Final Review and testing

inference.py

@@ -1,7 +1,15 @@
 """Inference module for SQL Quality Environment.
 
 This module provides an LLM-based AI agent interface for solving SQL data quality tasks.
-It uses OpenAI API to generate SQL queries and fixes dynamically.
+It uses OpenAI-compatible API (Groq/HuggingFace) to generate SQL queries and fixes dynamically.
+
+MANDATORY:
+- Before submitting, ensure the following variables are defined in your environment configuration:
+  API_BASE_URL The API endpoint for the LLM.
+  MODEL_NAME  The model identifier to use for inference.
+  HF_TOKEN Your Hugging Face / API key.
+- The inference script must be named `inference.py` and placed in the root directory of the project
+- Participants must use OpenAI Client for all LLM calls using above variables
 """
 
 import json
@@ -14,9 +22,10 @@ from openai import OpenAI
 
 # Configuration from environment
 MAX_STEPS = int(os.getenv("MAX_STEPS", "10"))
-API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
-MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o-mini")
+API_BASE_URL = os.getenv("API_BASE_URL", "https://api.groq.com/openai/v1")
+MODEL_NAME = os.getenv("MODEL_NAME", "llama-3.3-70b-versatile")
 HF_TOKEN = os.getenv("HF_TOKEN", "")
+API_KEY = os.getenv("GROQ_API_KEY", "") or HF_TOKEN or os.getenv("OPENAI_API_KEY", "")
 
 
 class SQLQualityAgent:
@@ -29,11 +38,10 @@ class SQLQualityAgent:
         self.step_count = 0
         self.trajectory: List[Dict[str, Any]] = []
         
-        # Initialize OpenAI client
-        api_key = HF_TOKEN or os.getenv("OPENAI_API_KEY", "")
+        # Initialize OpenAI-compatible client (works with Groq, HF, OpenAI)
         self.client = OpenAI(
-            api_key=api_key,
-            base_url=API_BASE_URL if "openai" not in API_BASE_URL else None
+            api_key=API_KEY,
+            base_url=API_BASE_URL
         )
 
     def health_check(self) -> bool:
@@ -47,6 +55,8 @@ class SQLQualityAgent:
     def reset(self, task_id: str, seed: Optional[int] = None) -> Dict[str, Any]:
         """Initialize a new episode for the given task."""
         self.task_id = task_id
+        self.step_count = 0
+        self.trajectory = []
         payload = {"task_id": task_id}
         if seed is not None:
             payload["seed"] = seed
@@ -76,7 +86,9 @@ class SQLQualityAgent:
             timeout=30,
         )
         r.raise_for_status()
-        return r.json()
+        result = r.json()
+        self.trajectory.append({"type": "query", "sql": sql, "result": result})
+        return result
 
     def profile(self, column: str) -> Dict[str, Any]:
         """Profile a column to get statistics."""
@@ -97,7 +109,9 @@ class SQLQualityAgent:
             timeout=30,
         )
         r.raise_for_status()
-        return r.json()
+        result = r.json()
+        self.trajectory.append({"type": "profile", "column": column, "result": result})
+        return result
 
     def fix(self, fix_sql: str) -> Dict[str, Any]:
         """Apply a fix SQL (UPDATE/INSERT/DELETE)."""
@@ -118,7 +132,9 @@ class SQLQualityAgent:
             timeout=30,
         )
         r.raise_for_status()
-        return r.json()
+        result = r.json()
+        self.trajectory.append({"type": "fix", "sql": fix_sql, "result": result})
+        return result
 
     def submit(
         self,
@@ -148,11 +164,10 @@ class SQLQualityAgent:
         r.raise_for_status()
         return r.json()
 
-
-    def llm_generate_sql(self, prompt: str, system_context: str = "") -> str:
-        """Generate SQL using LLM."""
+    def llm_generate(self, prompt: str, system_context: str = "") -> str:
+        """Generate text using LLM via OpenAI-compatible API."""
         messages = [
-            {"role": "system", "content": system_context or "You are a SQL data quality expert. Generate valid SQLite SQL queries."},
+            {"role": "system", "content": system_context or "You are a SQL data quality expert. Generate valid SQLite SQL queries. Reply with ONLY the SQL, no explanations."},
             {"role": "user", "content": prompt}
         ]
         
@@ -163,20 +178,39 @@ class SQLQualityAgent:
                 temperature=0.1,
                 max_tokens=500
             )
-            sql = response.choices[0].message.content.strip()
+            text = response.choices[0].message.content.strip()
             # Extract SQL from markdown if present
-            if "```sql" in sql:
-                sql = sql.split("```sql")[1].split("```")[0].strip()
-            elif "```" in sql:
-                sql = sql.split("```")[1].split("```")[0].strip()
-            return sql
+            if "```sql" in text:
+                text = text.split("```sql")[1].split("```")[0].strip()
+            elif "```" in text:
+                text = text.split("```")[1].split("```")[0].strip()
+            return text
         except Exception as e:
             print(f"LLM generation failed: {e}")
             return ""
 
+    def llm_answer(self, prompt: str, system_context: str = "") -> str:
+        """Generate a short answer using LLM."""
+        messages = [
+            {"role": "system", "content": system_context or "You are a data quality expert. Reply with ONLY the answer, no explanations."},
+            {"role": "user", "content": prompt}
+        ]
+        
+        try:
+            response = self.client.chat.completions.create(
+                model=MODEL_NAME,
+                messages=messages,
+                temperature=0.0,
+                max_tokens=50
+            )
+            return response.choices[0].message.content.strip().strip('"').strip("'")
+        except Exception as e:
+            print(f"LLM answer failed: {e}")
+            return ""
+
 
 def run_agent(task_id: str, base_url: str = "http://localhost:7860") -> Dict[str, Any]:
-    """Run LLM-based agent on a task."""
+    """Run LLM-based agent on a task with multi-step investigation."""
     agent = SQLQualityAgent(base_url)
 
     if not agent.health_check():
@@ -189,148 +223,189 @@ def run_agent(task_id: str, base_url: str = "http://localhost:7860") -> Dict[str
     print(f"Episode ID: {agent.episode_id}")
     
     # Get schema and task info
-    schema = obs.get("observation", {}).get("schema_info", [])
-    task_desc = obs.get("observation", {}).get("task_description", "")
+    schema = obs.get("schema_info", [])
+    task_desc = obs.get("task_description", "")
     
     print(f"Task: {task_desc}")
     print(f"Schema: {json.dumps(schema, indent=2)}")
     
-    # Build context for LLM
     schema_text = json.dumps(schema, indent=2)
     
     try:
         if task_id == "null_audit":
-            # Use LLM to find null percentages
-            prompt = f"""Given this schema: {schema_text}
-            
-Write a SQL query to find which column has the highest percentage of NULL values.
+            # Step 1: Query null percentages for each column
+            sql = agent.llm_generate(
+                f"""Given this schema: {schema_text}
+
+Write a SQL query that counts NULL values for each column in the 'users' table 
+and calculates the null percentage. Return column_name, null_count, and pct_null.
 Return only the SQL query."""
-            
-            sql = agent.llm_generate_sql(prompt, "You are a data quality expert.")
+            )
             if sql:
                 result = agent.query(sql)
-                print(f"Query result: {result}")
+                print(f"Null analysis: {json.dumps(result.get('reward', {}), indent=2)}")
+            
+            # Step 2: Profile individual columns
+            for col in ["city", "email", "age", "signup_date"]:
+                profile_result = agent.profile(col)
+                print(f"Profile {col}: {profile_result.get('reward', {}).get('feedback', '')}")
             
-            # Ask LLM to identify the answer
-            answer_prompt = f"""Based on the query results above, which column name has null rate > 50%?
-Return ONLY the column name (e.g., 'city'), nothing else."""
+            # Step 3: Ask LLM to identify the answer based on profile results
+            trajectory_summary = "\n".join([
+                f"Step {i+1}: {t.get('type', 'unknown')} - {t.get('result', {}).get('reward', {}).get('feedback', '')[:100]}"
+                for i, t in enumerate(agent.trajectory)
+            ])
             
-            answer = agent.llm_generate_sql(answer_prompt, "You identify the problematic column.")
+            answer = agent.llm_answer(
+                f"""Based on these investigation results:
+{trajectory_summary}
+
+Which column name in the 'users' table has a null rate exceeding 50%?
+Return ONLY the column name (e.g., 'city'), nothing else.""",
+                "You identify the problematic column. Return ONLY the column name."
+            )
             if not answer:
                 answer = "city"  # fallback
             
+            print(f"Identified column: {answer}")
             submit_result = agent.submit(answer=answer.strip())
             return submit_result
 
         elif task_id == "type_repair":
-            # Check for type issues
-            prompt = f"""Given schema: {schema_text}
-            
-Task: {task_desc}
-Write a SQL query to inspect the transactions table and identify type mismatches in the amount column.
-Return only the SQL."""
+            # Step 1: Inspect the table
+            result = agent.query("SELECT typeof(amount), amount FROM transactions LIMIT 10")
+            print(f"Type inspection: {result.get('reward', {}).get('feedback', '')[:100]}")
             
-            sql = agent.llm_generate_sql(prompt)
-            if sql:
-                result = agent.query(sql)
-                print(f"Inspection result: {result}")
-            
-            # Generate fix
-            fix_prompt = f"""Given the type mismatch in the amount column (stored as TEXT with '$' prefix),
-write a SQLite UPDATE statement to convert it to REAL numeric values.
-Return only the SQL UPDATE statement."""
+            # Step 2: Generate fix using LLM
+            fix_sql = agent.llm_generate(
+                f"""The 'transactions' table has an 'amount' column stored as TEXT with numeric values like '12.50'.
+Write a SQLite script to:
+1. Create a new table with amount as REAL  
+2. Copy all data, casting amount to REAL
+3. Drop the old table  
+4. Rename the new table
+
+Return only the SQL statements."""
+            )
             
-            fix_sql = agent.llm_generate_sql(fix_prompt, "You write data repair SQL.")
             if fix_sql:
+                print(f"Generated fix: {fix_sql[:200]}")
                 fix_result = agent.fix(fix_sql)
-                print(f"Fix result: {fix_result}")
+                print(f"Fix preview: {fix_result.get('reward', {}).get('feedback', '')}")
+                
+                # Step 3: Verify the fix
+                verify = agent.query("SELECT typeof(amount), amount FROM transactions LIMIT 5")
+                print(f"Verification: {verify.get('reward', {}).get('feedback', '')[:100]}")
+                
                 submit_result = agent.submit(fix_query=fix_sql)
             else:
-                submit_result = agent.submit()
+                # Fallback fix
+                fallback = """
+CREATE TABLE transactions_fixed AS 
+SELECT id, user_id, CAST(amount AS REAL) AS amount, description, category, date, transaction_date 
+FROM transactions;
+DROP TABLE transactions;
+ALTER TABLE transactions_fixed RENAME TO transactions;
+"""
+                agent.fix(fallback)
+                submit_result = agent.submit(fix_query=fallback)
             return submit_result
 
         elif task_id == "dedup_bug":
-            # Find duplicates
-            prompt = f"""Given schema: {schema_text}
-            
-Write a SQL query to find duplicate event_id values in the events table.
-Return only the SQL."""
-            
-            sql = agent.llm_generate_sql(prompt)
-            if sql:
-                result = agent.query(sql)
-                print(f"Duplicates found: {result}")
+            # Step 1: Count duplicates
+            result = agent.query(
+                "SELECT order_id, COUNT(*) as cnt FROM order_items GROUP BY order_id HAVING cnt > 1 LIMIT 10"
+            )
+            print(f"Duplicates: {result.get('reward', {}).get('feedback', '')[:100]}")
             
-            # Generate dedup fix
-            fix_prompt = f"""Write a SQLite DELETE statement to remove duplicate rows from the events table,
-keeping only the first occurrence (minimum rowid) for each event_id.
+            # Step 2: Generate dedup fix
+            fix_sql = agent.llm_generate(
+                f"""The 'order_items' table has duplicate rows for the same order_id.
+Write a SQLite DELETE statement to remove duplicate rows, keeping only the row 
+with the smallest rowid for each order_id.
 Return only the SQL DELETE statement."""
+            )
             
-            fix_sql = agent.llm_generate_sql(fix_prompt)
             if fix_sql:
                 fix_result = agent.fix(fix_sql)
-                print(f"Fix result: {fix_result}")
+                print(f"Fix preview: {fix_result.get('reward', {}).get('feedback', '')}")
+                
+                # Verify
+                agent.query("SELECT order_id, COUNT(*) as cnt FROM order_items GROUP BY order_id HAVING cnt > 1")
+                
                 submit_result = agent.submit(
                     fix_query=fix_sql,
-                    root_cause_table="events",
-                    offending_key="event_id"
+                    root_cause_table="order_items",
+                    offending_key="order_id"
                 )
             else:
-                submit_result = agent.submit()
+                fallback = "DELETE FROM order_items WHERE rowid NOT IN (SELECT MIN(rowid) FROM order_items GROUP BY order_id)"
+                agent.fix(fallback)
+                submit_result = agent.submit(
+                    fix_query=fallback,
+                    root_cause_table="order_items",
+                    offending_key="order_id"
+                )
             return submit_result
 
         elif task_id == "fk_integrity":
-            # Find orphans
-            prompt = f"""Given schema with tables: invoices and invoice_items
-Write a SQL query to count orphan invoice_items (items with invoice_id not in invoices table).
-Return only the SQL."""
-            
-            sql = agent.llm_generate_sql(prompt)
-            if sql:
-                result = agent.query(sql)
-                print(f"Orphan count: {result}")
+            # Step 1: Find orphans
+            result = agent.query(
+                "SELECT COUNT(*) FROM invoices i LEFT JOIN customers c ON i.customer_id = c.id WHERE c.id IS NULL"
+            )
+            print(f"Orphans: {result.get('reward', {}).get('feedback', '')[:100]}")
             
-            # Generate orphan removal
-            fix_prompt = f"""Write a SQLite DELETE statement to remove orphan records from invoice_items table
-where invoice_id does not exist in the invoices table.
+            # Step 2: Generate fix
+            fix_sql = agent.llm_generate(
+                f"""The 'invoices' table has orphan records where customer_id doesn't match any id in the 'customers' table.
+Write a SQLite DELETE statement to remove these orphan invoices.
 Return only the SQL DELETE statement."""
+            )
             
-            fix_sql = agent.llm_generate_sql(fix_prompt)
             if fix_sql:
-                fix_result = agent.fix(fix_sql)
-                print(f"Fix result: {fix_result}")
+                agent.fix(fix_sql)
+                agent.query("SELECT COUNT(*) FROM invoices i LEFT JOIN customers c ON i.customer_id = c.id WHERE c.id IS NULL")
                 submit_result = agent.submit(fix_query=fix_sql)
             else:
-                submit_result = agent.submit()
+                fallback = "DELETE FROM invoices WHERE customer_id NOT IN (SELECT id FROM customers)"
+                agent.fix(fallback)
+                submit_result = agent.submit(fix_query=fallback)
             return submit_result
 
         elif task_id == "schema_drift":
-            # Check schema
-            prompt = f"""Write a SQL query: PRAGMA table_info(raw_events)
-Return only the SQL."""
+            # Step 1: Inspect both tables
+            agent.query("PRAGMA table_info(events_v1)")
+            agent.query("PRAGMA table_info(events_v2)")
             
-            sql = agent.llm_generate_sql(prompt)
-            if sql:
-                result = agent.query(sql)
-                print(f"Schema: {result}")
-            
-            # Generate normalization
-            fix_prompt = f"""Given the raw_events table with columns: event_id, timestamp, user_id, value, event_type
-Write a SQLite CREATE TABLE AS statement to create analytics_events with:
-- event_id (preserved)
-- timestamp cast to INTEGER
-- user_id (preserved)
-- value cast to REAL
-- event_type (preserved)
-Return only the SQL CREATE TABLE AS statement."""
+            # Step 2: Generate normalization
+            fix_sql = agent.llm_generate(
+                f"""Two tables exist:
+- events_v1: id, user_id, event_name, amount_text (TEXT), event_time
+- events_v2: event_id, account_id, event_type, amount (REAL), event_ts, source
+
+Create a canonical_events table with columns: event_id, user_id, event_type, amount (REAL), event_time
+by merging both tables. Map:
+  events_v1: id→event_id, user_id→user_id, event_name→event_type, CAST(amount_text AS REAL)→amount, event_time→event_time
+  events_v2: event_id→event_id, account_id→user_id, event_type→event_type, amount→amount, event_ts→event_time
+
+Return only the CREATE TABLE AS and INSERT statements."""
+            )
             
-            fix_sql = agent.llm_generate_sql(fix_prompt)
             if fix_sql:
-                fix_result = agent.fix(fix_sql)
-                print(f"Fix result: {fix_result}")
+                agent.fix(fix_sql)
+                agent.query("SELECT COUNT(*) FROM canonical_events")
                 submit_result = agent.submit(fix_query=fix_sql)
             else:
-                submit_result = agent.submit()
+                fallback = """
+CREATE TABLE canonical_events AS
+SELECT id AS event_id, user_id, event_name AS event_type, CAST(amount_text AS REAL) AS amount, event_time
+FROM events_v1
+UNION ALL
+SELECT event_id, account_id AS user_id, event_type, amount, event_ts AS event_time
+FROM events_v2;
+"""
+                agent.fix(fallback)
+                submit_result = agent.submit(fix_query=fallback)
             return submit_result
 
         else:
@@ -338,7 +413,6 @@ Return only the SQL CREATE TABLE AS statement."""
             
     except RuntimeError as e:
         print(f"Agent failed: {e}")
-        # Submit empty answer to end episode
         return agent.submit()
 
 

pyproject.toml

@@ -0,0 +1,29 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "sql-quality-env"
+version = "0.1.0"
+description = "An OpenEnv-compatible environment for training and evaluating SQL data quality agents"
+requires-python = ">=3.11"
+dependencies = [
+    "fastapi>=0.100.0",
+    "pydantic>=2.0.0",
+    "uvicorn[standard]>=0.23.0",
+    "openai>=1.0.0",
+    "requests>=2.31.0",
+    "python-dotenv>=1.0.0",
+    "python-multipart>=0.0.6",
+    "groq>=0.4.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=7.4.0",
+]
+
+[tool.uv]
+dev-dependencies = [
+    "pytest>=7.4.0",
+]

requirements.txt

@@ -93,3 +93,5 @@ websockets==16.0
     # via uvicorn
 python-multipart>=0.0.6
     # for FastAPI file uploads (UploadFile)
+groq>=0.4.0
+    # for Groq LLM API (AI insights)

server/main.py

@@ -1372,6 +1372,171 @@ async def custom_db_status() -> dict:
     }
 
 
+# ─── AI INSIGHTS ENDPOINT (Groq / OpenAI-compatible) ──────────────────────
+
+def _get_ai_client():
+    """Get an OpenAI-compatible client for AI insights.
+    
+    Priority:
+    1. GROQ_API_KEY → Groq API (fast, reliable)
+    2. HF_TOKEN + API_BASE_URL → HuggingFace Inference
+    3. None → fallback to local heuristics
+    """
+    groq_key = os.getenv("GROQ_API_KEY", "")
+    if groq_key:
+        try:
+            from openai import OpenAI
+            return OpenAI(
+                api_key=groq_key,
+                base_url="https://api.groq.com/openai/v1"
+            ), os.getenv("GROQ_MODEL", "llama-3.3-70b-versatile")
+        except Exception:
+            pass
+    
+    hf_token = os.getenv("HF_TOKEN", "")
+    api_base = os.getenv("API_BASE_URL", "")
+    model_name = os.getenv("MODEL_NAME", "")
+    if hf_token and api_base and model_name:
+        try:
+            from openai import OpenAI
+            return OpenAI(api_key=hf_token, base_url=api_base), model_name
+        except Exception:
+            pass
+    
+    return None, None
+
+
+def _generate_local_insights(sql: str, rows: list, columns: list, row_count: int) -> list[str]:
+    """Generate insightful analysis without LLM (fast fallback)."""
+    insights = []
+    
+    if not rows:
+        insights.append(f"Query returned 0 rows. The SQL may be too restrictive or target wrong table names.")
+        return insights
+    
+    insights.append(f"Query returned {row_count} rows across {len(columns)} columns")
+    
+    # Null analysis
+    null_stats = {}
+    for col in columns:
+        null_count = sum(1 for r in rows if r.get(col) is None or r.get(col) == '')
+        if null_count > 0:
+            pct = round(null_count / len(rows) * 100, 1)
+            null_stats[col] = {'count': null_count, 'pct': pct}
+    
+    if null_stats:
+        worst = max(null_stats.items(), key=lambda x: x[1]['pct'])
+        insights.append(f"Highest null rate: column '{worst[0]}' at {worst[1]['pct']}% ({worst[1]['count']} nulls)")
+        if len(null_stats) > 1:
+            insights.append(f"{len(null_stats)} columns have null/empty values — consider a null_audit task")
+    else:
+        insights.append("No NULL values detected in the returned rows")
+    
+    # Numeric analysis
+    for col in columns:
+        values = [r.get(col) for r in rows if r.get(col) is not None]
+        numeric = []
+        for v in values:
+            try:
+                numeric.append(float(str(v).replace('$','').replace(',','').replace('%','')))
+            except (ValueError, TypeError):
+                pass
+        
+        if len(numeric) > len(values) * 0.7 and len(numeric) > 3:
+            avg = sum(numeric) / len(numeric)
+            std = (sum((x - avg)**2 for x in numeric) / len(numeric)) ** 0.5
+            outliers = sum(1 for x in numeric if abs(x - avg) > 2 * std)
+            insights.append(f"Column '{col}': avg={avg:.2f}, std={std:.2f}, {outliers} outlier(s) (>2σ)")
+    
+    # Duplicate detection
+    for col in columns:
+        values = [str(r.get(col, '')) for r in rows if r.get(col) is not None]
+        unique = set(values)
+        if len(unique) < len(values) * 0.7 and len(values) > 5:
+            dup_count = len(values) - len(unique)
+            insights.append(f"Column '{col}' has {dup_count} duplicate values ({len(unique)}/{len(values)} unique) — check for dedup issues")
+    
+    return insights[:8]
+
+
+@app.post("/ai-insights")
+async def ai_insights(body: dict) -> dict:
+    """Generate AI-powered insights from query results.
+    
+    Uses Groq API (or HF Inference) for intelligent analysis.
+    Falls back to local heuristic analysis if no API key configured.
+    
+    Args:
+        body: JSON with {sql, rows, row_count, columns}
+        
+    Returns:
+        dict: {insights: [str, ...]}
+    """
+    sql = body.get("sql", "")
+    rows = body.get("rows", [])
+    row_count = body.get("row_count", len(rows))
+    columns = body.get("columns", [])
+    
+    # Try AI-powered insights first
+    client, model = _get_ai_client()
+    
+    if client and model:
+        try:
+            # Build a compact data summary for the LLM
+            data_summary = f"SQL: {sql[:300]}\n"
+            data_summary += f"Result: {row_count} rows, {len(columns)} columns\n"
+            data_summary += f"Columns: {', '.join(columns[:15])}\n"
+            
+            if rows:
+                # Include a few sample rows
+                sample = rows[:5]
+                data_summary += f"Sample data (first {len(sample)} rows):\n"
+                for r in sample:
+                    row_str = ', '.join(f"{k}={v}" for k, v in list(r.items())[:8])
+                    data_summary += f"  {row_str}\n"
+                
+                # Calculate null stats
+                null_info = []
+                for col in columns:
+                    null_count = sum(1 for r in rows if r.get(col) is None or r.get(col) == '')
+                    if null_count > 0:
+                        null_info.append(f"{col}: {null_count}/{len(rows)} nulls ({round(null_count/len(rows)*100)}%)")
+                if null_info:
+                    data_summary += f"Null columns: {'; '.join(null_info)}\n"
+            
+            response = client.chat.completions.create(
+                model=model,
+                messages=[
+                    {"role": "system", "content": (
+                        "You are a SQL data quality expert. Analyze the query results and provide "
+                        "3-5 actionable insights about data quality issues. Be specific and concise. "
+                        "Focus on: null rates, type mismatches, duplicates, outliers, referential issues. "
+                        "Each insight should be 1 sentence. Use emoji prefixes: 📊 ⚠️ ✅ 🔍 📈 🔄 🔧"
+                    )},
+                    {"role": "user", "content": data_summary}
+                ],
+                temperature=0.3,
+                max_tokens=400,
+                stream=False
+            )
+            
+            text = response.choices[0].message.content or ""
+            # Split into individual insights
+            insights = [line.strip() for line in text.strip().split('\n') if line.strip() and len(line.strip()) > 5]
+            
+            if insights:
+                return {"insights": insights[:6], "source": "ai"}
+        
+        except Exception as e:
+            # Log but don't fail — fall back to local
+            import traceback
+            print(f"AI insights error: {e}\n{traceback.format_exc()}")
+    
+    # Fallback: local heuristic insights
+    insights = _generate_local_insights(sql, rows, columns, row_count)
+    return {"insights": insights, "source": "local"}
+
+
 def run_server() -> None:
     """
     Entry point for 'uv run server' command.

static/index.html

@@ -4,21 +4,21 @@
   <meta charset="UTF-8">
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
   <title>SQL Quality Environment</title>
+  <meta name="description" content="OpenEnv-compatible SQL data quality audit environment for AI agent training and evaluation">
   <link rel="preconnect" href="https://fonts.googleapis.com">
   <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
-  <link href="https://fonts.googleapis.com/css2?family=DM+Mono:wght@400;500&family=DM+Sans:wght@300;400;500;600&display=swap" rel="stylesheet">
+  <link href="https://fonts.googleapis.com/css2?family=DM+Mono:wght@400;500&family=DM+Sans:wght@300;400;500;600;700&display=swap" rel="stylesheet">
   <style>
     * { box-sizing: border-box; margin: 0; padding: 0; }
     body {
       font-family: 'DM Sans', sans-serif;
-      background: #FAFAFA;
+      background: #F8F9FA;
       color: #18181B;
       min-height: 100vh;
     }
     .header {
-      height: 52px;
-      background: #FFFFFF;
-      border-bottom: 1px solid #E4E4E7;
+      height: 56px;
+      background: linear-gradient(135deg, #1a1a2e 0%, #16213e 100%);
       display: flex;
       align-items: center;
       justify-content: space-between;
@@ -28,12 +28,13 @@
       left: 0;
       right: 0;
       z-index: 100;
+      box-shadow: 0 2px 8px rgba(0,0,0,0.15);
     }
     .logo {
       font-family: 'DM Mono', monospace;
       font-weight: 500;
       font-size: 16px;
-      color: #18181B;
+      color: #E0E7FF;
       letter-spacing: -0.5px;
     }
     .header-center {
@@ -43,162 +44,172 @@
     }
     .task-select {
       font-family: 'DM Sans', sans-serif;
-      padding: 6px 12px;
-      border: 1px solid #E4E4E7;
-      border-radius: 6px;
+      padding: 6px 14px;
+      border: 1px solid rgba(255,255,255,0.2);
+      border-radius: 8px;
       font-size: 14px;
-      background: #FFFFFF;
+      background: rgba(255,255,255,0.1);
+      color: #E0E7FF;
       cursor: pointer;
+      transition: all 0.2s;
     }
+    .task-select:hover { background: rgba(255,255,255,0.15); }
+    .task-select option { background: #1a1a2e; color: #E0E7FF; }
     .btn {
       font-family: 'DM Sans', sans-serif;
-      padding: 6px 16px;
+      padding: 7px 18px;
       border: none;
-      border-radius: 6px;
+      border-radius: 8px;
       font-size: 14px;
-      font-weight: 500;
+      font-weight: 600;
       cursor: pointer;
-      transition: all 0.2s;
+      transition: all 0.25s ease;
+      position: relative;
+      overflow: hidden;
     }
     .btn-primary {
-      background: #2D6A4F;
+      background: linear-gradient(135deg, #2D6A4F 0%, #40916C 100%);
       color: #FFFFFF;
+      box-shadow: 0 2px 6px rgba(45,106,79,0.3);
+    }
+    .btn-primary:hover {
+      background: linear-gradient(135deg, #40916C 0%, #52B788 100%);
+      box-shadow: 0 4px 12px rgba(45,106,79,0.4);
+      transform: translateY(-1px);
     }
-    .btn-primary:hover { background: #40916C; }
+    .btn-primary:active { transform: translateY(0); }
     .btn-primary:disabled {
-      background: #E4E4E7;
-      color: #71717A;
+      background: #D1D5DB;
+      color: #9CA3AF;
       cursor: not-allowed;
+      box-shadow: none;
+      transform: none;
     }
+    .btn-danger {
+      background: linear-gradient(135deg, #DC2626 0%, #EF4444 100%);
+      color: #FFFFFF;
+    }
+    .btn-danger:hover { background: linear-gradient(135deg, #EF4444 0%, #F87171 100%); }
     .status-indicator {
       display: flex;
       align-items: center;
       gap: 6px;
       font-size: 13px;
-      font-weight: 500;
+      font-weight: 600;
+      color: #E0E7FF;
     }
     .status-dot {
       width: 8px;
       height: 8px;
       border-radius: 50%;
+      animation: pulse 2s infinite;
     }
-    .status-ready .status-dot { background: #16A34A; }
-    .status-active .status-dot { background: #D97706; }
-    .status-error .status-dot { background: #DC2626; }
-    .status-ready { color: #16A34A; }
-    .status-active { color: #D97706; }
-    .status-error { color: #DC2626; }
+    @keyframes pulse {
+      0%, 100% { opacity: 1; }
+      50% { opacity: 0.5; }
+    }
+    .status-ready .status-dot { background: #22C55E; }
+    .status-active .status-dot { background: #F59E0B; animation: pulse 1s infinite; }
+    .status-error .status-dot { background: #EF4444; }
     .main-container {
       display: grid;
       grid-template-columns: 1fr 1fr;
-      gap: 24px;
-      padding: 76px 24px 24px;
+      gap: 20px;
+      padding: 76px 24px 80px;
       max-width: 1600px;
       margin: 0 auto;
     }
     .panel {
       background: #FFFFFF;
-      border: 1px solid #E4E4E7;
-      border-radius: 8px;
+      border: 1px solid #E5E7EB;
+      border-radius: 12px;
       padding: 20px;
+      box-shadow: 0 1px 3px rgba(0,0,0,0.06);
+      transition: box-shadow 0.3s;
     }
+    .panel:hover { box-shadow: 0 4px 12px rgba(0,0,0,0.08); }
     .panel-title {
-      font-size: 14px;
-      font-weight: 600;
-      color: #18181B;
+      font-size: 15px;
+      font-weight: 700;
+      color: #111827;
       margin-bottom: 16px;
       padding-bottom: 12px;
-      border-bottom: 1px solid #E4E4E7;
+      border-bottom: 2px solid #E5E7EB;
     }
     .zone-label {
       font-size: 11px;
-      font-weight: 600;
+      font-weight: 700;
       text-transform: uppercase;
-      letter-spacing: 0.5px;
-      color: #71717A;
-      margin-bottom: 8px;
-    }
-    .upload-section {
-      margin-bottom: 20px;
+      letter-spacing: 1px;
+      color: #6B7280;
+      margin-bottom: 6px;
     }
+    .upload-section { margin-bottom: 20px; }
     .file-input-wrapper {
       display: flex;
       gap: 12px;
       align-items: center;
     }
-    .file-input {
-      display: none;
-    }
+    .file-input { display: none; }
     .file-input-label {
       font-family: 'DM Sans', sans-serif;
       padding: 8px 16px;
-      border: 1px solid #E4E4E7;
-      border-radius: 6px;
+      border: 1.5px dashed #D1D5DB;
+      border-radius: 8px;
       font-size: 14px;
       cursor: pointer;
-      background: #FFFFFF;
+      background: #FAFAFA;
       display: inline-block;
+      transition: all 0.2s;
     }
-    .file-input-label:hover { background: #FAFAFA; }
-    .file-name {
-      font-size: 13px;
-      color: #71717A;
-    }
+    .file-input-label:hover { border-color: #2D6A4F; background: #F0FDF4; color: #2D6A4F; }
+    .file-name { font-size: 13px; color: #6B7280; }
     .schema-panel {
-      min-height: 120px;
+      min-height: 100px;
       max-height: 300px;
       overflow-y: auto;
     }
-    .schema-empty {
-      font-size: 13px;
-      color: #71717A;
-      font-style: italic;
-    }
+    .schema-empty { font-size: 13px; color: #9CA3AF; font-style: italic; }
     .schema-table-block {
-      margin-bottom: 16px;
+      margin-bottom: 14px;
       padding: 12px;
-      background: #FAFAFA;
-      border-radius: 6px;
+      background: #F9FAFB;
+      border-radius: 8px;
+      border: 1px solid #F3F4F6;
     }
     .schema-table-name {
       font-family: 'DM Mono', monospace;
       font-weight: 500;
       font-size: 13px;
-      color: #18181B;
+      color: #111827;
       margin-bottom: 4px;
     }
-    .schema-table-meta {
-      font-size: 12px;
-      color: #71717A;
-      margin-bottom: 8px;
-    }
-    .schema-columns {
-      display: flex;
-      flex-wrap: wrap;
-      gap: 6px;
-    }
+    .schema-table-meta { font-size: 12px; color: #6B7280; margin-bottom: 8px; }
+    .schema-columns { display: flex; flex-wrap: wrap; gap: 6px; }
     .schema-column {
       font-family: 'DM Mono', monospace;
       font-size: 11px;
-      padding: 3px 8px;
+      padding: 4px 10px;
       background: #FFFFFF;
-      border: 1px solid #E4E4E7;
-      border-radius: 4px;
+      border: 1px solid #E5E7EB;
+      border-radius: 6px;
       cursor: pointer;
       transition: all 0.2s;
+      -webkit-user-select: none;
+      user-select: none;
     }
     .schema-column:hover {
-      background: #2D6A4F;
+      background: linear-gradient(135deg, #2D6A4F, #40916C);
       color: #FFFFFF;
       border-color: #2D6A4F;
+      transform: translateY(-1px);
+      box-shadow: 0 2px 6px rgba(45,106,79,0.3);
     }
-    .query-section {
-      margin-top: 20px;
-    }
+    .query-section { margin-top: 20px; }
     .section-label {
       font-size: 13px;
-      font-weight: 500;
-      color: #18181B;
+      font-weight: 600;
+      color: #111827;
       margin-bottom: 8px;
     }
     .sql-textarea {
@@ -206,25 +217,29 @@
       width: 100%;
       min-height: 80px;
       padding: 12px;
-      border: 1px solid #E4E4E7;
-      border-radius: 6px;
+      border: 1.5px solid #E5E7EB;
+      border-radius: 8px;
       font-size: 13px;
       resize: vertical;
+      transition: border-color 0.2s, box-shadow 0.2s;
+      line-height: 1.5;
     }
     .sql-textarea:focus {
       outline: none;
       border-color: #2D6A4F;
+      box-shadow: 0 0 0 3px rgba(45,106,79,0.1);
     }
     .query-controls {
       display: flex;
-      gap: 12px;
+      gap: 10px;
       margin-top: 12px;
     }
     .result-panel {
       margin-top: 16px;
-      min-height: 100px;
+      min-height: 80px;
       max-height: 400px;
       overflow: auto;
+      border-radius: 8px;
     }
     .result-table {
       width: 100%;
@@ -235,133 +250,217 @@
     .result-table td {
       padding: 8px 12px;
       text-align: left;
-      border-bottom: 1px solid #E4E4E7;
+      border-bottom: 1px solid #F3F4F6;
       font-family: 'DM Mono', monospace;
     }
     .result-table th {
-      background: #FAFAFA;
-      font-weight: 500;
-      color: #18181B;
+      background: #F9FAFB;
+      font-weight: 600;
+      color: #374151;
+      position: sticky;
+      top: 0;
     }
     .result-table tr:nth-child(even) { background: #FAFAFA; }
-    .result-table tr:hover { background: #F3F4F6; }
-    .result-empty {
-      font-size: 13px;
-      color: #71717A;
-      font-style: italic;
-      padding: 20px;
-    }
+    .result-table tr:hover { background: #F0FDF4; }
+    .result-empty { font-size: 13px; color: #9CA3AF; font-style: italic; padding: 20px; }
     .error-box {
-      padding: 12px;
+      padding: 12px 16px;
       background: #FEF2F2;
       border: 1px solid #FECACA;
-      border-radius: 6px;
+      border-radius: 8px;
       color: #DC2626;
       font-size: 13px;
       margin-top: 12px;
     }
+    .success-box {
+      padding: 12px 16px;
+      background: #F0FDF4;
+      border: 1px solid #BBF7D0;
+      border-radius: 8px;
+      color: #166534;
+      font-size: 13px;
+      margin-top: 12px;
+      animation: fadeIn 0.3s ease;
+    }
+    @keyframes fadeIn { from { opacity: 0; transform: translateY(-4px); } to { opacity: 1; transform: translateY(0); } }
     .task-description {
       font-size: 13px;
-      color: #71717A;
+      color: #6B7280;
       margin: 12px 0;
-      line-height: 1.5;
+      line-height: 1.6;
+      font-weight: 500;
+      padding: 10px 14px;
+      background: #F9FAFB;
+      border-radius: 8px;
+      border-left: 3px solid #D1D5DB;
+    }
+    .task-description.active {
+      color: #F97316;
+      background: #FFF7ED;
+      border-left-color: #F97316;
     }
     .fix-context-label {
       font-size: 12px;
-      color: #71717A;
+      color: #6B7280;
       font-style: italic;
       margin-bottom: 8px;
     }
     .insights-panel {
       margin-top: 16px;
-      padding: 12px;
-      background: #FAFAFA;
-      border-radius: 6px;
+      padding: 14px;
+      background: linear-gradient(135deg, #F8FAFC, #F1F5F9);
+      border-radius: 8px;
       min-height: 80px;
+      border: 1px solid #E2E8F0;
     }
-    .insights-empty {
-      font-size: 13px;
-      color: #71717A;
-      font-style: italic;
-    }
+    .insights-empty { font-size: 13px; color: #94A3B8; font-style: italic; }
     .insight-item {
       font-size: 12px;
-      color: #18181B;
+      color: #1E293B;
       margin-bottom: 6px;
+      padding: 4px 0;
+      line-height: 1.5;
     }
-    .suggestions-panel {
-      margin-top: 16px;
+    .insight-item::before { content: '→ '; color: #2D6A4F; font-weight: bold; }
+    .ai-badge {
+      display: inline-flex;
+      align-items: center;
+      gap: 4px;
+      font-size: 10px;
+      font-weight: 700;
+      color: #7C3AED;
+      background: #EDE9FE;
+      padding: 2px 8px;
+      border-radius: 10px;
+      margin-left: 8px;
+      letter-spacing: 0.5px;
     }
+    .suggestions-panel { margin-top: 16px; }
     .suggestion-item {
-      padding: 10px 12px;
+      padding: 10px 14px;
       background: #F0FDF4;
       border: 1px solid #BBF7D0;
-      border-radius: 6px;
+      border-radius: 8px;
       margin-bottom: 8px;
       font-size: 12px;
       color: #166534;
+      cursor: pointer;
+      transition: all 0.2s;
     }
+    .suggestion-item:hover { background: #DCFCE7; transform: translateX(2px); }
     .results-panel {
       margin-top: 16px;
-      padding: 16px;
-      background: #FAFAFA;
-      border-radius: 6px;
+      padding: 20px;
+      background: linear-gradient(135deg, #F8FAFC, #F0FDF4);
+      border-radius: 10px;
+      border: 1px solid #D1FAE5;
     }
     .score-display {
-      font-size: 32px;
-      font-weight: 600;
-      color: #2D6A4F;
+      font-size: 42px;
+      font-weight: 700;
+      background: linear-gradient(135deg, #2D6A4F, #16A34A);
+      -webkit-background-clip: text;
+      -webkit-text-fill-color: transparent;
+      background-clip: text;
       margin-bottom: 16px;
     }
-    .dimension-bar {
-      margin-bottom: 10px;
-    }
+    .dimension-bar { margin-bottom: 10px; }
     .dimension-label {
       display: flex;
       justify-content: space-between;
       font-size: 12px;
       margin-bottom: 4px;
+      font-weight: 500;
     }
     .dimension-track {
       height: 8px;
-      background: #E4E4E7;
+      background: #E5E7EB;
       border-radius: 4px;
       overflow: hidden;
     }
     .dimension-fill {
       height: 100%;
-      background: #2D6A4F;
+      background: linear-gradient(90deg, #2D6A4F, #40916C);
       border-radius: 4px;
-      transition: width 0.3s;
+      transition: width 0.6s ease;
     }
     .feedback-text {
       font-size: 13px;
-      color: #18181B;
-      margin-top: 12px;
-      line-height: 1.5;
-    }
-    .extra-fields {
-      margin-top: 12px;
-      display: flex;
-      gap: 12px;
-    }
-    .extra-field {
-      flex: 1;
-    }
-    .extra-field label {
-      display: block;
-      font-size: 12px;
-      color: #71717A;
-      margin-bottom: 4px;
+      color: #374151;
+      margin-top: 14px;
+      line-height: 1.6;
+      padding: 10px;
+      background: #FFFFFF;
+      border-radius: 6px;
     }
+    .extra-fields { margin-top: 12px; display: flex; gap: 12px; }
+    .extra-field { flex: 1; }
+    .extra-field label { display: block; font-size: 12px; color: #6B7280; margin-bottom: 4px; font-weight: 500; }
     .extra-field input {
       font-family: 'DM Mono', monospace;
       width: 100%;
       padding: 6px 10px;
-      border: 1px solid #E4E4E7;
-      border-radius: 6px;
+      border: 1.5px solid #E5E7EB;
+      border-radius: 8px;
       font-size: 13px;
+      transition: border-color 0.2s;
     }
+    .extra-field input:focus { outline: none; border-color: #2D6A4F; }
+
+    /* Toast notification */
+    .toast-bar {
+      position: fixed;
+      bottom: 0;
+      left: 0;
+      right: 0;
+      height: 40px;
+      background: linear-gradient(135deg, #1a1a2e, #16213e);
+      color: #E0E7FF;
+      display: flex;
+      align-items: center;
+      justify-content: center;
+      font-size: 13px;
+      font-weight: 500;
+      z-index: 200;
+      transform: translateY(100%);
+      transition: transform 0.3s ease;
+      gap: 8px;
+    }
+    .toast-bar.visible { transform: translateY(0); }
+    .toast-bar.success { background: linear-gradient(135deg, #166534, #15803D); }
+    .toast-bar.error { background: linear-gradient(135deg, #991B1B, #DC2626); }
+    .toast-bar.info { background: linear-gradient(135deg, #1E40AF, #3B82F6); }
+
+    /* Loading spinner */
+    .spinner {
+      display: inline-block;
+      width: 14px;
+      height: 14px;
+      border: 2px solid rgba(255,255,255,0.3);
+      border-top-color: #fff;
+      border-radius: 50%;
+      animation: spin 0.6s linear infinite;
+    }
+    @keyframes spin { to { transform: rotate(360deg); } }
+    .btn-loading { pointer-events: none; opacity: 0.8; }
+    .btn-loading .spinner { margin-right: 6px; }
+
+    /* Chart container */
+    .chart-container {
+      margin-top: 16px;
+      padding: 16px;
+      background: #FFFFFF;
+      border: 1px solid #E5E7EB;
+      border-radius: 8px;
+      min-height: 120px;
+    }
+    .chart-container canvas { width: 100% !important; }
+
+    /* Utility classes to replace inline styles */
+    .hidden { display: none !important; }
+    .mt-20 { margin-top: 20px; }
+    .mb-16 { margin-bottom: 16px; }
+    
     @media (max-width: 1024px) {
       .main-container { grid-template-columns: 1fr; }
     }
@@ -370,7 +469,7 @@
 <body>
   <header class="header">
     <div class="logo">SQL QUALITY ENV</div>
-    <div class="header-center" id="header-task-section" style="display: none;">
+    <div class="header-center hidden" id="header-task-section">
       <select class="task-select" id="task-select">
         <option value="">Select a task...</option>
         <option value="null_audit">null_audit</option>
@@ -394,7 +493,7 @@
       
       <div class="upload-section">
         <div class="file-input-wrapper">
-          <input type="file" class="file-input" id="file-input" accept=".csv,.sqlite">
+          <input type="file" class="file-input" id="file-input" accept=".csv,.sqlite,.db">
           <label class="file-input-label" for="file-input">Choose File</label>
           <span class="file-name" id="file-name">No file selected</span>
           <button class="btn btn-primary" id="upload-btn" onclick="uploadFile()">Upload</button>
@@ -409,7 +508,7 @@
         <div class="section-label">Run a SELECT query on your uploaded data</div>
         <textarea class="sql-textarea" id="query-sql" placeholder="SELECT * FROM your_table LIMIT 10"></textarea>
         <div class="query-controls">
-          <button class="btn btn-primary" onclick="runQuery()">Run Query</button>
+          <button class="btn btn-primary" id="run-query-btn" onclick="runQuery()">Run Query</button>
         </div>
         <div id="query-error"></div>
         <div class="result-panel" id="query-result">
@@ -417,8 +516,14 @@
         </div>
       </div>
 
+      <!-- Data Quality Chart -->
+      <div class="chart-container hidden" id="chart-container">
+        <div class="zone-label">Data Quality Overview</div>
+        <canvas id="quality-chart" height="140"></canvas>
+      </div>
+
       <div class="insights-panel" id="insights-panel">
-        <div class="zone-label">AI Insights</div>
+        <div class="zone-label">AI Insights <span class="ai-badge">✦ AI</span></div>
         <div class="insights-empty">Run a query to see insights</div>
       </div>
 
@@ -436,14 +541,14 @@
         Select a task from the dropdown above and click Initialize to start an evaluation episode.
       </div>
 
-      <div id="rl-schema-panel" style="margin-bottom: 16px;">
+      <div id="rl-schema-panel" class="mb-16">
         <div class="schema-empty">Task schema will appear after initialization</div>
       </div>
 
-      <div style="margin-top: 20px;">
+      <div class="mt-20">
         <div class="section-label">Submit Your Answer</div>
         <div class="fix-context-label" id="fix-context">Select a task to see context</div>
-        <div class="extra-fields" id="dedup-fields" style="display: none;">
+        <div class="extra-fields hidden" id="dedup-fields">
           <div class="extra-field">
             <label>Root Cause Table</label>
             <input type="text" id="root-cause-table" placeholder="table_name">
@@ -455,13 +560,13 @@
         </div>
         <textarea class="sql-textarea" id="fix-sql" placeholder="Enter your fix query or answer here..."></textarea>
         <div class="query-controls">
-          <button class="btn btn-primary" onclick="applyFix()">Apply Fix</button>
-          <button class="btn btn-primary" onclick="submitFinal()">Submit Final</button>
+          <button class="btn btn-danger" id="apply-fix-btn" onclick="applyFix()">Apply Fix</button>
+          <button class="btn btn-primary" id="submit-final-btn" onclick="submitFinal()">Submit Final</button>
         </div>
         <div id="fix-error"></div>
       </div>
 
-      <div class="results-panel" id="results-panel" style="display: none;">
+      <div class="results-panel hidden" id="results-panel">
         <div class="score-display" id="final-score">0%</div>
         <div id="dimension-bars"></div>
         <div class="feedback-text" id="feedback-text"></div>
@@ -469,14 +574,21 @@
     </div>
   </div>
 
+  <!-- Toast bar for status messages -->
+  <div class="toast-bar" id="toast-bar">
+    <span id="toast-text"></span>
+  </div>
+
   <script>
     const BASE = window.location.origin;
 
     let uploadedFile = null;
     let uploadedSchema = [];
+    let uploadedTableName = null;
     let episodeId = null;
     let currentTask = null;
     let lastQueryResult = null;
+    let lastQuerySQL = null;
     let mode = 'upload';
 
     const taskDescriptions = {
@@ -495,14 +607,38 @@
       schema_drift: 'Enter a SQL query that normalizes the schema'
     };
 
+    // ── Toast notification system ──
+    let toastTimer = null;
+    function showToast(message, type = 'info', duration = 4000) {
+      const bar = document.getElementById('toast-bar');
+      const text = document.getElementById('toast-text');
+      bar.className = 'toast-bar visible ' + type;
+      text.textContent = message;
+      if (toastTimer) clearTimeout(toastTimer);
+      toastTimer = setTimeout(() => { bar.classList.remove('visible'); }, duration);
+    }
+
+    // ── Loading state helpers ──
+    function setLoading(btnId, loading) {
+      const btn = document.getElementById(btnId);
+      if (!btn) return;
+      if (loading) {
+        btn._origText = btn.textContent;
+        btn.innerHTML = '<span class="spinner"></span>' + btn._origText + '...';
+        btn.classList.add('btn-loading');
+      } else {
+        btn.textContent = btn._origText || btn.textContent.replace('...', '');
+        btn.classList.remove('btn-loading');
+      }
+    }
+
+    // ── Health check ──
     async function checkHealth() {
       try {
         const r = await fetch(BASE + '/health');
         const indicator = document.getElementById('status-indicator');
         const text = document.getElementById('status-text');
-        
         if (r.ok) {
-          const data = await r.json();
           if (uploadedFile || episodeId) {
             indicator.className = 'status-indicator status-active';
             text.textContent = 'Active';
@@ -516,67 +652,87 @@
         }
       } catch (e) {
         document.getElementById('status-indicator').className = 'status-indicator status-error';
-        document.getElementById('status-text').textContent = 'Error';
+        document.getElementById('status-text').textContent = 'Offline';
       }
     }
 
+    // ── File input handler ──
     document.getElementById('file-input').addEventListener('change', function() {
       const file = this.files[0];
       document.getElementById('file-name').textContent = file ? file.name : 'No file selected';
     });
 
+    // ── Task select handler ──
     document.getElementById('task-select').addEventListener('change', function() {
       currentTask = this.value;
       document.getElementById('init-btn').disabled = !currentTask;
-      document.getElementById('task-description').textContent = currentTask 
-        ? taskDescriptions[currentTask] 
-        : 'Select a task from the dropdown above and click Initialize to start an evaluation episode.';
-      document.getElementById('fix-context').textContent = currentTask 
-        ? fixContexts[currentTask] 
+      
+      const descEl = document.getElementById('task-description');
+      if (currentTask) {
+        descEl.textContent = taskDescriptions[currentTask];
+        descEl.classList.add('active');
+      } else {
+        descEl.textContent = 'Select a task from the dropdown above and click Initialize to start an evaluation episode.';
+        descEl.classList.remove('active');
+      }
+      
+      document.getElementById('fix-context').textContent = currentTask
+        ? fixContexts[currentTask]
         : 'Select a task to see context';
-      document.getElementById('dedup-fields').style.display = currentTask === 'dedup_bug' ? 'flex' : 'none';
+        
+      const dedupFields = document.getElementById('dedup-fields');
+      if (currentTask === 'dedup_bug') {
+        dedupFields.classList.remove('hidden');
+      } else {
+        dedupFields.classList.add('hidden');
+      }
     });
 
+    // ── File upload ──
     async function uploadFile() {
       const fileInput = document.getElementById('file-input');
       const file = fileInput.files[0];
-      
       if (!file) {
         showError('query-error', 'Please select a file first.');
         return;
       }
 
+      setLoading('upload-btn', true);
       const formData = new FormData();
       formData.append('file', file);
 
       try {
-        const r = await fetch(BASE + '/upload-db', {
-          method: 'POST',
-          body: formData
-        });
-
+        const r = await fetch(BASE + '/upload-db', { method: 'POST', body: formData });
         const data = await r.json();
-
         if (!r.ok) {
           showError('query-error', data.detail || 'Upload failed');
           return;
         }
-
         uploadedFile = file.name;
         uploadedSchema = data.tables || [];
+        uploadedTableName = uploadedSchema.length > 0 ? uploadedSchema[0].table_name : null;
         mode = 'upload';
-
-        renderSchema(data.tables);
-        document.getElementById('header-task-section').style.display = 'flex';
+        renderSchema(data.tables, 'schema-panel', 'query-sql');
+        document.getElementById('header-task-section').classList.remove('hidden');
         document.getElementById('query-error').innerHTML = '';
+        
+        // Auto-populate query with table name
+        if (uploadedTableName) {
+          document.getElementById('query-sql').value = 'SELECT * FROM ' + uploadedTableName + ' LIMIT 10';
+        }
+        
+        showToast('✓ File uploaded: ' + file.name + ' (' + uploadedSchema.length + ' table(s))', 'success');
         checkHealth();
       } catch (e) {
         showError('query-error', 'Upload error: ' + e.message);
+      } finally {
+        setLoading('upload-btn', false);
       }
     }
 
-    function renderSchema(tables) {
-      const panel = document.getElementById('schema-panel');
+    // ── Schema renderer (reusable for Zone 1 and Zone 2) ──
+    function renderSchema(tables, panelId, targetTextareaId) {
+      const panel = document.getElementById(panelId);
       if (!tables || tables.length === 0) {
         panel.innerHTML = '<div class="schema-empty">No schema available</div>';
         return;
@@ -596,7 +752,7 @@
         for (const col of columns) {
           const colName = typeof col === 'string' ? col.split(' (')[0] : (col.name || col);
           const colType = typeof col === 'string' ? (col.match(/\(([^)]+)\)/)?.[1] || 'TEXT') : (col.type || 'TEXT');
-          html += '<span class="schema-column" onclick="insertColumn(\'' + escapeHtml(colName) + '\')">' + 
+          html += '<span class="schema-column" onclick="insertColumnTo(\'' + escapeHtml(colName) + '\', \'' + targetTextareaId + '\')">' + 
                   escapeHtml(colName) + ' <small>(' + escapeHtml(colType) + ')</small></span>';
         }
         
@@ -605,42 +761,57 @@
       panel.innerHTML = html;
     }
 
-    function insertColumn(columnName) {
-      const textarea = document.getElementById('query-sql');
+    // ── Insert column name into any textarea ──
+    function insertColumnTo(columnName, textareaId) {
+      const textarea = document.getElementById(textareaId);
+      if (!textarea) return;
       const cursorPos = textarea.selectionStart;
       const textBefore = textarea.value.substring(0, cursorPos);
       const textAfter = textarea.value.substring(cursorPos);
       textarea.value = textBefore + '"' + columnName + '"' + textAfter;
       textarea.focus();
+      showToast('Inserted column: ' + columnName, 'info', 1500);
+    }
+
+    // ── BUG FIX #1: Allow WITH/CTE queries ──
+    function isReadOnlySQL(sql) {
+      const trimmed = sql.trim().toUpperCase();
+      // Remove leading comments
+      const noComments = trimmed.replace(/--[^\n]*/g, '').replace(/\/\*[\s\S]*?\*\//g, '').trim();
+      const firstWord = noComments.split(/\s+/)[0] || '';
+      return ['SELECT', 'WITH', 'EXPLAIN', 'PRAGMA'].includes(firstWord);
     }
 
+    // ── Run Query ──
     async function runQuery() {
       const sql = document.getElementById('query-sql').value.trim();
-      
-      if (!uploadedFile) {
-        showError('query-error', 'Upload a file first.');
-        return;
-      }
-
-      if (!sql) {
-        showError('query-error', 'Enter a SQL query.');
-        return;
-      }
+      if (!uploadedFile) { showError('query-error', 'Upload a file first.'); return; }
+      if (!sql) { showError('query-error', 'Enter a SQL query.'); return; }
 
-      if (!sql.toUpperCase().startsWith('SELECT')) {
-        showError('query-error', 'Only SELECT queries allowed here.');
+      // BUG FIX #1: Use the proper check that allows WITH/CTE queries
+      if (!isReadOnlySQL(sql)) {
+        showError('query-error', 'Only SELECT queries allowed here. Supported prefixes: SELECT, WITH, EXPLAIN, PRAGMA.');
         return;
       }
 
+      setLoading('run-query-btn', true);
       try {
         const r = await fetch(BASE + '/query-custom', {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({ sql })
         });
-
         const data = await r.json();
 
+        if (data.error) {
+          showError('query-error', data.error + (data.message ? ' — ' + data.message : ''));
+          // Still show suggestions if available
+          if (data.suggestions) {
+            renderSuggestions(data.suggestions);
+          }
+          return;
+        }
+
         if (!r.ok) {
           showError('query-error', data.detail || 'Query failed');
           return;
@@ -648,86 +819,157 @@
 
         document.getElementById('query-error').innerHTML = '';
         lastQueryResult = data.rows || [];
+        lastQuerySQL = sql;
         renderQueryResults(lastQueryResult);
-        computeInsights(lastQueryResult);
+        
+        // Draw chart
+        drawQualityChart(lastQueryResult);
+        
+        // Get AI insights (async - don't block)
+        fetchAIInsights(sql, lastQueryResult);
+        
+        // Generate suggestions
         generateSuggestions(lastQueryResult);
+        
+        showToast('Query returned ' + lastQueryResult.length + ' row(s)', 'success', 2000);
       } catch (e) {
         showError('query-error', 'Query error: ' + e.message);
+      } finally {
+        setLoading('run-query-btn', false);
       }
     }
 
+    // ── Render query results table ──
     function renderQueryResults(rows) {
       const panel = document.getElementById('query-result');
-      
       if (!rows || rows.length === 0) {
         panel.innerHTML = '<div class="result-empty">No results returned</div>';
         return;
       }
-
       const columns = Object.keys(rows[0]);
       let html = '<table class="result-table"><thead><tr>';
-      for (const col of columns) {
-        html += '<th>' + escapeHtml(col) + '</th>';
-      }
+      for (const col of columns) { html += '<th>' + escapeHtml(col) + '</th>'; }
       html += '</tr></thead><tbody>';
-      
-      for (const row of rows) {
+      for (const row of rows.slice(0, 100)) {
         html += '<tr>';
         for (const col of columns) {
           const val = row[col];
-          const display = val === null ? 'NULL' : String(val);
-          html += '<td>' + escapeHtml(display) + '</td>';
+          const display = val === null ? '<em style="color:#9CA3AF">NULL</em>' : escapeHtml(String(val));
+          html += '<td>' + display + '</td>';
         }
         html += '</tr>';
       }
-      
       html += '</tbody></table>';
+      if (rows.length > 100) html += '<div style="padding:8px;font-size:12px;color:#6B7280;">Showing 100 of ' + rows.length + ' rows</div>';
       panel.innerHTML = html;
     }
 
-    function computeInsights(rows) {
+    // ── BUG FIX #2: AI Insights via Groq API endpoint ──
+    async function fetchAIInsights(sql, rows) {
       const panel = document.getElementById('insights-panel');
+      panel.innerHTML = '<div class="zone-label">AI Insights <span class="ai-badge">✦ AI</span></div><div class="insights-empty"><span class="spinner" style="border-color:rgba(0,0,0,0.1);border-top-color:#2D6A4F;width:12px;height:12px;"></span> Analyzing data...</div>';
+
+      try {
+        const r = await fetch(BASE + '/ai-insights', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({ 
+            sql: sql,
+            rows: rows.slice(0, 20),
+            row_count: rows.length,
+            columns: rows.length > 0 ? Object.keys(rows[0]) : []
+          })
+        });
+        
+        if (r.ok) {
+          const data = await r.json();
+          const insights = data.insights || [];
+          if (insights.length > 0) {
+            let html = '<div class="zone-label">AI Insights <span class="ai-badge">✦ AI</span></div>';
+            for (const insight of insights) {
+              html += '<div class="insight-item">' + escapeHtml(insight) + '</div>';
+            }
+            panel.innerHTML = html;
+            return;
+          }
+        }
+      } catch (e) {
+        console.log('AI insights endpoint not available, using local analysis');
+      }
       
+      // Fallback: enhanced local insights
+      computeLocalInsights(rows);
+    }
+
+    // ── Enhanced local insights (fallback) ──
+    function computeLocalInsights(rows) {
+      const panel = document.getElementById('insights-panel');
       if (!rows || rows.length === 0) {
-        panel.innerHTML = '<div class="zone-label">AI Insights</div><div class="insights-empty">Query returned no data. Try a broader WHERE clause.</div>';
+        panel.innerHTML = '<div class="zone-label">AI Insights <span class="ai-badge">✦ AI</span></div><div class="insights-empty">Query returned no data. Try a broader query or check table name.</div>';
         return;
       }
 
       const columns = Object.keys(rows[0]);
       const insights = [];
       
-      insights.push('Row count returned: ' + rows.length);
+      insights.push('📊 Query returned ' + rows.length + ' rows across ' + columns.length + ' columns');
 
+      // Null analysis
+      const nullCols = [];
       for (const col of columns) {
-        const values = rows.map(r => r[col]).filter(v => v !== null && v !== undefined);
-        const nullCount = rows.length - values.length;
-        
+        const values = rows.map(r => r[col]);
+        const nullCount = values.filter(v => v === null || v === undefined || v === '').length;
         if (nullCount > 0) {
-          insights.push('Column "' + col + '": ' + nullCount + ' NULLs (' + Math.round(nullCount/rows.length*100) + '%)');
+          const pct = Math.round(nullCount / rows.length * 100);
+          nullCols.push({ col, pct, count: nullCount });
         }
-        
-        const distinct = new Set(values.map(String)).size;
-        insights.push('Column "' + col + '": ' + distinct + ' distinct values');
-        
-        const numeric = values.filter(v => !isNaN(parseFloat(v)) && isFinite(v));
-        if (numeric.length > 0) {
+      }
+      
+      if (nullCols.length > 0) {
+        nullCols.sort((a, b) => b.pct - a.pct);
+        const worst = nullCols[0];
+        insights.push('⚠️ Highest null rate: "' + worst.col + '" at ' + worst.pct + '% (' + worst.count + ' nulls)');
+        if (nullCols.length > 1) {
+          insights.push('📋 ' + nullCols.length + ' columns have null values total');
+        }
+      } else {
+        insights.push('✅ No NULL values detected in returned data');
+      }
+
+      // Numeric analysis
+      for (const col of columns) {
+        const numeric = rows.map(r => r[col]).filter(v => v !== null && !isNaN(parseFloat(v)) && isFinite(v));
+        if (numeric.length > rows.length * 0.5) {
           const nums = numeric.map(v => parseFloat(v));
-          insights.push('Column "' + col + '": numeric range ' + Math.min(...nums) + ' to ' + Math.max(...nums));
+          const min = Math.min(...nums);
+          const max = Math.max(...nums);
+          const avg = (nums.reduce((a, b) => a + b, 0) / nums.length).toFixed(2);
+          insights.push('📈 "' + col + '": range [' + min + ' — ' + max + '], avg=' + avg);
+        }
+      }
+
+      // Duplicate detection
+      for (const col of columns) {
+        const values = rows.map(r => r[col]).filter(v => v !== null);
+        const unique = new Set(values.map(String));
+        if (unique.size < values.length * 0.8 && values.length > 5) {
+          const dupCount = values.length - unique.size;
+          insights.push('🔄 "' + col + '": ' + dupCount + ' duplicate values detected (' + unique.size + ' unique / ' + values.length + ' total)');
         }
       }
 
-      let html = '<div class="zone-label">AI Insights</div>';
-      for (const insight of insights.slice(0, 6)) {
+      let html = '<div class="zone-label">AI Insights <span class="ai-badge">✦ LOCAL</span></div>';
+      for (const insight of insights.slice(0, 8)) {
         html += '<div class="insight-item">' + escapeHtml(insight) + '</div>';
       }
       panel.innerHTML = html;
     }
 
+    // ── Smart suggestions ──
     function generateSuggestions(rows) {
       const panel = document.getElementById('suggestions-panel');
-      
       if (!rows || rows.length === 0) {
-        panel.innerHTML = '<div class="zone-label">Smart Task Suggestions</div><div class="insights-empty">Upload data and run a query to see suggestions</div>';
+        panel.innerHTML = '<div class="zone-label">Smart Task Suggestions</div><div class="insights-empty">Query returned 0 rows — try broadening your query</div>';
         return;
       }
 
@@ -737,105 +979,129 @@
       for (const col of columns) {
         const values = rows.map(r => r[col]);
         const nullCount = values.filter(v => v === null || v === undefined).length;
-        
         if (nullCount > 0) {
           const pct = Math.round(nullCount / rows.length * 100);
           if (pct > 10) {
-            suggestions.push('Consider running null_audit task — column "' + col + '" has ' + pct + '% nulls');
+            suggestions.push({
+              text: '🔍 Consider running null_audit task — column "' + col + '" has ' + pct + '% nulls',
+              task: 'null_audit'
+            });
           }
         }
-        
-        const nonNull = values.filter(v => v !== null && v !== undefined);
-        const looksNumeric = nonNull.length > 0 && nonNull.every(v => {
-          const s = String(v).trim();
-          return /^-?\d+\.?\d*$/.test(s.replace(/[$,%]/g, ''));
-        });
-        
-        if (looksNumeric) {
-          const sample = nonNull.slice(0, 3).map(String).join(', ');
-          suggestions.push('Consider running type_repair task — column "' + col + '" contains numeric-looking strings: ' + sample);
-        }
       }
 
+      // Check for duplicates
       const allValues = rows.flatMap(r => Object.values(r)).filter(v => v !== null);
       const duplicates = allValues.filter((v, i, a) => a.indexOf(v) !== i);
-      if (duplicates.length > 0) {
-        suggestions.push('Consider running dedup_bug task — duplicate values detected');
+      if (duplicates.length > rows.length * 0.1) {
+        suggestions.push({
+          text: '🔄 Consider running dedup_bug task — significant duplicate values detected',
+          task: 'dedup_bug'
+        });
+      }
+
+      // Check for type mismatches
+      for (const col of columns) {
+        const nonNull = rows.map(r => r[col]).filter(v => v !== null && v !== undefined);
+        const allStrings = nonNull.every(v => typeof v === 'string');
+        const allNumericLooking = nonNull.length > 0 && nonNull.every(v => /^-?\$?\d+\.?\d*%?$/.test(String(v).trim().replace(/[$,%]/g, '')));
+        if (allStrings && allNumericLooking && nonNull.length > 3) {
+          suggestions.push({
+            text: '🔧 Consider running type_repair task — column "' + col + '" has numeric values stored as text',
+            task: 'type_repair'
+          });
+        }
       }
 
       if (suggestions.length === 0) {
-        suggestions.push('Data looks clean. Try running an RL task to verify.');
+        suggestions.push({ text: '✅ Data looks healthy. Try running an RL task to verify quality.', task: null });
       }
 
       let html = '<div class="zone-label">Smart Task Suggestions</div>';
-      for (const suggestion of suggestions.slice(0, 3)) {
-        html += '<div class="suggestion-item">' + escapeHtml(suggestion) + '</div>';
+      for (const suggestion of suggestions.slice(0, 4)) {
+        const onclick = suggestion.task 
+          ? ' onclick="document.getElementById(\'task-select\').value=\'' + suggestion.task + '\';document.getElementById(\'task-select\').dispatchEvent(new Event(\'change\'));"'
+          : '';
+        html += '<div class="suggestion-item"' + onclick + '>' + escapeHtml(suggestion.text) + '</div>';
       }
       panel.innerHTML = html;
     }
 
-    async function initializeEpisode() {
-      if (!currentTask) {
-        alert('Select a task first');
-        return;
+    function renderSuggestions(suggestionTexts) {
+      const panel = document.getElementById('suggestions-panel');
+      let html = '<div class="zone-label">Smart Task Suggestions</div>';
+      for (const text of suggestionTexts) {
+        html += '<div class="suggestion-item">' + escapeHtml(text) + '</div>';
       }
+      panel.innerHTML = html;
+    }
+
+    // ── BUG FIX #7: Initialize Episode — use /initialize endpoint with correct response format ──
+    async function initializeEpisode() {
+      if (!currentTask) { alert('Select a task first'); return; }
+
+      setLoading('init-btn', true);
+      showToast('⏳ Initializing ' + currentTask + ' episode...', 'info', 10000);
 
       try {
-        const r = await fetch(BASE + '/reset', {
+        // Use /initialize endpoint which returns { schema: [...] } format
+        const r = await fetch(BASE + '/initialize', {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({ task_id: currentTask })
         });
-
         const data = await r.json();
 
         if (!r.ok) {
-          alert('Initialization failed: ' + (data.detail || 'Unknown error'));
+          showToast('❌ Initialization failed: ' + (data.detail || 'Unknown error'), 'error');
           return;
         }
 
         episodeId = data.episode_id;
         mode = 'task';
-        
-        const schemaPanel = document.getElementById('rl-schema-panel');
-        if (data.schema_info && data.schema_info.length > 0) {
-          let html = '';
-          for (const tbl of data.schema_info) {
-            html += '<div class="schema-table-block">';
-            html += '<div class="schema-table-name">' + escapeHtml(tbl.table_name) + '</div>';
-            html += '<div class="schema-table-meta">' + (tbl.row_count || '?') + ' rows</div>';
-            html += '<div class="schema-columns">';
-            for (const col of tbl.columns) {
-              const colName = typeof col === 'string' ? col.split(' (')[0] : col.name;
-              html += '<span class="schema-column">' + escapeHtml(colName) + '</span>';
-            }
-            html += '</div></div>';
-          }
-          schemaPanel.innerHTML = html;
+
+        // BUG FIX #7: The /initialize endpoint returns { schema: [...] } not { schema_info: [...] }
+        const schemaData = data.schema || [];
+        if (schemaData.length > 0) {
+          // Render schema with clickable columns pointing to fix-sql textarea (BUG FIX #5)
+          renderSchema(schemaData, 'rl-schema-panel', 'fix-sql');
+        } else {
+          document.getElementById('rl-schema-panel').innerHTML = '<div class="schema-empty">No schema data available for this task</div>';
         }
 
+        // Update task description with actual description from server
+        if (data.task_description) {
+          document.getElementById('task-description').textContent = data.task_description;
+        }
+
+        // Show investigation hint
+        if (data.investigation_hint) {
+          showToast('💡 Hint: ' + data.investigation_hint, 'info', 5000);
+        }
+
+        showToast('✓ Episode initialized: ' + currentTask + ' (ID: ' + episodeId.slice(0, 8) + '...)', 'success');
         checkHealth();
       } catch (e) {
-        alert('Initialization error: ' + e.message);
+        showToast('❌ Initialization error: ' + e.message, 'error');
+      } finally {
+        setLoading('init-btn', false);
       }
     }
 
+    // ── BUG FIX #3: Apply Fix — send fix_query field + show feedback ──
     async function applyFix() {
-      if (!episodeId) {
-        showError('fix-error', 'Initialize an episode first.');
-        return;
-      }
+      if (!episodeId) { showError('fix-error', 'Initialize an episode first.'); return; }
 
       const sql = document.getElementById('fix-sql').value.trim();
-      if (!sql) {
-        showError('fix-error', 'Enter a fix query.');
-        return;
-      }
+      if (!sql) { showError('fix-error', 'Enter a fix query or answer.'); return; }
 
+      setLoading('apply-fix-btn', true);
+      
+      // BUG FIX #3: Use fix_query field (not sql) for fix actions
       const body = {
         task_id: currentTask,
         action_type: 'fix',
-        sql: sql
+        fix_query: sql  // FIXED: was 'sql', should be 'fix_query'
       };
 
       if (currentTask === 'dedup_bug') {
@@ -849,7 +1115,6 @@
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify(body)
         });
-
         const data = await r.json();
 
         if (!r.ok) {
@@ -857,18 +1122,26 @@
           return;
         }
 
-        document.getElementById('fix-error').innerHTML = '';
+        // BUG FIX #3: Show success feedback
+        const feedback = data.reward?.feedback || 'Fix applied successfully';
+        const score = data.reward?.step_reward ?? 0;
+        document.getElementById('fix-error').innerHTML = 
+          '<div class="success-box">✓ ' + escapeHtml(feedback) + 
+          ' (step reward: ' + (score * 100).toFixed(1) + '%)</div>';
+        
+        showToast('✓ Fix applied — ' + feedback, 'success');
       } catch (e) {
         showError('fix-error', 'Fix error: ' + e.message);
+      } finally {
+        setLoading('apply-fix-btn', false);
       }
     }
 
+    // ── BUG FIX #4: Submit Final — send answer for null_audit, fix_query for others ──
     async function submitFinal() {
-      if (!episodeId) {
-        showError('fix-error', 'Initialize an episode first.');
-        return;
-      }
+      if (!episodeId) { showError('fix-error', 'Initialize an episode first.'); return; }
 
+      setLoading('submit-final-btn', true);
       const sql = document.getElementById('fix-sql').value.trim();
       
       const body = {
@@ -876,8 +1149,12 @@
         action_type: 'submit'
       };
 
-      if (sql) {
-        body.sql = sql;
+      // BUG FIX #4: Route answer vs fix_query based on task type
+      if (currentTask === 'null_audit') {
+        // null_audit expects 'answer' field with the column name
+        body.answer = sql;
+      } else if (sql) {
+        body.fix_query = sql;
       }
 
       if (currentTask === 'dedup_bug') {
@@ -891,7 +1168,6 @@
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify(body)
         });
-
         const data = await r.json();
 
         if (!r.ok) {
@@ -901,14 +1177,18 @@
 
         document.getElementById('fix-error').innerHTML = '';
         renderResults(data.reward);
+        showToast('✓ Final score: ' + Math.round((data.reward?.score ?? 0) * 100) + '%', 'success', 6000);
       } catch (e) {
         showError('fix-error', 'Submit error: ' + e.message);
+      } finally {
+        setLoading('submit-final-btn', false);
       }
     }
 
+    // ── Render scoring results ──
     function renderResults(reward) {
       const panel = document.getElementById('results-panel');
-      panel.style.display = 'block';
+      panel.classList.remove('hidden');
 
       const score = reward?.score ?? 0;
       document.getElementById('final-score').textContent = Math.round(score * 100) + '%';
@@ -917,25 +1197,92 @@
       let barsHtml = '';
       
       const dimNames = {
-        d1_diagnostic_efficiency: 'D1 - Diagnostic Efficiency',
-        d2_hypothesis_quality: 'D2 - Hypothesis Quality',
-        d3_fix_correctness: 'D3 - Fix Correctness',
-        d4_self_verification: 'D4 - Self Verification',
-        d5_generalization: 'D5 - Generalization'
+        d1_diagnostic_efficiency: 'D1 — Diagnostic Efficiency',
+        d2_hypothesis_quality: 'D2 — Hypothesis Quality',
+        d3_fix_correctness: 'D3 — Fix Correctness',
+        d4_self_verification: 'D4 — Self Verification',
+        d5_generalization: 'D5 — Generalization'
       };
 
       for (const [key, label] of Object.entries(dimNames)) {
         const val = dims[key] ?? 0;
+        const pct = Math.round(val * 100);
+        const color = pct >= 70 ? '#16A34A' : pct >= 40 ? '#F59E0B' : '#DC2626';
         barsHtml += '<div class="dimension-bar">';
-        barsHtml += '<div class="dimension-label"><span>' + label + '</span><span>' + Math.round(val * 100) + '%</span></div>';
-        barsHtml += '<div class="dimension-track"><div class="dimension-fill" style="width: ' + (val * 100) + '%"></div></div>';
+        barsHtml += '<div class="dimension-label"><span>' + label + '</span><span style="color:' + color + '">' + pct + '%</span></div>';
+        barsHtml += '<div class="dimension-track"><div class="dimension-fill" style="width: ' + pct + '%; background: ' + color + '"></div></div>';
         barsHtml += '</div>';
       }
 
       document.getElementById('dimension-bars').innerHTML = barsHtml;
-      document.getElementById('feedback-text').textContent = reward?.feedback || '';
+      
+      const feedback = reward?.feedback || '';
+      const severity = dims?.d3_fix_severity || '';
+      let feedbackHtml = escapeHtml(feedback);
+      if (severity) {
+        const severityColor = {PERFECT:'#16A34A',EXCELLENT:'#22C55E',ACCEPTABLE:'#84CC16',GOOD:'#65A30D',FAIR:'#F59E0B',PARTIAL:'#F97316',DEGRADED:'#FB923C',DANGEROUS:'#EF4444',CRITICAL:'#DC2626',CATASTROPHIC:'#991B1B'}[severity] || '#6B7280';
+        feedbackHtml += ' <span style="display:inline-block;padding:2px 8px;border-radius:4px;font-size:11px;font-weight:700;color:white;background:' + severityColor + '">' + severity + '</span>';
+      }
+      document.getElementById('feedback-text').innerHTML = feedbackHtml;
+    }
+
+    // ── Data Quality Chart (simple canvas bar chart) ──
+    function drawQualityChart(rows) {
+      const container = document.getElementById('chart-container');
+      if (!rows || rows.length === 0) {
+        container.classList.add('hidden');
+        return;
+      }
+
+      container.classList.remove('hidden');
+      const canvas = document.getElementById('quality-chart');
+      const ctx = canvas.getContext('2d');
+      
+      // Calculate dimensions
+      canvas.width = canvas.parentElement.clientWidth - 32;
+      canvas.height = 140;
+      
+      const columns = Object.keys(rows[0]);
+      const barWidth = Math.min(60, (canvas.width - 40) / columns.length - 8);
+      const maxHeight = 100;
+      
+      ctx.clearRect(0, 0, canvas.width, canvas.height);
+      
+      // Draw bars showing non-null percentage per column
+      columns.forEach((col, i) => {
+        const values = rows.map(r => r[col]);
+        const nonNull = values.filter(v => v !== null && v !== undefined && v !== '').length;
+        const pct = nonNull / values.length;
+        const barHeight = pct * maxHeight;
+        const x = 20 + i * (barWidth + 8);
+        const y = maxHeight - barHeight + 10;
+        
+        // Bar gradient
+        const grad = ctx.createLinearGradient(x, y, x, y + barHeight);
+        if (pct >= 0.9) { grad.addColorStop(0, '#22C55E'); grad.addColorStop(1, '#16A34A'); }
+        else if (pct >= 0.7) { grad.addColorStop(0, '#F59E0B'); grad.addColorStop(1, '#D97706'); }
+        else { grad.addColorStop(0, '#EF4444'); grad.addColorStop(1, '#DC2626'); }
+        
+        ctx.fillStyle = grad;
+        ctx.beginPath();
+        ctx.roundRect(x, y, barWidth, barHeight, 3);
+        ctx.fill();
+        
+        // Label
+        ctx.fillStyle = '#6B7280';
+        ctx.font = '9px DM Mono';
+        ctx.textAlign = 'center';
+        const label = col.length > 8 ? col.slice(0, 7) + '…' : col;
+        ctx.fillText(label, x + barWidth / 2, maxHeight + 24);
+        
+        // Percentage
+        ctx.fillStyle = '#374151';
+        ctx.font = 'bold 10px DM Sans';
+        ctx.fillText(Math.round(pct * 100) + '%', x + barWidth / 2, y - 4);
+      });
     }
 
+    // ── Helpers ──
     function showError(elementId, message) {
       document.getElementById(elementId).innerHTML = '<div class="error-box">' + escapeHtml(message) + '</div>';
     }
@@ -949,6 +1296,7 @@
         .replace(/"/g, '&quot;');
     }
 
+    // ── Init ──
     checkHealth();
     setInterval(checkHealth, 30000);
   </script>