Tune action detection sensitivity, fix behavior label display

Stage 6: REQUIRED_CONSECUTIVE 8→12, decay -2 on miss, confidence
gate >0.4, tighten clap (0.3×scale) and touch_head (0.25×scale).
Dot counter updated to 12.

Stage 8: Behavior label now recomputed from actual probabilities —
shows "Normal Activity" when P(non_autistic) > 50% instead of
always showing hand_flapping due to model bias.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Partha-dev01

DOCS.md

@@ -612,3 +612,32 @@ npx playwright test    # Run all 30 tests
 
 **Files:**
 - Modified: `app/intake/video-capture/page.tsx` (setCamReady moved, stream re-attach effect, description text fix)
+
+### v1.6.2 — 2026-03-05 (Action Detection Tuning, Behavior Label Fix)
+
+**Fixed:**
+- **Stage 6 (Action Challenge) hypersensitive detection**: Actions were confirming too quickly (~0.5s) due to low `REQUIRED_CONSECUTIVE` (8) and slow decay (-1 per miss). False positives from YOLO keypoint noise accumulated faster than they decayed.
+  - `REQUIRED_CONSECUTIVE`: 8 → 12 (~0.8s sustained detection at 15fps)
+  - Decay on miss: -1 → -2 (noise drops out faster)
+  - Added confidence gate: only count hits with confidence > 0.4
+  - Tightened clap threshold: 0.4×scale → 0.3×scale
+  - Tightened touch_head threshold: 0.3×scale → 0.25×scale (less overlap with touch_nose)
+- **Stage 8 (Video Capture) "Hand Flapping" always shown as behavior label**: The body TCN model is biased toward hand_flapping (F1: 0.68) due to training data class imbalance (non_autistic F1 only 0.33). The video overlay now recomputes the display label from actual probabilities:
+  - When P(non_autistic) > 50%: shows "Normal Activity" with a green badge
+  - Otherwise: shows the highest ASD-related behavior class (indices 0-4)
+  - Eliminates frame-mismatch display inconsistencies
+
+**Training Model Notes (from `Autism_code/` analysis):**
+- Body TCN best F1: 0.384 (macro-averaged, 6 classes)
+- Classes 2 (head_banging) and 4 (toe_walking): 0.0 F1 — zero validation samples
+- Class 0 (hand_flapping): 0.68 F1 — strongest, causes bias
+- Class 5 (non_autistic): 0.33 F1 — weak recognition
+- Long-term fix: retrain with balanced data. Current UI-side fix handles the display bias.
+
+**CI Playwright Warnings (NOT a bug):**
+- `CredentialsProviderError` messages in CI logs are expected — CI lacks AWS credentials. Bedrock/Polly APIs fail gracefully and tests verify fallback/mock responses. All 31 tests pass.
+
+**Files:**
+- Modified: `app/lib/actions/actionDetector.ts` (REQUIRED_CONSECUTIVE 12, decay -2, confidence gate, tighter thresholds)
+- Modified: `app/intake/preparation/page.tsx` (12-dot counter, status text thresholds)
+- Modified: `app/components/DetectorVideoCanvas.tsx` (behavior label recomputed from probs, "Normal Activity" display)

app/components/DetectorVideoCanvas.tsx

@@ -6,6 +6,7 @@
 "use client";
 import { useRef, useEffect } from "react";
 import type { PipelineResult } from "../types/inference";
+import { BEHAVIOR_CLASSES } from "../types/inference";
 
 // COCO-17 skeleton connections
 const SKELETON: [number, number][] = [
@@ -163,21 +164,41 @@ export default function DetectorVideoCanvas({ videoRef, canvasRef, result, isCam
         </div>
       )}
 
-      {/* Behavior label */}
-      {result?.behavior && (
-        <div style={{
-          position: "absolute", top: 8, left: 8, padding: "6px 14px",
-          borderRadius: "var(--r-md)", fontSize: "0.85rem", fontWeight: 700,
-          fontFamily: "'Fredoka',sans-serif",
-          background: "rgba(0,0,0,0.65)", color: "white",
-          border: "1px solid rgba(255,255,255,0.15)",
-        }}>
-          {fmt(result.behavior.className)}
-          <span style={{ marginLeft: 8, fontSize: "0.75rem", opacity: 0.7 }}>
-            {(result.behavior.probabilities[result.behavior.predictedClass] * 100).toFixed(0)}%
-          </span>
-        </div>
-      )}
+      {/* Behavior label — show "Normal Activity" when non_autistic > 50%, otherwise top ASD class */}
+      {result?.behavior && result.behavior.probabilities.length >= 6 && (() => {
+        const probs = result.behavior!.probabilities;
+        const nonAutisticProb = probs[5];
+        let displayClass: string;
+        let displayProb: number;
+        if (nonAutisticProb > 0.5) {
+          displayClass = "Normal Activity";
+          displayProb = nonAutisticProb;
+        } else {
+          // Find highest ASD behavior (indices 0-4)
+          let maxIdx = 0, maxP = probs[0];
+          for (let i = 1; i < 5; i++) {
+            if (probs[i] > maxP) { maxP = probs[i]; maxIdx = i; }
+          }
+          displayClass = fmt(BEHAVIOR_CLASSES[maxIdx]);
+          displayProb = maxP;
+        }
+        const isNormal = nonAutisticProb > 0.5;
+        return (
+          <div style={{
+            position: "absolute", top: 8, left: 8, padding: "6px 14px",
+            borderRadius: "var(--r-md)", fontSize: "0.85rem", fontWeight: 700,
+            fontFamily: "'Fredoka',sans-serif",
+            background: isNormal ? "rgba(58,99,68,0.75)" : "rgba(0,0,0,0.65)",
+            color: "white",
+            border: "1px solid rgba(255,255,255,0.15)",
+          }}>
+            {displayClass}
+            <span style={{ marginLeft: 8, fontSize: "0.75rem", opacity: 0.7 }}>
+              {(displayProb * 100).toFixed(0)}%
+            </span>
+          </div>
+        );
+      })()}
     </div>
   );
 }

app/intake/preparation/page.tsx

@@ -410,12 +410,12 @@ export default function PreparationPage() {
                   </div>
                 )}
 
-                {/* 8-dot frame counter (matches REQUIRED_CONSECUTIVE = 8) */}
+                {/* 12-dot frame counter (matches REQUIRED_CONSECUTIVE = 12) */}
                 {actionPhase === "detecting" && (
-                  <div style={{ display: "flex", gap: 4, justifyContent: "center", marginBottom: 8 }}>
-                    {Array.from({ length: 8 }, (_, i) => (
+                  <div style={{ display: "flex", gap: 3, justifyContent: "center", marginBottom: 8 }}>
+                    {Array.from({ length: 12 }, (_, i) => (
                       <div key={i} style={{
-                        width: 16, height: 16, borderRadius: "50%",
+                        width: 14, height: 14, borderRadius: "50%",
                         background: i < consecutiveHits ? "var(--sage-500)" : "var(--bg-elevated)",
                         border: "2px solid var(--sage-300)",
                         transition: "background 0.2s",
@@ -429,12 +429,12 @@ export default function PreparationPage() {
                   <p style={{
                     fontSize: "0.9rem", fontWeight: 700, marginBottom: 8,
                     color: !hasKeypoints ? "var(--text-muted)"
-                      : consecutiveHits >= 3 ? "var(--sage-600)"
+                      : consecutiveHits >= 5 ? "var(--sage-600)"
                       : (actionResult?.confidence || 0) > 0 ? "var(--sky-400)"
                       : "var(--text-muted)",
                   }}>
                     {!hasKeypoints ? "Step into view so we can see you!"
-                      : consecutiveHits >= 5 ? "Almost there! Keep holding..."
+                      : consecutiveHits >= 7 ? "Almost there! Keep holding..."
                       : (actionResult?.confidence || 0) > 0.3 ? "Getting closer!"
                       : `Looking for: ${meta.label}...`}
                   </p>

app/lib/actions/actionDetector.ts

@@ -128,7 +128,7 @@ function detectClap(
   if (!confOk(conf, L_WRIST, R_WRIST))
     return { hit: false, proximity: 0 };
   const d = dist(kp(kps, L_WRIST), kp(kps, R_WRIST));
-  const threshold = 0.4 * scale; // Relaxed: hands near each other (YOLO wrist keypoints are imprecise during clap)
+  const threshold = 0.3 * scale; // Tightened: require hands closer together
   return { hit: d < threshold, proximity: Math.max(0, 1 - d / threshold) };
 }
 
@@ -156,7 +156,7 @@ function detectTouchHead(
   const dL = confOk(conf, L_WRIST) ? dist(kp(kps, L_WRIST), nose) : Infinity;
   const dR = confOk(conf, R_WRIST) ? dist(kp(kps, R_WRIST), nose) : Infinity;
   const minD = Math.min(dL, dR);
-  const threshold = 0.3 * scale;
+  const threshold = 0.25 * scale;
   return { hit: minD < threshold, proximity: Math.max(0, 1 - minD / threshold) };
 }
 
@@ -229,7 +229,7 @@ export function detectAction(
 
 // ── Sustained detection tracker ─────────────────────────────────────
 
-const REQUIRED_CONSECUTIVE = 8;
+const REQUIRED_CONSECUTIVE = 12;
 
 export class ActionTracker {
   private consecutiveHits = 0;
@@ -253,10 +253,10 @@ export class ActionTracker {
 
     const result = detectAction(keypoints, confidence, action, this.history);
 
-    if (result.detected) {
+    if (result.detected && result.confidence > 0.4) {
       this.consecutiveHits++;
     } else {
-      this.consecutiveHits = Math.max(0, this.consecutiveHits - 1);
+      this.consecutiveHits = Math.max(0, this.consecutiveHits - 2);
     }
 
     if (this.consecutiveHits >= REQUIRED_CONSECUTIVE) {