feat: add primary descriptive headline and ensure entirely emoji-free UI

server/app.py

@@ -39,11 +39,17 @@ if ENABLE_WEB_INTERFACE:
             import gradio as gr
             with gr.Blocks() as blocks:
                 gr.Markdown("## Agent Learning Dashboard")
+                gr.Markdown("**This system simulates how RLHF agents learn from human feedback in real time.**")
                 gr.Markdown(
                     "This dashboard transforms the basic interface into an intelligent view of the RLHF agent's decision-making process. "
                     "You can observe reward signals, evaluation rationale, and training progression."
                 )
 
+                with gr.Row():
+                    best_reward_disp = gr.Markdown("### Best Reward: --")
+                    reward_delta_disp = gr.Markdown("### Recent Delta: --")
+                    confidence_disp = gr.Markdown("### Confidence: --")
+
                 with gr.Row():
                     with gr.Column(scale=2):
                         reward_plot = gr.LinePlot(
@@ -61,7 +67,9 @@ if ENABLE_WEB_INTERFACE:
                             improvement_tip = gr.Textbox(label="Agent Suggestion", lines=2)
                             
                     with gr.Column(scale=1):
-                        refresh_btn = gr.Button("Sync Agent State", variant="primary")
+                        with gr.Row():
+                            refresh_btn = gr.Button("Sync Agent State", variant="primary")
+                            demo_btn = gr.Button("Run Guided Demo", variant="secondary")
                         
                         agent_thinking = gr.Markdown(
                             "### Agent Process:\n"
@@ -86,7 +94,7 @@ if ENABLE_WEB_INTERFACE:
                     # Always ensure graph shows at least one point
                     if not data:
                         df = pd.DataFrame({"Step": [0], "Reward": [0.0]})
-                        return df, "Awaiting first agent action...", "Waiting...", "Dataset: <b>Pending</b>", "### Episode Summary\n_No steps yet._"
+                        return df, "Awaiting first agent action...", "Waiting...", "### Agent Process\n_Waiting for agent actions..._", "Dataset: <b>Pending</b>", "### Episode Summary\n_No steps yet._", "### Best Reward: --", "### Recent Delta: --", "### Confidence: --"
                         
                     df = pd.DataFrame(data)
                     latest_reward = data[-1]["Reward"]
@@ -126,14 +134,78 @@ if ENABLE_WEB_INTERFACE:
                         f"- **Improvement:** {improvement:+.1f}%\n"
                         f"- **Steps:** {latest_step}"
                     )
+                    # Dynamic Agent Thinking Engine
+                    task_type = getattr(last_log.observation, "task_type", "unknown") if hasattr(last_log, "observation") else "unknown"
                     
-                    return df, exp, tip, f"Dataset: <b>{dataset_str.upper()}</b>", summary
+                    thinking = f"### Agent Process (Step {latest_step}):\n"
+                    thinking += f"- Received `{task_type}` observation\n"
+                    if task_type == "pairwise":
+                        thinking += "- Compared Response A and B against Gold Standard\n"
+                    elif task_type == "likert":
+                        thinking += "- Evaluated response on 4 heuristic axes (Helpfulness, Honesty, etc)\n"
+                    elif task_type == "consistency":
+                        thinking += "- Checked consistency rankings for transitivity faults\n"
+                    else:
+                        thinking += "- Parsing standard input features\n"
+                        
+                    if latest_reward > 0.8:
+                        thinking += "- Decision matched gold labels almost perfectly\n"
+                        thinking += "- Issuing high positive reinforcement"
+                    elif latest_reward > 0.5:
+                        thinking += "- Decision showed partial alignment\n"
+                        thinking += "- Issuing moderate reinforcement"
+                    else:
+                        thinking += "- Decision strongly contradicted gold labels\n"
+                        thinking += "- Issuing negative reinforcement penalty"
+                    
+                    # KPI Visualizations
+                    best_reward = max([d["Reward"] for d in data])
+                    if len(data) > 1:
+                        delta = latest_reward - data[-2]["Reward"]
+                        delta_str = f"+{delta:.2f}" if delta >= 0 else f"{delta:.2f}"
+                    else:
+                        delta_str = "--"
 
+                    conf = 0.8
+                    if hasattr(last_log, "action") and last_log.action is not None:
+                        if hasattr(last_log.action, "confidence"):
+                            conf = last_log.action.confidence
+                        elif isinstance(last_log.action, dict) and "confidence" in last_log.action:
+                            conf = last_log.action["confidence"]
+                    conf_str = f"{int(conf * 100)}%"
+                    
+                    return df, exp, tip, thinking, f"Dataset: <b>{dataset_str.upper()}</b>", summary, f"### Best Reward: {best_reward:.2f}", f"### Recent Delta: {delta_str}", f"### Confidence: {conf_str}"
+
+                # Manual safe refresh mapping
+                # Manual safe refresh mapping
                 # Manual safe refresh mapping
                 refresh_btn.click(
                     fn=update_dashboard,
                     inputs=None,
-                    outputs=[reward_plot, reward_explanation, improvement_tip, dataset_vis, session_summary]
+                    outputs=[reward_plot, reward_explanation, improvement_tip, agent_thinking, dataset_vis, session_summary, best_reward_disp, reward_delta_disp, confidence_disp]
+                )
+
+                def run_demo_mode():
+                    import time
+                    import pandas as pd
+                    # Step 1
+                    df1 = pd.DataFrame([{"Step": 1, "Reward": 0.2}])
+                    yield df1, "Poor response, lacks relevance", "Focus on correctness", "### Agent Process (Demo):\n- Parsing standard input features\n- Decision strongly contradicted gold labels\n- Issuing negative reinforcement penalty", "Dataset: <b>SYNTHETIC DEMO</b>", "### Episode Summary\n- **Final Reward:** 0.20\n- **Improvement:** 0.0%\n- **Steps:** 1", "### Best Reward: 0.20", "### Recent Delta: --", "### Confidence: 20%"
+                    time.sleep(2)
+                    
+                    # Step 2
+                    df2 = pd.DataFrame([{"Step": 1, "Reward": 0.2}, {"Step": 2, "Reward": 0.55}])
+                    yield df2, "Decent response but can be improved in clarity", "Improve structure and clarity", "### Agent Process (Demo):\n- Compared Response A and B against Gold Standard\n- Decision showed partial alignment\n- Issuing moderate reinforcement", "Dataset: <b>SYNTHETIC DEMO</b>", "### Episode Summary\n- **Final Reward:** 0.55\n- **Improvement:** +175.0%\n- **Steps:** 2", "### Best Reward: 0.55", "### Recent Delta: +0.35", "### Confidence: 60%"
+                    time.sleep(2)
+                    
+                    # Step 3
+                    df3 = pd.DataFrame([{"Step": 1, "Reward": 0.2}, {"Step": 2, "Reward": 0.55}, {"Step": 3, "Reward": 0.99}])
+                    yield df3, "High quality response, well aligned with user intent", "Try making the response more concise", "### Agent Process (Demo):\n- Evaluated response on 4 heuristic axes (Helpfulness, Honesty, etc)\n- Decision matched gold labels almost perfectly\n- Issuing high positive reinforcement", "Dataset: <b>SYNTHETIC DEMO</b>", "### Episode Summary\n- **Final Reward:** 0.99\n- **Improvement:** +395.0%\n- **Steps:** 3", "### Best Reward: 0.99", "### Recent Delta: +0.44", "### Confidence: 95%"
+
+                demo_btn.click(
+                    fn=run_demo_mode,
+                    inputs=None,
+                    outputs=[reward_plot, reward_explanation, improvement_tip, agent_thinking, dataset_vis, session_summary, best_reward_disp, reward_delta_disp, confidence_disp]
                 )
             return blocks