fix: hybird not functioning. Change back from xgboost+geo to MLP+geo for hybird

checkpoints/hybrid_focus_config.json

@@ -1,14 +1,10 @@
 {
-  "use_xgb": true,
+  "use_xgb": false,
   "w_mlp": 0.3,
-  "w_xgb": 0.3,
   "w_geo": 0.7,
-  "threshold": 0.46117913373775393,
+  "threshold": 0.35,
   "use_yawn_veto": true,
   "geo_face_weight": 0.7,
   "geo_eye_weight": 0.3,
-  "mar_yawn_threshold": 0.55,
-  "metric": "f1",
-  "combiner": "logistic",
-  "combiner_path": "hybrid_combiner.joblib"
+  "mar_yawn_threshold": 0.55
 }

src/components/FocusPageLocal.jsx

@@ -43,14 +43,14 @@ const MODEL_INFO = {
   hybrid: {
     label: 'Hybrid',
     tagline: 'Best overall — combines ML with geometric scoring',
-    how: 'Fuses XGBoost predictions (30%) with geometric face/eye scores (70%). Uses a weighted blend tuned with LOPO evaluation.',
+    how: 'Fuses MLP predictions (30%) with geometric face/eye scores (70%). Uses a weighted blend tuned with LOPO evaluation.',
     accuracy: 'N/A',
     f1: '0.8409',
     auc: 'N/A',
-    threshold: '0.46',
+    threshold: '0.35',
     evaluation: 'LOPO tuning (9 participants, 144K frames)',
     features: '10 features: head deviation, face score, eye scores (EAR), gaze offset, pitch, horizontal gaze, PERCLOS',
-    strengths: 'Most robust across different people. Latest LOPO mean F1 is 0.8409 at w_mlp=0.3.',
+    strengths: 'Most robust across different people. LOPO mean F1 is 0.8409 at w_mlp=0.3, w_geo=0.7.',
     badge: 'Recommended',
   },
   xgboost: {
@@ -109,10 +109,6 @@ function FocusPageLocal({ videoManager, sessionResult, setSessionResult, isActiv
   const [calibration, setCalibration] = useState(null);
   const [l2csBoost, setL2csBoost] = useState(false);
   const [l2csBoostAvailable, setL2csBoostAvailable] = useState(false);
-  const [showEyeGazeModal, setShowEyeGazeModal] = useState(false);
-  const [eyeGazeDontShow, setEyeGazeDontShow] = useState(
-    () => localStorage.getItem('focusguard_eyegaze_noshowalert') === 'true'
-  );
 
   const localVideoRef = useRef(null);
   const displayCanvasRef = useRef(null);
@@ -331,20 +327,7 @@ function FocusPageLocal({ videoManager, sessionResult, setSessionResult, isActiv
 
   const handleEyeGazeToggle = async () => {
     const next = !l2csBoost;
-    if (next && !eyeGazeDontShow) {
-      // Show the warning/calibration modal before enabling
-      setShowEyeGazeModal(true);
-      return;
-    }
-    await applyEyeGazeChange(next, true);
-  };
-
-  const handleEyeGazeModalAction = async (withCalibration) => {
-    if (eyeGazeDontShow) {
-      localStorage.setItem('focusguard_eyegaze_noshowalert', 'true');
-    }
-    setShowEyeGazeModal(false);
-    await applyEyeGazeChange(true, withCalibration);
+    await applyEyeGazeChange(next, false);
   };
 
   const handleStart = async () => {
@@ -700,91 +683,9 @@ function FocusPageLocal({ videoManager, sessionResult, setSessionResult, isActiv
     return null;
   };
 
-  const renderEyeGazeModal = () => {
-    if (!showEyeGazeModal) return null;
-    return (
-      <div className="focus-flow-overlay" style={{ zIndex: 2000 }}>
-        <div className="focus-flow-card">
-          <div className="focus-flow-header">
-            <div>
-              <div className="focus-flow-eyebrow">Eye Gaze Tracking</div>
-              <h2>Before you enable</h2>
-            </div>
-            <div className="focus-flow-icon">
-              <svg width="96" height="96" viewBox="0 0 96 96" aria-hidden="true">
-                <ellipse cx="48" cy="48" rx="38" ry="24" fill="none" stroke="#007BFF" strokeWidth="5" />
-                <circle cx="48" cy="48" r="13" fill="none" stroke="#007BFF" strokeWidth="5" />
-                <circle cx="48" cy="48" r="5" fill="#007BFF" />
-              </svg>
-            </div>
-          </div>
-
-          <p className="focus-flow-lead">
-            Eye gaze tracking runs an additional deep neural network (L2CS-Net) alongside your current model.
-            Please read the notes below before proceeding.
-          </p>
-
-          <div className="focus-flow-grid">
-            <article className="focus-flow-panel focus-flow-panel-warn">
-              <h3>Performance impact</h3>
-              <p>Enabling eye gaze tracking increases CPU usage and may reduce frame rate. If the system feels sluggish, consider disabling it.</p>
-            </article>
-            <article className="focus-flow-panel">
-              <h3>Calibration (recommended)</h3>
-              <p>For best accuracy, calibrate by looking at 9 screen positions one at a time, followed by 1 validation point. The whole process takes about 30 seconds.</p>
-            </article>
-          </div>
-
-          <div className="focus-flow-steps">
-            <div className="focus-flow-step">
-              <div className="focus-flow-step-number">1</div>
-              <div className="focus-flow-step-copy">
-                <h3>Click "Start Calibration"</h3>
-                <p>A dot will appear on screen. Look directly at it and keep your gaze steady. It will cycle through 9 positions then show a final validation dot.</p>
-              </div>
-            </div>
-            <div className="focus-flow-step">
-              <div className="focus-flow-step-number">2</div>
-              <div className="focus-flow-step-copy">
-                <h3>Or skip for now</h3>
-                <p>Click "Skip" to enable eye gaze tracking without calibrating. You can recalibrate at any time using the "Recalibrate" button during a session.</p>
-              </div>
-            </div>
-          </div>
-
-          <label className="eye-gaze-modal-checkbox">
-            <input
-              type="checkbox"
-              checked={eyeGazeDontShow}
-              onChange={(e) => setEyeGazeDontShow(e.target.checked)}
-            />
-            Don't show this again
-          </label>
-
-          <div className="focus-flow-footer">
-            <button
-              type="button"
-              className="focus-flow-secondary"
-              onClick={() => handleEyeGazeModalAction(false)}
-            >
-              Skip
-            </button>
-            <button
-              className="focus-flow-button"
-              onClick={() => handleEyeGazeModalAction(true)}
-            >
-              Start Calibration
-            </button>
-          </div>
-        </div>
-      </div>
-    );
-  };
-
   return (
     <main id="page-b" className="page" style={pageStyle}>
       {renderIntroCard()}
-      {renderEyeGazeModal()}
 
       <section id="display-area" className="focus-display-shell">
         <video

src/components/Help.jsx

@@ -52,7 +52,7 @@ function Help() {
 
         <section className="help-section">
           <h2>Available Models</h2>
-          <p><strong>Hybrid</strong> <em>(Recommended)</em>: Combines XGBoost predictions with geometric face/eye scoring using a trained logistic regression combiner. Most robust across different people. LOPO F1: 0.8409.</p>
+          <p><strong>Hybrid</strong> <em>(Recommended)</em>: Blends MLP predictions (30%) with geometric face/eye scoring (70%) using a weighted average tuned with LOPO evaluation. Most robust across different people. LOPO F1: 0.8409.</p>
           <p><strong>XGBoost:</strong> Gradient-boosted tree model using 10 selected features. Highest raw accuracy (95.87% pooled, LOPO AUC 0.8695). Strong on tabular data with fast inference.</p>
           <p><strong>MLP:</strong> Two-layer neural network (10→64→32 neurons) trained with PyTorch. Good balance of speed and accuracy (92.92% pooled, LOPO AUC 0.8624). Fastest inference.</p>
           <p><strong>Geometric:</strong> Rule-based scoring using head pose and eye openness. No ML model needed — lightweight fallback when model checkpoints are unavailable. LOPO F1: 0.8195.</p>
@@ -66,13 +66,13 @@ function Help() {
           <p>The <strong>Eye Gaze</strong> button enables L2CS-Net, a deep neural network that estimates your gaze direction from the eye region. It runs alongside your selected base model and can improve focus detection accuracy.</p>
           <p style={{ marginTop: '8px' }}><strong>Performance note:</strong> Eye gaze tracking increases CPU usage and may reduce frame rate. If the system feels sluggish, disable it.</p>
           <h3 style={{ marginTop: '14px', fontSize: '1rem' }}>Calibration</h3>
-          <p>For best accuracy, calibrate when prompted:</p>
+          <p>For best accuracy, calibrate when Eye Gaze is active:</p>
           <ol>
-            <li>Click "Start Calibration" in the warning dialog (or use the "Recalibrate" button during a session)</li>
+            <li>Start a session, then click the <strong>"Recalibrate"</strong> button that appears in the model strip when Eye Gaze is on</li>
             <li>Look directly at each dot as it appears on screen — there are <strong>9 calibration points</strong> across the screen</li>
             <li>A final <strong>validation point</strong> confirms accuracy before calibration is applied</li>
           </ol>
-          <p>You can skip calibration and recalibrate at any time using the "Recalibrate" button, which appears in the model strip when Eye Gaze is on and a session is running.</p>
+          <p>You can recalibrate at any time using the "Recalibrate" button, which appears in the model strip when Eye Gaze is on and a session is running.</p>
         </section>
 
         <section className="help-section">
@@ -134,8 +134,8 @@ function Help() {
             <p>The face mesh overlay updates each time the server returns a detection result. The camera feed itself renders locally. Any visible lag depends on network latency and server processing time. Reducing the frame rate slider can help if lag is noticeable.</p>
           </details>
           <details>
-            <summary>The Hybrid model doesn't seem to work differently from XGBoost — why?</summary>
-            <p>The Hybrid model uses a trained logistic regression combiner on top of XGBoost and geometric scores. If the combiner file wasn't available at startup, it falls back to a simple weighted average which behaves similarly to XGBoost. Check the server logs for "[HYBRID]" messages to confirm the combiner loaded correctly.</p>
+            <summary>The Hybrid model doesn't seem to work differently from MLP — why?</summary>
+            <p>The Hybrid model blends MLP (30%) and geometric (70%) scores using a fixed weighted average. Because its geometric component dominates, it tends to be more conservative than raw MLP — especially when head pose or eye openness signals are borderline. It is tuned to be the most consistent across different people rather than the most aggressive.</p>
           </details>
         </section>
 
@@ -143,7 +143,7 @@ function Help() {
           <h2>Technical Info</h2>
           <p><strong>Face Detection:</strong> MediaPipe Face Mesh (478 landmarks)</p>
           <p><strong>Feature Extraction:</strong> Head pose (yaw/pitch/roll), EAR, MAR, gaze offset, PERCLOS, blink rate — 10 features selected via LOFO analysis</p>
-          <p><strong>ML Models:</strong> PyTorch MLP (10→64→32→2), XGBoost (600 trees), Geometric (rule-based), Hybrid (LR combiner over XGBoost+Geo)</p>
+          <p><strong>ML Models:</strong> PyTorch MLP (10→64→32→2), XGBoost (600 trees), Geometric (rule-based), Hybrid (MLP 30% + Geometric 70% weighted blend)</p>
           <p><strong>Eye Gaze:</strong> L2CS-Net (ResNet50 backbone, trained on Gaze360) with 9-point polynomial calibration</p>
           <p><strong>Storage:</strong> SQLite database (sessions, events, settings)</p>
           <p><strong>Framework:</strong> FastAPI + React (Vite) + WebSocket</p>