value=initial_ui["data_size_value"], interactive=initial_ui["data_size_interactive"]),
+    )
+
+
+# -------------------------------------------------------------------------
+# Conclusion helpers (dark/light mode aware)
+# -------------------------------------------------------------------------
+def build_final_conclusion_html(best_score, submissions, rank, first_score, feature_set):
+    """
+    Build the final conclusion HTML with performance summary.
+    Colors are handled via CSS classes so that light/dark mode work correctly.
+    """
+    unlocked_tiers = min(3, max(0, submissions - 1))  # 0..3
+    tier_names = ["Trainee", "Junior", "Senior", "Lead"]
+    reached = tier_names[: unlocked_tiers + 1]
+    tier_line = " → ".join([f"{t}{' ✅' if t in reached else ''}" for t in tier_names])
+
+    improvement = (best_score - first_score) if (first_score is not None and submissions > 1) else 0.0
+    strong_predictors = {"age", "length_of_stay", "priors_count", "age_cat"}
+    strong_used = [f for f in feature_set if f in strong_predictors]
+
+    ethical_note = (
+        "You unlocked powerful predictors. Consider: Would removing demographic fields change fairness? "
+        "In the next section we will begin to investigate this question further."
+    )
+
+    # Tailor message for very few submissions
+    tip_html = ""
+    if submissions < 2:
+        tip_html = """
+        <div class="final-conclusion-tip">
+          <b>Tip:</b> Try at least 2–3 submissions changing ONE setting at a time to see clear cause/effect.
+        </div>
+        """
+
+    # Add note if user reached the attempt cap
+    attempt_cap_html = ""
+    if submissions >= ATTEMPT_LIMIT:
+        attempt_cap_html = f"""
+        <div class="final-conclusion-attempt-cap">
+          <p style="margin:0;">
+            <b>📊 Attempt Limit Reached:</b> You used all {ATTEMPT_LIMIT} allowed submission attempts for this session.
+            We will open up submissions again after you complete some new activities next.
+          </p>
+        </div>
+        """
+
+    return f"""
+    <div class="final-conclusion-root">
+      <h1 class="final-conclusion-title">🎉 Engineering Phase Complete</h1>
+      <div class="final-conclusion-card">
+        <h2 class="final-conclusion-subtitle">Your Performance Snapshot</h2>
+        <ul class="final-conclusion-list">
+          <li>🏁 <b>Best Accuracy:</b> {(best_score * 100):.2f}%</li>
+          <li>📊 <b>Rank Achieved:</b> {('#' + str(rank)) if rank > 0 else '—'}</li>
+          <li>🔁 <b>Submissions Made This Session:</b> {submissions}{' / ' + str(ATTEMPT_LIMIT) if submissions >= ATTEMPT_LIMIT else ''}</li>
+          <li>🧗 <b>Improvement Over First Score This Session:</b> {(improvement * 100):+.2f}</li>
+          <li>🎖️ <b>Tier Progress:</b> {tier_line}</li>
+          <li>🧪 <b>Strong Predictors Used:</b> {len(strong_used)} ({', '.join(strong_used) if strong_used else 'None yet'})</li>
+        </ul>
+
+        {tip_html}
+
+        <div class="final-conclusion-ethics">
+          <p style="margin:0;"><b>Ethical Reflection:</b> {ethical_note}</p>
+        </div>
+
+        {attempt_cap_html}
+
+        <hr class="final-conclusion-divider" />
+
+        <div class="final-conclusion-next">
+          <h2>➡️ Next: Real-World Consequences</h2>
+          <p>Scroll below this app to continue. You'll examine how models like yours shape judicial outcomes.</p>
+          <h1 class="final-conclusion-scroll">👇 SCROLL DOWN 👇</h1>
+        </div>
+      </div>
+    </div>
+    """
+
+
+
+def build_conclusion_from_state(best_score, submissions, rank, first_score, feature_set):
+    return build_final_conclusion_html(best_score, submissions, rank, first_score, feature_set)
+def create_model_building_game_es_final_app(theme_primary_hue: str = "indigo") -> "gr.Blocks":
+    """
+    Create (but do not launch) the model building game app.
+    """
+    start_background_init()
+
+    # Add missing globals (FIX)
+    global submit_button, submission_feedback_display, team_leaderboard_display
+    global individual_leaderboard_display, last_submission_score_state, last_rank_state
+    global best_score_state, submission_count_state, first_submission_score_state
+    global rank_message_display, model_type_radio, complexity_slider
+    global feature_set_checkbox, data_size_radio
+    global login_username, login_password, login_submit, login_error
+    global attempts_tracker_display, team_name_state
+    global username_state, token_state  # <-- Added
+    global readiness_state, was_preview_state, kpi_meta_state  # <-- Added for parameter shadowing guards
+    global last_seen_ts_state  # <-- Added for timestamp tracking
+    
+    css = """
+    /* ------------------------------
+      Shared Design Tokens (local)
+      ------------------------------ */
+
+    /* We keep everything driven by Gradio theme vars:
+      --body-background-fill, --body-text-color, --secondary-text-color,
+      --border-color-primary, --block-background-fill, --color-accent,
+      --shadow-drop, --prose-background-fill
+    */
+
+    :root {
+        --slide-radius-md: 12px;
+        --slide-radius-lg: 16px;
+        --slide-radius-xl: 18px;
+        --slide-spacing-lg: 24px;
+
+        /* Local, non-brand tokens built *on top of* theme vars */
+        --card-bg-soft: var(--block-background-fill);
+        --card-bg-strong: var(--prose-background-fill, var(--block-background-fill));
+        --card-border-subtle: var(--border-color-primary);
+        --accent-strong: var(--color-accent);
+        --text-main: var(--body-text-color);
+        --text-muted: var(--secondary-text-color);
+    }
+
+    /* ------------------------------------------------------------------
+      Base Layout Helpers
+      ------------------------------------------------------------------ */
+
+    .slide-content {
+        max-width: 900px;
+        margin-left: auto;
+        margin-right: auto;
+    }
+
+    /* Shared card-like panels used throughout slides */
+    .panel-box {
+        background: var(--card-bg-soft);
+        padding: 20px;
+        border-radius: var(--slide-radius-lg);
+        border: 2px solid var(--card-border-subtle);
+        margin-bottom: 18px;
+        color: var(--text-main);
+        box-shadow: var(--shadow-drop, 0 2px 4px rgba(0,0,0,0.04));
+    }
+
+    .leaderboard-box {
+        background: var(--card-bg-soft);
+        padding: 20px;
+        border-radius: var(--slide-radius-lg);
+        border: 1px solid var(--card-border-subtle);
+        margin-top: 12px;
+        color: var(--text-main);
+    }
+
+    /* For “explanatory UI” scaffolding */
+    .mock-ui-box {
+        background: var(--card-bg-strong);
+        border: 2px solid var(--card-border-subtle);
+        padding: 24px;
+        border-radius: var(--slide-radius-lg);
+        color: var(--text-main);
+    }
+
+    .mock-ui-inner {
+        background: var(--block-background-fill);
+        border: 1px solid var(--card-border-subtle);
+        padding: 24px;
+        border-radius: var(--slide-radius-md);
+    }
+
+    /* “Control box” inside the mock UI */
+    .mock-ui-control-box {
+        padding: 12px;
+        background: var(--block-background-fill);
+        border-radius: 8px;
+        border: 1px solid var(--card-border-subtle);
+    }
+
+    /* Little radio / check icons */
+    .mock-ui-radio-on {
+        font-size: 1.5rem;
+        vertical-align: middle;
+        color: var(--accent-strong);
+    }
+
+    .mock-ui-radio-off {
+        font-size: 1.5rem;
+        vertical-align: middle;
+        color: var(--text-muted);
+    }
+
+    .mock-ui-slider-text {
+        font-size: 1.5rem;
+        margin: 0;
+        color: var(--accent-strong);
+        letter-spacing: 4px;
+    }
+
+    .mock-ui-slider-bar {
+        color: var(--text-muted);
+    }
+
+    /* Simple mock button representation */
+    .mock-button {
+        width: 100%;
+        font-size: 1.25rem;
+        font-weight: 600;
+        padding: 16px 24px;
+        background-color: var(--accent-strong);
+        color: var(--body-background-fill);
+        border: none;
+        border-radius: 8px;
+        cursor: not-allowed;
+    }
+
+    /* Step visuals on slides */
+    .step-visual {
+        display: flex;
+        flex-wrap: wrap;
+        justify-content: space-around;
+        align-items: center;
+        margin: 24px 0;
+        text-align: center;
+        font-size: 1rem;
+    }
+
+    .step-visual-box {
+        padding: 16px;
+        background: var(--block-background-fill);   /* ✅ theme-aware */
+        border-radius: 8px;
+        border: 2px solid var(--border-color-primary);
+        margin: 5px;
+        color: var(--body-text-color);              /* optional, safe */
+    }
+
+    .step-visual-arrow {
+        font-size: 2rem;
+        margin: 5px;
+        /* no explicit color – inherit from theme or override in dark mode */
+    }
+
+    /* ------------------------------------------------------------------
+      KPI Card (score feedback)
+      ------------------------------------------------------------------ */
+
+    .kpi-card {
+        background: var(--card-bg-strong);
+        border: 2px solid var(--accent-strong);
+        padding: 24px;
+        border-radius: var(--slide-radius-lg);
+        text-align: center;
+        max-width: 600px;
+        margin: auto;
+        color: var(--text-main);
+        box-shadow: var(--shadow-drop, 0 4px 6px -1px rgba(0,0,0,0.08));
+        min-height: 200px; /* prevent layout shift */
+    }
+
+    .kpi-card-body {
+        display: flex;
+        flex-wrap: wrap;
+        justify-content: space-around;
+        align-items: flex-end;
+        margin-top: 24px;
+    }
+
+    .kpi-metric-box {
+        min-width: 150px;
+        margin: 10px;
+    }
+
+    .kpi-label {
+        font-size: 1rem;
+        color: var(--text-muted);
+        margin: 0;
+    }
+
+    .kpi-score {
+        font-size: 3rem;
+        font-weight: 700;
+        margin: 0;
+        line-height: 1.1;
+        color: var(--accent-strong);
+    }
+
+    .kpi-subtext-muted {
+        font-size: 1.2rem;
+        font-weight: 500;
+        color: var(--text-muted);
+        margin: 0;
+        padding-top: 8px;
+    }
+
+    /* Small variants to hint semantic state without hard-coded colors */
+    .kpi-card--neutral {
+        border-color: var(--card-border-subtle);
+    }
+
+    .kpi-card--subtle-accent {
+        border-color: var(--accent-strong);
+    }
+
+    .kpi-score--muted {
+        color: var(--text-muted);
+    }
+
+    /* ------------------------------------------------------------------
+      Leaderboard Table + Placeholder
+      ------------------------------------------------------------------ */
+
+    .leaderboard-html-table {
+        width: 100%;
+        border-collapse: collapse;
+        text-align: left;
+        font-size: 1rem;
+        color: var(--text-main);
+        min-height: 300px; /* Stable height */
+    }
+
+    .leaderboard-html-table thead {
+        background: var(--block-background-fill);
+    }
+
+    .leaderboard-html-table th {
+        padding: 12px 16px;
+        font-size: 0.9rem;
+        color: var(--text-muted);
+        font-weight: 500;
+    }
+
+    .leaderboard-html-table tbody tr {
+        border-bottom: 1px solid var(--card-border-subtle);
+    }
+
+    .leaderboard-html-table td {
+        padding: 12px 16px;
+    }
+
+    .leaderboard-html-table .user-row-highlight {
+        background: rgba( var(--color-accent-rgb, 59,130,246), 0.1 );
+        font-weight: 600;
+        color: var(--accent-strong);
+    }
+
+    /* Static placeholder (no shimmer, no animation) */
+    .lb-placeholder {
+        min-height: 300px;
+        display: flex;
+        flex-direction: column;
+        align-items: center;
+        justify-content: center;
+        background: var(--block-background-fill);
+        border: 1px solid var(--card-border-subtle);
+        border-radius: 12px;
+        padding: 40px 20px;
+        text-align: center;
+    }
+
+    .lb-placeholder-title {
+        font-size: 1.25rem;
+        font-weight: 500;
+        color: var(--text-muted);
+        margin-bottom: 8px;
+    }
+
+    .lb-placeholder-sub {
+        font-size: 1rem;
+        color: var(--text-muted);
+    }
+
+    /* ------------------------------------------------------------------
+      Processing / “Experiment running” status
+      ------------------------------------------------------------------ */
+
+    .processing-status {
+        background: var(--block-background-fill);
+        border: 2px solid var(--accent-strong);
+        border-radius: 16px;
+        padding: 30px;
+        text-align: center;
+        box-shadow: var(--shadow-drop, 0 4px 6px rgba(0,0,0,0.12));
+        animation: pulse-indigo 2s infinite;
+        color: var(--text-main);
+    }
+
+    .processing-icon {
+        font-size: 4rem;
+        margin-bottom: 10px;
+        display: block;
+        animation: spin-slow 3s linear infinite;
+    }
+
+    .processing-text {
+        font-size: 1.5rem;
+        font-weight: 700;
+        color: var(--accent-strong);
+    }
+
+    .processing-subtext {
+        font-size: 1.1rem;
+        color: var(--text-muted);
+        margin-top: 8px;
+    }
+
+    /* Pulse & spin animations */
+    @keyframes pulse-indigo {
+        0%   { box-shadow: 0 0 0 0 rgba(99, 102, 241, 0.4); }
+        70%  { box-shadow: 0 0 0 15px rgba(99, 102, 241, 0); }
+        100% { box-shadow: 0 0 0 0 rgba(99, 102, 241, 0); }
+    }
+
+    @keyframes spin-slow {
+        from { transform: rotate(0deg); }
+        to   { transform: rotate(360deg); }
+    }
+
+    /* Conclusion arrow pulse */
+    @keyframes pulseArrow {
+        0%   { transform: scale(1);     opacity: 1; }
+        50%  { transform: scale(1.08);  opacity: 0.85; }
+        100% { transform: scale(1);     opacity: 1; }
+    }
+
+    @media (prefers-reduced-motion: reduce) {
+        [style*='pulseArrow'] {
+            animation: none !important;
+        }
+        .processing-status,
+        .processing-icon {
+            animation: none !important;
+        }
+    }
+
+    /* ------------------------------------------------------------------
+      Attempts Tracker + Init Banner + Alerts
+      ------------------------------------------------------------------ */
+
+    .init-banner {
+        background: var(--card-bg-strong);
+        padding: 12px;
+        border-radius: 8px;
+        text-align: center;
+        margin-bottom: 16px;
+        border: 1px solid var(--card-border-subtle);
+        color: var(--text-main);
+    }
+
+    .init-banner__text {
+        margin: 0;
+        font-weight: 500;
+        color: var(--text-muted);
+    }
+
+    /* Attempts tracker shell */
+    .attempts-tracker {
+        text-align: center;
+        padding: 8px;
+        margin: 8px 0;
+        background: var(--block-background-fill);
+        border-radius: 8px;
+        border: 1px solid var(--card-border-subtle);
+    }
+
+    .attempts-tracker__text {
+        margin: 0;
+        font-weight: 600;
+        font-size: 1rem;
+        color: var(--accent-strong);
+    }
+
+    /* Limit reached variant – we *still* stick to theme colors */
+    .attempts-tracker--limit .attempts-tracker__text {
+        color: var(--text-main);
+    }
+
+    /* Generic alert helpers used in inline login messages */
+    .alert {
+        padding: 12px 16px;
+        border-radius: 8px;
+        margin-top: 12px;
+        text-align: left;
+        font-size: 0.95rem;
+    }
+
+    .alert--error {
+        border-left: 4px solid var(--accent-strong);
+        background: var(--block-background-fill);
+        color: var(--text-main);
+    }
+
+    .alert--success {
+        border-left: 4px solid var(--accent-strong);
+        background: var(--block-background-fill);
+        color: var(--text-main);
+    }
+
+    .alert__title {
+        margin: 0;
+        font-weight: 600;
+        color: var(--text-main);
+    }
+
+    .alert__body {
+        margin: 8px 0 0 0;
+        color: var(--text-muted);
+    }
+
+    /* ------------------------------------------------------------------
+      Navigation Loading Overlay
+      ------------------------------------------------------------------ */
+
+    #nav-loading-overlay {
+        position: fixed;
+        top: 0;
+        left: 0;
+        width: 100%;
+        height: 100%;
+        background: color-mix(in srgb, var(--body-background-fill) 90%, transparent);
+        z-index: 9999;
+        display: none;
+        flex-direction: column;
+        align-items: center;
+        justify-content: center;
+        opacity: 0;
+        transition: opacity 0.3s ease;
+    }
+
+    .nav-spinner {
+        width: 50px;
+        height: 50px;
+        border: 5px solid var(--card-border-subtle);
+        border-top: 5px solid var(--accent-strong);
+        border-radius: 50%;
+        animation: nav-spin 1s linear infinite;
+        margin-bottom: 20px;
+    }
+
+    @keyframes nav-spin {
+        0%   { transform: rotate(0deg); }
+        100% { transform: rotate(360deg); }
+    }
+
+    #nav-loading-text {
+        font-size: 1.3rem;
+        font-weight: 600;
+        color: var(--accent-strong);
+    }
+
+    /* ------------------------------------------------------------------
+      Utility: Image inversion for dark mode (if needed)
+      ------------------------------------------------------------------ */
+
+    .dark-invert-image {
+        filter: invert(0);
+    }
+
+    @media (prefers-color-scheme: dark) {
+        .dark-invert-image {
+            filter: invert(1) hue-rotate(180deg);
+        }
+    }
+
+    /* ------------------------------------------------------------------
+      Dark Mode Specific Fine Tuning
+      ------------------------------------------------------------------ */
+
+    @media (prefers-color-scheme: dark) {
+        .panel-box,
+        .leaderboard-box,
+        .mock-ui-box,
+        .mock-ui-inner,
+        .processing-status,
+        .kpi-card {
+            background: color-mix(in srgb, var(--block-background-fill) 85%, #000 15%);
+            border-color: color-mix(in srgb, var(--card-border-subtle) 70%, var(--accent-strong) 30%);
+        }
+
+        .leaderboard-html-table thead {
+            background: color-mix(in srgb, var(--block-background-fill) 75%, #000 25%);
+        }
+
+        .lb-placeholder {
+            background: color-mix(in srgb, var(--block-background-fill) 75%, #000 25%);
+        }
+
+        #nav-loading-overlay {
+            background: color-mix(in srgb, #000 70%, var(--body-background-fill) 30%);
+        }
+    }
+    
+    /* ---------- Conclusion Card Theme Tokens ---------- */
+
+    /* Light theme defaults */
+    :root,
+    :root[data-theme="light"] {
+        --conclusion-card-bg: #e0f2fe;          /* light sky */
+        --conclusion-card-border: #0369a1;      /* sky-700 */
+        --conclusion-card-fg: #0f172a;          /* slate-900 */
+
+        --conclusion-tip-bg: #fef9c3;           /* amber-100 */
+        --conclusion-tip-border: #f59e0b;       /* amber-500 */
+        --conclusion-tip-fg: #713f12;           /* amber-900 */
+
+        --conclusion-ethics-bg: #fef2f2;        /* red-50 */
+        --conclusion-ethics-border: #ef4444;    /* red-500 */
+        --conclusion-ethics-fg: #7f1d1d;        /* red-900 */
+
+        --conclusion-attempt-bg: #fee2e2;       /* red-100 */
+        --conclusion-attempt-border: #ef4444;   /* red-500 */
+        --conclusion-attempt-fg: #7f1d1d;       /* red-900 */
+
+        --conclusion-next-fg: #0f172a;          /* main text color */
+    }
+
+    /* Dark theme overrides – keep contrast high on dark background */
+    [data-theme="dark"] {
+        --conclusion-card-bg: #020617;          /* slate-950 */
+        --conclusion-card-border: #38bdf8;      /* sky-400 */
+        --conclusion-card-fg: #e5e7eb;          /* slate-200 */
+
+        --conclusion-tip-bg: rgba(250, 204, 21, 0.08);   /* soft amber tint */
+        --conclusion-tip-border: #facc15;                /* amber-400 */
+        --conclusion-tip-fg: #facc15;
+
+        --conclusion-ethics-bg: rgba(248, 113, 113, 0.10); /* soft red tint */
+        --conclusion-ethics-border: #f97373;               /* red-ish */
+        --conclusion-ethics-fg: #fecaca;
+
+        --conclusion-attempt-bg: rgba(248, 113, 113, 0.16);
+        --conclusion-attempt-border: #f97373;
+        --conclusion-attempt-fg: #fee2e2;
+
+        --conclusion-next-fg: #e5e7eb;
+    }
+
+    /* ---------- Conclusion Layout ---------- */
+
+    .app-conclusion-wrapper {
+        text-align: center;
+    }
+
+    .app-conclusion-title {
+        font-size: 2.4rem;
+        margin: 0;
+    }
+
+    .app-conclusion-card {
+        margin-top: 24px;
+        max-width: 950px;
+        margin-left: auto;
+        margin-right: auto;
+        padding: 28px;
+        border-radius: 18px;
+        border-width: 3px;
+        border-style: solid;
+        background: var(--conclusion-card-bg);
+        border-color: var(--conclusion-card-border);
+        color: var(--conclusion-card-fg);
+        box-shadow: 0 20px 40px rgba(15, 23, 42, 0.25);
+    }
+
+    .app-conclusion-subtitle {
+        margin-top: 0;
+        font-size: 1.5rem;
+    }
+
+    .app-conclusion-metrics {
+        list-style: none;
+        padding: 0;
+        font-size: 1.05rem;
+        text-align: left;
+        max-width: 640px;
+        margin: 20px auto;
+    }
+
+    /* ---------- Generic panel helpers reused here ---------- */
+
+    .app-panel-tip,
+    .app-panel-critical,
+    .app-panel-warning {
+        padding: 16px;
+        border-radius: 12px;
+        border-left-width: 6px;
+        border-left-style: solid;
+        text-align: left;
+        font-size: 0.98rem;
+        line-height: 1.4;
+        margin-top: 16px;
+    }
+
+    .app-panel-title {
+        margin: 0 0 4px 0;
+        font-weight: 700;
+    }
+
+    .app-panel-body {
+        margin: 0;
+    }
+
+    /* Specific variants */
+
+    .app-conclusion-tip.app-panel-tip {
+        background: var(--conclusion-tip-bg);
+        border-left-color: var(--conclusion-tip-border);
+        color: var(--conclusion-tip-fg);
+    }
+
+    .app-conclusion-ethics.app-panel-critical {
+        background: var(--conclusion-ethics-bg);
+        border-left-color: var(--conclusion-ethics-border);
+        color: var(--conclusion-ethics-fg);
+    }
+
+    .app-conclusion-attempt-cap.app-panel-warning {
+        background: var(--conclusion-attempt-bg);
+        border-left-color: var(--conclusion-attempt-border);
+        color: var(--conclusion-attempt-fg);
+    }
+
+    /* Divider + next section */
+
+    .app-conclusion-divider {
+        margin: 28px 0;
+        border: 0;
+        border-top: 2px solid rgba(148, 163, 184, 0.8); /* slate-400-ish */
+    }
+
+    .app-conclusion-next-title {
+        margin: 0;
+        color: var(--conclusion-next-fg);
+    }
+
+    .app-conclusion-next-body {
+        font-size: 1rem;
+        color: var(--conclusion-next-fg);
+    }
+
+    /* Arrow inherits the same color, keeps pulse animation defined earlier */
+    .app-conclusion-arrow {
+        margin: 12px 0;
+        font-size: 3rem;
+        animation: pulseArrow 2.5s infinite;
+        color: var(--conclusion-next-fg);
+    }
+
+    /* ---------------------------------------------------- */
+    /* Final Conclusion Slide (Light Mode Defaults)         */
+    /* ---------------------------------------------------- */
+
+    .final-conclusion-root {
+        text-align: center;
+        color: var(--body-text-color);
+    }
+
+    .final-conclusion-title {
+        font-size: 2.4rem;
+        margin: 0;
+    }
+
+    .final-conclusion-card {
+        background-color: var(--block-background-fill);
+        color: var(--body-text-color);
+        padding: 28px;
+        border-radius: 18px;
+        border: 2px solid var(--border-color-primary);
+        margin-top: 24px;
+        max-width: 950px;
+        margin-left: auto;
+        margin-right: auto;
+        box-shadow: var(--shadow-drop, 0 4px 10px rgba(15, 23, 42, 0.08));
+    }
+
+    .final-conclusion-subtitle {
+        margin-top: 0;
+        margin-bottom: 8px;
+    }
+
+    .final-conclusion-list {
+        list-style: none;
+        padding: 0;
+        font-size: 1.05rem;
+        text-align: left;
+        max-width: 640px;
+        margin: 20px auto;
+    }
+
+    .final-conclusion-list li {
+        margin: 4px 0;
+    }
+
+    .final-conclusion-tip {
+        margin-top: 16px;
+        padding: 16px;
+        border-radius: 12px;
+        border-left: 6px solid var(--color-accent);
+        background-color: color-mix(in srgb, var(--color-accent) 12%, transparent);
+        text-align: left;
+        font-size: 0.98rem;
+        line-height: 1.4;
+    }
+
+    .final-conclusion-ethics {
+        margin-top: 16px;
+        padding: 18px;
+        border-radius: 12px;
+        border-left: 6px solid #ef4444;
+        background-color: color-mix(in srgb, #ef4444 10%, transparent);
+        text-align: left;
+        font-size: 0.98rem;
+        line-height: 1.4;
+    }
+
+    .final-conclusion-attempt-cap {
+        margin-top: 16px;
+        padding: 16px;
+        border-radius: 12px;
+        border-left: 6px solid #ef4444;
+        background-color: color-mix(in srgb, #ef4444 16%, transparent);
+        text-align: left;
+        font-size: 0.98rem;
+        line-height: 1.4;
+    }
+
+    .final-conclusion-divider {
+        margin: 28px 0;
+        border: 0;
+        border-top: 2px solid var(--border-color-primary);
+    }
+
+    .final-conclusion-next h2 {
+        margin: 0;
+    }
+
+    .final-conclusion-next p {
+        font-size: 1rem;
+        margin-top: 4px;
+        margin-bottom: 0;
+    }
+
+    .final-conclusion-scroll {
+        margin: 12px 0 0 0;
+        font-size: 3rem;
+        animation: pulseArrow 2.5s infinite;
+    }
+
+    /* ---------------------------------------------------- */
+    /* Dark Mode Overrides for Final Slide                  */
+    /* ---------------------------------------------------- */
+
+    @media (prefers-color-scheme: dark) {
+        .final-conclusion-card {
+            background-color: #0b1120;        /* deep slate */
+            color: white;                     /* 100% contrast confidence */
+            border-color: #38bdf8;
+            box-shadow: none;
+        }
+
+        .final-conclusion-tip {
+            background-color: rgba(56, 189, 248, 0.18);
+        }
+
+        .final-conclusion-ethics {
+            background-color: rgba(248, 113, 113, 0.18);
+        }
+
+        .final-conclusion-attempt-cap {
+            background-color: rgba(248, 113, 113, 0.26);
+        }
+    }
+    /* ---------------------------------------------------- */
+    /* Slide 3: INPUT → MODEL → OUTPUT flow (theme-aware)   */
+    /* ---------------------------------------------------- */
+
+
+    .model-flow {
+        text-align: center;
+        font-weight: 600;
+        font-size: 1.2rem;
+        margin: 20px 0;
+        /* No explicit color – inherit from the card */
+    }
+
+    .model-flow-label {
+        padding: 0 0.1rem;
+        /* No explicit color – inherit */
+    }
+
+    .model-flow-arrow {
+        margin: 0 0.35rem;
+        font-size: 1.4rem;
+        /* No explicit color – inherit */
+    }
+
+    @media (prefers-color-scheme: dark) {
+        .model-flow {
+            color: var(--body-text-color);
+        }
+        .model-flow-arrow {
+            /* In dark mode, nudge arrows toward accent for contrast/confidence */
+            color: color-mix(in srgb, var(--color-accent) 75%, var(--body-text-color) 25%);
+        }
+    }
+    """
+
+
+    # Define globals for yield
+    global submit_button, submission_feedback_display, team_leaderboard_display
+    # --- THIS IS THE FIXED LINE ---
+    global individual_leaderboard_display, last_submission_score_state, last_rank_state, best_score_state, submission_count_state, first_submission_score_state
+    # --- END OF FIX ---
+    global rank_message_display, model_type_radio, complexity_slider
+    global feature_set_checkbox, data_size_radio
+    global login_username, login_password, login_submit, login_error
+    global attempts_tracker_display, team_name_state
+
+    with gr.Blocks(theme=gr.themes.Soft(primary_hue="indigo"), css=css) as demo:
+        # Persistent top anchor for scroll-to-top navigation
+        gr.HTML("<div id='app_top_anchor' style='height:0;'></div>")
+        
+        # Navigation loading overlay with spinner and dynamic message
+        gr.HTML("""
+            <div id='nav-loading-overlay'>
+                <div class='nav-spinner'></div>
+                <span id='nav-loading-text'>Loading...</span>
+            </div>
+        """)
+
+        # Concurrency Note: Do NOT read per-user state from os.environ here.
+        # Username and other per-user data are managed via gr.State objects
+        # and populated during handle_load_with_session_auth.
+
+        # Loading screen
+        with gr.Column(visible=False) as loading_screen:
+            gr.Markdown(
+                """
+                <div style='text-align:center; padding:100px 0;'>
+                    <h2 style='font-size:2rem; color:#6b7280;'>⏳ Loading...</h2>
+                </div>
+                """
+            )
+
+        # --- Briefing Slideshow (Updated with New Cards) ---
+
+        # Slide 7: The Final Transition
+        with gr.Column(visible=True, elem_id="intro-slide") as intro_slide:
+            gr.Markdown("<h1 style='text-align:center;'>🚀 The Final Frontier</h1>")
+            
+            gr.HTML(
+                """
+                <div class='slide-content'>
+                    <div class='panel-box'>
+                        <div style="text-align:center; margin-bottom: 25px;">
+                            <p style="font-size:1.15rem; line-height:1.6;">
+                                You have explored the ethics. You understand the risks.
+                                <br>
+                                Now, it is time to prove you have the technical <strong>Skill</strong>.
+                            </p>
+                        </div>
+
+                        <div style="background:linear-gradient(to right, #eff6ff, white); border:2px solid #3b82f6; border-radius:12px; padding:24px; margin-bottom: 25px;">
+                            <h3 style="margin-top:0; color:#1e40af; text-align:center; font-size:1.4rem;">🛠️ The Accuracy Competition</h3>
+                            <div style="font-size:1.1rem; line-height:1.6; color:#1f2937;">
+                                <p>Your <strong>final mission</strong> is to compete against your peers to build the <strong>most accurate model possible</strong>.</p>
+                                
+                                <p>✨ <strong>Unrestricted Access:</strong> You are now a Lead Engineer. All data inputs and modeling tools are <strong>unlocked immediately</strong>.</p>
+                                
+                                <p>Use every tool at your disposal to climb the leaderboard, but remember the lessons you just learned: 
+                                <em>Accuracy is the goal, but data choices have consequences.</em></p>
+                            </div>
+                        </div>
+
+                        <div style="text-align:center; margin-top:20px; padding-top:10px; border-top: 1px solid #e5e7eb;">
+                            <p style="font-size:1.2rem; font-weight:700; color:#4b5563; margin-bottom:5px;">
+                                Ready to begin?
+                            </p>
+                            <p style="font-size:1rem; color:#6b7280; margin-top:0;">
+                                👇 Click the <b>"Enter the Arena"</b> button below.
+                            </p>
+                        </div>
+                    </div>
+                </div>
+                """
+            )
+            
+            # Only ONE button needed now
+            intro_next_btn = gr.Button("Enter the Arena ▶️", variant="primary", size="lg")
+
+        # --- End Briefing Slideshow ---
+
+
+        # Model Building App (Main Interface)
+        with gr.Column(visible=False, elem_id="model-step") as model_building_step:
+            gr.Markdown("<h1 style='text-align:center;'>🛠️ Model Building Arena</h1>")
+            
+            # Status panel for initialization progress - HIDDEN
+            init_status_display = gr.HTML(value="", visible=False)
+            
+            # Banner for UI state
+
+            init_banner = gr.HTML(
+              value=(
+                  "<div class='init-banner'>"
+                  "<p class='init-banner__text'>"
+                  "⏳ Initializing data & leaderboard… you can explore but must wait for readiness to submit."
+                  "</p>"
+                  "</div>"
+              ),
+              visible=True)
+
+            # Session-based authentication state objects
+            # Concurrency Note: These are initialized to None/empty and populated
+            # during handle_load_with_session_auth. Do NOT use os.environ here.
+            username_state = gr.State(None)
+            token_state = gr.State(None)
+            
+            team_name_state = gr.State(None)  # Populated via handle_load_with_session_auth
+            last_submission_score_state = gr.State(0.0)
+            last_rank_state = gr.State(0)
+            best_score_state = gr.State(0.0)
+            submission_count_state = gr.State(0)
+            first_submission_score_state = gr.State(None)
+            
+            # New states for readiness gating and preview tracking
+            readiness_state = gr.State(False)
+            was_preview_state = gr.State(False)
+            kpi_meta_state = gr.State({})
+            last_seen_ts_state = gr.State(None)  # Track last seen user timestamp
+
+            # Buffered states for all dynamic inputs
+            model_type_state = gr.State(DEFAULT_MODEL)
+            complexity_state = gr.State(2)
+            feature_set_state = gr.State(DEFAULT_FEATURE_SET)
+            data_size_state = gr.State(DEFAULT_DATA_SIZE)
+
+            rank_message_display = gr.Markdown("### Rank loading...")
+            with gr.Row():
+                with gr.Column(scale=1):
+
+                    model_type_radio = gr.Radio(
+                        label="1. Model Strategy",
+                        choices=[],
+                        value=None,
+                        interactive=False
+                    )
+                    model_card_display = gr.Markdown(get_model_card(DEFAULT_MODEL))
+
+                    gr.Markdown("---") # Separator
+
+                    complexity_slider = gr.Slider(
+                        label="2. Model Complexity (1–10)",
+                        minimum=1, maximum=3, step=1, value=2,
+                        info="Higher values allow deeper pattern learning; very high values may overfit."
+                    )
+
+                    gr.Markdown("---") # Separator
+
+                    feature_set_checkbox = gr.CheckboxGroup(
+                        label="3. Select Data Ingredients",
+                        choices=FEATURE_SET_ALL_OPTIONS,
+                        value=DEFAULT_FEATURE_SET,
+                        interactive=False,
+                        info="More ingredients unlock as you rank up!"
+                    )
+
+                    gr.Markdown("---") # Separator
+
+                    data_size_radio = gr.Radio(
+                        label="4. Data Size",
+                        choices=[DEFAULT_DATA_SIZE],
+                        value=DEFAULT_DATA_SIZE,
+                        interactive=False
+                    )
+
+                    gr.Markdown("---") # Separator
+
+                    # Attempt tracker display
+                    attempts_tracker_display = gr.HTML(
+                        value="<div style='text-align:center; padding:8px; margin:8px 0; background:#f0f9ff; border-radius:8px; border:1px solid #bae6fd;'>"
+                        "<p style='margin:0; color:#0369a1; font-weight:600; font-size:1rem;'>📊 Attempts used: 0/10</p>"
+                        "</div>",
+                        visible=True
+                    )
+
+                    submit_button = gr.Button(
+                        value="5. 🔬 Build & Submit Model",
+                        variant="primary",
+                        size="lg"
+                    )
+
+                with gr.Column(scale=1):
+                    gr.HTML(
+                        """
+                        <div class='leaderboard-box'>
+                            <h3 style='margin-top:0;'>🏆 Live Standings</h3>
+                            <p style='margin:0;'>Submit a model to see your rank.</p>
+                        </div>
+                        """
+                    )
+
+                    # KPI Card
+                    submission_feedback_display = gr.HTML(
+                        "<p style='text-align:center; color:#6b7280; padding:20px 0;'>Submit your first model to get feedback!</p>"
+                    )
+                    
+                    # Inline Login Components (initially hidden)
+                    login_username = gr.Textbox(
+                        label="Username",
+                        placeholder="Enter your modelshare.ai username",
+                        visible=False
+                    )
+                    login_password = gr.Textbox(
+                        label="Password",
+                        type="password",
+                        placeholder="Enter your password",
+                        visible=False
+                    )
+                    login_submit = gr.Button(
+                        "Sign In & Submit",
+                        variant="primary",
+                        visible=False
+                    )
+                    login_error = gr.HTML(
+                        value="",
+                        visible=False
+                    )
+
+                    with gr.Tabs():
+                        with gr.TabItem("Team Standings"):
+                            team_leaderboard_display = gr.HTML(
+                                "<p style='text-align:center; color:#6b7280; padding-top:20px;'>Submit a model to see team rankings.</p>"
+                            )
+                        with gr.TabItem("Individual Standings"):
+                            individual_leaderboard_display = gr.HTML(
+                                "<p style='text-align:center; color:#6b7280; padding-top:20px;'>Submit a model to see individual rankings.</p>"
+                            )
+
+            # REMOVED: Ethical Reminder HTML Block
+            step_2_next = gr.Button("Finish & Reflect ▶️", variant="secondary")
+
+        # Conclusion Step
+        with gr.Column(visible=False, elem_id="conclusion-step") as conclusion_step:
+            gr.Markdown("<h1 style='text-align:center;'>✅ Section Complete</h1>")
+            final_score_display = gr.HTML(value="<p>Preparing final summary...</p>")
+            step_3_back = gr.Button("◀️ Back to Experiment")
+
+        # --- Navigation Logic ---
+        all_steps_nav = [
+            intro_slide, 
+            model_building_step, 
+            conclusion_step, 
+            loading_screen
+        ]
+
+        def create_nav(current_step, next_step):
+            """
+            Simplified navigation: directly switches visibility without artificial loading screen.
+            Loading screen only shown when entering arena if not yet ready.
+            """
+            def _nav():
+                # Direct single-step navigation
+                updates = {next_step: gr.update(visible=True)}
+                for s in all_steps_nav:
+                    if s != next_step:
+                        updates[s] = gr.update(visible=False)
+                return updates
+            return _nav
+
+        def finalize_and_show_conclusion(best_score, submissions, rank, first_score, feature_set):
+            """Build dynamic conclusion HTML and navigate to conclusion step."""
+            html = build_final_conclusion_html(best_score, submissions, rank, first_score, feature_set)
+            updates = {
+                conclusion_step: gr.update(visible=True),
+                final_score_display: gr.update(value=html)
+            }
+            for s in all_steps_nav:
+                if s != conclusion_step:
+                    updates[s] = gr.update(visible=False)
+            return [updates[s] if s in updates else gr.update() for s in all_steps_nav] + [html]
+
+        # Helper function to generate navigation JS with loading overlay
+        def nav_js(target_id: str, message: str, min_show_ms: int = 1200) -> str:
+            """
+            Generate JavaScript for enhanced slide navigation with loading overlay.
+            
+            Args:
+                target_id: Element ID of the target slide (e.g., 'slide-2', 'model-step')
+                message: Loading message to display during transition
+                min_show_ms: Minimum time to show overlay (prevents flicker)
+            
+            Returns:
+                JavaScript arrow function string for Gradio's js parameter
+            """
+            return f"""
+()=>{{
+  try {{
+    // Show overlay immediately
+    const overlay = document.getElementById('nav-loading-overlay');
+    const messageEl = document.getElementById('nav-loading-text');
+    if(overlay && messageEl) {{
+      messageEl.textContent = '{message}';
+      overlay.style.display = 'flex';
+      setTimeout(() => {{ overlay.style.opacity = '1'; }}, 10);
+    }}
+    
+    const startTime = Date.now();
+    
+    // Scroll to top after brief delay
+    setTimeout(() => {{
+      const anchor = document.getElementById('app_top_anchor');
+      const container = document.querySelector('.gradio-container') || document.scrollingElement || document.documentElement;
+      
+      function doScroll() {{
+        if(anchor) {{ anchor.scrollIntoView({{behavior:'smooth', block:'start'}}); }}
+        else {{ container.scrollTo({{top:0, behavior:'smooth'}}); }}
+        
+        // Best-effort Colab iframe scroll
+        try {{
+          if(window.parent && window.parent !== window && window.frameElement) {{
+            const top = window.frameElement.getBoundingClientRect().top + window.parent.scrollY;
+            window.parent.scrollTo({{top: Math.max(top - 10, 0), behavior:'smooth'}});
+          }}
+        }} catch(e2) {{}}
+      }}
+      
+      doScroll();
+      // Retry scroll to combat layout shifts
+      let scrollAttempts = 0;
+      const scrollInterval = setInterval(() => {{
+        scrollAttempts++;
+        doScroll();
+        if(scrollAttempts >= 3) clearInterval(scrollInterval);
+      }}, 130);
+    }}, 40);
+    
+    // Poll for target visibility and minimum display time
+    const targetId = '{target_id}';
+    const minShowMs = {min_show_ms};
+    let pollCount = 0;
+    const maxPolls = 77; // ~7 seconds max
+    
+    const pollInterval = setInterval(() => {{
+      pollCount++;
+      const elapsed = Date.now() - startTime;
+      const target = document.getElementById(targetId);
+      const isVisible = target && target.offsetParent !== null && 
+                       window.getComputedStyle(target).display !== 'none';
+      
+      // Hide overlay when target is visible AND minimum time elapsed
+      if((isVisible && elapsed >= minShowMs) || pollCount >= maxPolls) {{
+        clearInterval(pollInterval);
+        if(overlay) {{
+          overlay.style.opacity = '0';
+          setTimeout(() => {{ overlay.style.display = 'none'; }}, 300);
+        }}
+      }}
+    }}, 90);
+    
+  }} catch(e) {{ console.warn('nav-js error', e); }}
+}}
+"""
+        # Final wiring
+        intro_next_btn.click(
+            fn=create_nav(intro_slide, model_building_step),
+            inputs=None, outputs=all_steps_nav,
+            js=nav_js("model-step", "Entering model arena...")
+        )
+
+        # App -> Conclusion (unchanged)
+        step_2_next.click(
+            fn=finalize_and_show_conclusion,
+            inputs=[
+                best_score_state,
+                submission_count_state,
+                last_rank_state,
+                first_submission_score_state,
+                feature_set_state
+            ],
+            outputs=all_steps_nav + [final_score_display],
+            js=nav_js("conclusion-step", "Generating performance summary...")
+        )
+
+        # Conclusion -> App (unchanged)
+        step_3_back.click(
+            fn=create_nav(conclusion_step, model_building_step),
+            inputs=None, outputs=all_steps_nav,
+            js=nav_js("model-step", "Returning to experiment workspace...")
+        )
+
+        # Events
+        model_type_radio.change(
+            fn=get_model_card,
+            inputs=model_type_radio,
+            outputs=model_card_display
+        )
+        model_type_radio.change(
+            fn=lambda v: v or DEFAULT_MODEL,
+            inputs=model_type_radio,
+            outputs=model_type_state
+        )
+        complexity_slider.change(fn=lambda v: v, inputs=complexity_slider, outputs=complexity_state)
+
+        feature_set_checkbox.change(
+            fn=lambda v: v or [],
+            inputs=feature_set_checkbox,
+            outputs=feature_set_state
+        )
+        data_size_radio.change(
+            fn=lambda v: v or DEFAULT_DATA_SIZE,
+            inputs=data_size_radio,
+            outputs=data_size_state
+        )
+
+        all_outputs = [
+            submission_feedback_display,
+            team_leaderboard_display,
+            individual_leaderboard_display,
+            last_submission_score_state,
+            last_rank_state,
+            best_score_state,
+            submission_count_state,
+            first_submission_score_state,
+            rank_message_display,
+            model_type_radio,
+            complexity_slider,
+            feature_set_checkbox,
+            data_size_radio,
+            submit_button,
+            login_username,
+            login_password,
+            login_submit,
+            login_error,
+            attempts_tracker_display,
+            was_preview_state,
+            kpi_meta_state,
+            last_seen_ts_state
+        ]
+
+        # Wire up login button
+        login_submit.click(
+            fn=perform_inline_login,
+            inputs=[login_username, login_password],
+            outputs=[
+                login_username, 
+                login_password, 
+                login_submit, 
+                login_error, 
+                submit_button, 
+                submission_feedback_display, 
+                team_name_state,
+                username_state,  # NEW
+                token_state      # NEW
+            ]
+        )
+
+        # Removed gr.State(username) from the inputs list
+        submit_button.click(
+            fn=run_experiment,
+            inputs=[
+                model_type_state,
+                complexity_state,
+                feature_set_state,
+                data_size_state,
+                team_name_state,
+                last_submission_score_state,
+                last_rank_state,
+                submission_count_state,
+                first_submission_score_state,
+                best_score_state,
+                username_state,  # NEW: Session-based auth
+                token_state,     # NEW: Session-based auth
+                readiness_state, # Renamed to readiness_flag in function signature
+                was_preview_state, # Renamed to was_preview_prev in function signature
+                # kpi_meta_state removed from inputs - used only as output
+            ],
+            outputs=all_outputs,
+            show_progress="full",
+            js=nav_js("model-step", "Running experiment...", 500)
+        )
+
+        # Timer for polling initialization status
+        status_timer = gr.Timer(value=0.5, active=True)  # Poll every 0.5 seconds
+        
+        def update_init_status():
+            """
+            Poll initialization status and update UI elements.
+            Returns status HTML, banner visibility, submit button state, data size choices, and readiness_state.
+            """
+            status_html, ready = poll_init_status()
+            
+            # Update banner visibility - hide when ready
+            banner_visible = not ready
+            
+            # Update submit button
+            if ready:
+                submit_label = "5. 🔬 Build & Submit Model"
+                submit_interactive = True
+            else:
+                submit_label = "⏳ Waiting for data..."
+                submit_interactive = False
+            
+            # Get available data sizes based on init progress
+            available_sizes = get_available_data_sizes()
+            
+            # Stop timer once fully initialized
+            timer_active = not (ready and INIT_FLAGS.get("pre_samples_full", False))
+            
+            return (
+                status_html,
+                gr.update(visible=banner_visible),
+                gr.update(value=submit_label, interactive=submit_interactive),
+                gr.update(choices=available_sizes),
+                timer_active,
+                ready  # readiness_state
+            )
+        
+        status_timer.tick(
+            fn=update_init_status,
+            inputs=None,
+            outputs=[init_status_display, init_banner, submit_button, data_size_radio, status_timer, readiness_state]
+        )
+
+        # Handle session-based authentication on page load
+        def handle_load_with_session_auth(request: "gr.Request"):
+            """
+            Check for session token, auto-login if present, then load initial UI with stats.
+            
+            Concurrency Note: This function does NOT set per-user values in os.environ.
+            All authentication state is returned via gr.State objects (username_state,
+            token_state, team_name_state) to prevent cross-user data leakage.
+            """
+            success, username, token = _try_session_based_auth(request)
+            
+            if success and username and token:
+                _log(f"Session auth successful on load for {username}")
+                
+                # Get user stats and team from cache/leaderboard
+                stats = _compute_user_stats(username, token)
+                team_name = stats.get("team_name", "")
+                
+                # Concurrency Note: Do NOT set os.environ for per-user values.
+                # Return state via gr.State objects exclusively.
+                
+                # Hide login form since user is authenticated via session
+                # Return initial load results plus login form hidden
+                # Pass token explicitly for authenticated leaderboard fetch
+                initial_results = on_initial_load(username, token=token, team_name=team_name)
+                return initial_results + (
+                    gr.update(visible=False),  # login_username
+                    gr.update(visible=False),  # login_password  
+                    gr.update(visible=False),  # login_submit
+                    gr.update(visible=False),  # login_error (hide any messages)
+                    username,  # username_state
+                    token,     # token_state
+                    team_name, # team_name_state
+                )
+            else:
+                _log("No valid session on load, showing login form")
+                # No valid session, proceed with normal load (show login form)
+                # No token available, call without token
+                initial_results = on_initial_load(None, token=None, team_name="")
+                return initial_results + (
+                    gr.update(visible=True),   # login_username
+                    gr.update(visible=True),   # login_password
+                    gr.update(visible=True),   # login_submit
+                    gr.update(visible=False),  # login_error
+                    None,  # username_state
+                    None,  # token_state
+                    "",    # team_name_state
+                )
+        
+        demo.load(
+            fn=handle_load_with_session_auth,
+            inputs=None,  # Request is auto-injected
+            outputs=[
+                model_card_display,
+                team_leaderboard_display, 
+                individual_leaderboard_display, 
+                rank_message_display,
+                model_type_radio,
+                complexity_slider,
+                feature_set_checkbox,
+                data_size_radio,
+                login_username,
+                login_password,
+                login_submit,
+                login_error,
+                username_state,  # NEW
+                token_state,     # NEW
+                team_name_state, # NEW
+            ]
+        )
+
+    return demo
+
+# -------------------------------------------------------------------------
+# 4. Convenience Launcher
+# -------------------------------------------------------------------------
+
+def launch_model_building_game_es_final_app(height: int = 1200, share: bool = False, debug: bool = False) -> None:
+    """
+    Create and directly launch the Model Building Game app inline (e.g., in notebooks).
+    """
+    global playground, X_TRAIN_RAW, X_TEST_RAW, Y_TRAIN, Y_TEST
+    if playground is None:
+        try:
+            playground = Competition(MY_PLAYGROUND_ID)
+        except Exception as e:
+            print(f"WARNING: Could not connect to playground: {e}")
+            playground = None
+
+    if X_TRAIN_RAW is None:
+        X_TRAIN_RAW, X_TEST_RAW, Y_TRAIN, Y_TEST = load_and_prep_data()
+
+    demo = create_model_building_game_es_final_app()
+    port = int(os.environ.get("PORT", 8080))
+    demo.launch(share=share, inline=True, debug=debug, height=height, server_port=port)