rollingscan very clean

likelyhood -> find similar
REFRESHCLAUD.MD 20260126

app/src/main/java/com/placeholder/sherpai2/data/local/AppDatabase.kt

@@ -10,6 +10,10 @@ import com.placeholder.sherpai2.data.local.entity.*
 /**
  * AppDatabase - Complete database for SherpAI2
  *
+ * VERSION 12 - Distribution-based rejection stats
+ * - Added similarityStdDev, similarityMin to FaceModelEntity
+ * - Enables self-calibrating threshold for face matching
+ *
  * VERSION 10 - User Feedback Loop
  * - Added UserFeedbackEntity for storing user corrections
  * - Enables cluster refinement before training
@@ -52,7 +56,7 @@ import com.placeholder.sherpai2.data.local.entity.*
         CollectionImageEntity::class,
         CollectionFilterEntity::class
     ],
-    version = 11,  // INCREMENTED for person statistics
+    version = 12,  // INCREMENTED for distribution-based rejection stats
     exportSchema = false
 )
 abstract class AppDatabase : RoomDatabase() {
@@ -272,13 +276,32 @@ val MIGRATION_10_11 = object : Migration(10, 11) {
     }
 }
 
+/**
+ * MIGRATION 11 → 12 (Distribution-based Rejection Stats)
+ *
+ * Changes:
+ * 1. Add similarityStdDev column to face_models (default 0.05)
+ * 2. Add similarityMin column to face_models (default 0.6)
+ *
+ * These fields enable self-calibrating thresholds during scanning.
+ * During training, we compute stats from training sample similarities
+ * and use (mean - 2*stdDev) as a floor for matching.
+ */
+val MIGRATION_11_12 = object : Migration(11, 12) {
+    override fun migrate(database: SupportSQLiteDatabase) {
+        // Add distribution stats columns with sensible defaults for existing models
+        database.execSQL("ALTER TABLE face_models ADD COLUMN similarityStdDev REAL NOT NULL DEFAULT 0.05")
+        database.execSQL("ALTER TABLE face_models ADD COLUMN similarityMin REAL NOT NULL DEFAULT 0.6")
+    }
+}
+
 /**
  * PRODUCTION MIGRATION NOTES:
  *
  * Before shipping to users, update DatabaseModule to use migrations:
  *
  * Room.databaseBuilder(context, AppDatabase::class.java, "sherpai.db")
- *     .addMigrations(MIGRATION_7_8, MIGRATION_8_9, MIGRATION_9_10, MIGRATION_10_11)  // Add all migrations
+ *     .addMigrations(MIGRATION_7_8, MIGRATION_8_9, MIGRATION_9_10, MIGRATION_10_11, MIGRATION_11_12)  // Add all migrations
  *     // .fallbackToDestructiveMigration()  // Remove this
  *     .build()
  */

app/src/main/java/com/placeholder/sherpai2/data/local/entity/Facerecognitionentities.kt

@@ -143,6 +143,13 @@ data class FaceModelEntity(
     @ColumnInfo(name = "averageConfidence")
     val averageConfidence: Float,
 
+    // Distribution stats for self-calibrating rejection
+    @ColumnInfo(name = "similarityStdDev")
+    val similarityStdDev: Float = 0.05f,  // Default for backwards compat
+
+    @ColumnInfo(name = "similarityMin")
+    val similarityMin: Float = 0.6f,      // Default for backwards compat
+
     @ColumnInfo(name = "createdAt")
     val createdAt: Long,
 
@@ -157,26 +164,29 @@ data class FaceModelEntity(
 ) {
     companion object {
         /**
-         * Backwards compatible create() method
-         * Used by existing FaceRecognitionRepository code
+         * Create with distribution stats for self-calibrating rejection
          */
         fun create(
             personId: String,
             embeddingArray: FloatArray,
             trainingImageCount: Int,
-            averageConfidence: Float
+            averageConfidence: Float,
+            similarityStdDev: Float = 0.05f,
+            similarityMin: Float = 0.6f
         ): FaceModelEntity {
-            return createFromEmbedding(personId, embeddingArray, trainingImageCount, averageConfidence)
+            return createFromEmbedding(personId, embeddingArray, trainingImageCount, averageConfidence, similarityStdDev, similarityMin)
         }
 
         /**
-         * Create from single embedding (backwards compatible)
+         * Create from single embedding with distribution stats
          */
         fun createFromEmbedding(
             personId: String,
             embeddingArray: FloatArray,
             trainingImageCount: Int,
-            averageConfidence: Float
+            averageConfidence: Float,
+            similarityStdDev: Float = 0.05f,
+            similarityMin: Float = 0.6f
         ): FaceModelEntity {
             val now = System.currentTimeMillis()
             val centroid = TemporalCentroid(
@@ -194,6 +204,8 @@ data class FaceModelEntity(
                 centroidsJson = serializeCentroids(listOf(centroid)),
                 trainingImageCount = trainingImageCount,
                 averageConfidence = averageConfidence,
+                similarityStdDev = similarityStdDev,
+                similarityMin = similarityMin,
                 createdAt = now,
                 updatedAt = now,
                 lastUsed = null,

app/src/main/java/com/placeholder/sherpai2/data/repository/Facerecognitionrepository.kt

@@ -99,11 +99,19 @@ class FaceRecognitionRepository @Inject constructor(
         }
         val avgConfidence = confidences.average().toFloat()
 
+        // Compute distribution stats for self-calibrating rejection
+        val stdDev = kotlin.math.sqrt(
+            confidences.map { (it - avgConfidence).toDouble().let { d -> d * d } }.average()
+        ).toFloat()
+        val minSimilarity = confidences.minOrNull() ?: 0f
+
         val faceModel = FaceModelEntity.create(
             personId = personId,
             embeddingArray = personEmbedding,
             trainingImageCount = validImages.size,
-            averageConfidence = avgConfidence
+            averageConfidence = avgConfidence,
+            similarityStdDev = stdDev,
+            similarityMin = minSimilarity
         )
 
         faceModelDao.insertFaceModel(faceModel)

app/src/main/java/com/placeholder/sherpai2/domain/clustering/Facequalityfilter.kt

@@ -29,6 +29,64 @@ import kotlin.math.sqrt
  */
 object FaceQualityFilter {
 
+    /**
+     * Age group estimation for filtering (child vs adult detection)
+     */
+    enum class AgeGroup { CHILD, ADULT, UNCERTAIN }
+
+    /**
+     * Estimate whether a face belongs to a child or adult based on facial proportions.
+     *
+     * Uses two heuristics:
+     * 1. Eye position ratio - Children have larger foreheads, so eyes are lower (~45% from top)
+     *    Adults have eyes at ~35% from top
+     * 2. Face roundness (width/height ratio) - Children: ~0.85-1.0, Adults: ~0.7-0.85
+     *
+     * @return AgeGroup.CHILD, AgeGroup.ADULT, or AgeGroup.UNCERTAIN
+     */
+    fun estimateAgeGroup(face: Face, imageWidth: Int, imageHeight: Int): AgeGroup {
+        val leftEye = face.getLandmark(FaceLandmark.LEFT_EYE)
+        val rightEye = face.getLandmark(FaceLandmark.RIGHT_EYE)
+
+        if (leftEye == null || rightEye == null) {
+            return AgeGroup.UNCERTAIN
+        }
+
+        // Eye-to-face height ratio (where eyes sit relative to face top)
+        val faceHeight = face.boundingBox.height().toFloat()
+        val faceTop = face.boundingBox.top.toFloat()
+        val eyeY = (leftEye.position.y + rightEye.position.y) / 2
+        val eyePositionRatio = (eyeY - faceTop) / faceHeight
+
+        // Children: eyes at ~45% from top (larger forehead proportionally)
+        // Adults: eyes at ~35% from top
+        // Score: higher = more child-like
+
+        // Face roundness (width/height)
+        val faceWidth = face.boundingBox.width().toFloat()
+        val faceRatio = faceWidth / faceHeight
+        // Children: ratio ~0.85-1.0 (rounder faces)
+        // Adults: ratio ~0.7-0.85 (longer/narrower faces)
+
+        var childScore = 0
+
+        // Eye position scoring
+        if (eyePositionRatio > 0.45f) childScore += 2  // Strong child signal
+        else if (eyePositionRatio > 0.42f) childScore += 1  // Mild child signal
+        else if (eyePositionRatio < 0.35f) childScore -= 1  // Adult signal
+
+        // Face roundness scoring
+        if (faceRatio > 0.90f) childScore += 2  // Very round = child
+        else if (faceRatio > 0.82f) childScore += 1  // Somewhat round
+        else if (faceRatio < 0.75f) childScore -= 1  // Long face = adult
+
+        return when {
+            childScore >= 3 -> AgeGroup.CHILD
+            childScore <= 0 -> AgeGroup.ADULT
+            else -> AgeGroup.UNCERTAIN
+        }
+    }
+
     /**
      * Validate face for Discovery/Clustering
      *

app/src/main/java/com/placeholder/sherpai2/ui/modelinventory/Personinventoryviewmodel.kt

@@ -19,6 +19,7 @@ import com.placeholder.sherpai2.data.local.entity.PersonEntity
 import com.placeholder.sherpai2.data.local.entity.PhotoFaceTagEntity
 import com.placeholder.sherpai2.ml.FaceNetModel
 import com.placeholder.sherpai2.ml.ThresholdStrategy
+import com.placeholder.sherpai2.domain.clustering.FaceQualityFilter
 import dagger.hilt.android.lifecycle.HiltViewModel
 import dagger.hilt.android.qualifiers.ApplicationContext
 import kotlinx.coroutines.Dispatchers
@@ -142,7 +143,7 @@ class PersonInventoryViewModel @Inject constructor(
 
                 val detectorOptions = FaceDetectorOptions.Builder()
                     .setPerformanceMode(FaceDetectorOptions.PERFORMANCE_MODE_ACCURATE)
-                    .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_NONE)
+                    .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_ALL)  // Needed for age estimation
                     .setClassificationMode(FaceDetectorOptions.CLASSIFICATION_MODE_NONE)
                     .setMinFaceSize(0.15f)
                     .build()
@@ -159,9 +160,23 @@ class PersonInventoryViewModel @Inject constructor(
                 }
 
                 val faceNetModel = FaceNetModel(context)
-                // Production threshold - balance precision vs recall
-                val baseThreshold = 0.58f
-                android.util.Log.d("PersonScan", "Using threshold: $baseThreshold, centroids: ${modelCentroids.size}")
+                // Production threshold - STRICT to avoid false positives
+                // Solo face photos: 0.62, Group photos: 0.68
+                val baseThreshold = 0.62f
+                val groupPhotoThreshold = 0.68f  // Higher bar for multi-face images
+
+                // Load ALL other models for "best match wins" comparison
+                val allModels = faceModelDao.getAllActiveFaceModels()
+                val otherModelCentroids = allModels
+                    .filter { it.id != faceModel.id }
+                    .map { model -> model.id to model.getCentroids().map { it.getEmbeddingArray() } }
+
+                // Distribution-based minimum threshold (self-calibrating)
+                val distributionMin = (faceModel.averageConfidence - 2 * faceModel.similarityStdDev)
+                    .coerceAtLeast(faceModel.similarityMin - 0.05f)
+                    .coerceAtLeast(0.50f)  // Never go below 0.50 absolute floor
+
+                android.util.Log.d("PersonScan", "Using threshold: solo=$baseThreshold, group=$groupPhotoThreshold, distributionMin=$distributionMin (avgConf=${faceModel.averageConfidence}, stdDev=${faceModel.similarityStdDev}), centroids: ${modelCentroids.size}, competing models: ${otherModelCentroids.size}, isChild=${person.isChild}")
 
                 val completed = AtomicInteger(0)
                 val facesFound = AtomicInteger(0)
@@ -173,7 +188,7 @@ class PersonInventoryViewModel @Inject constructor(
                     val jobs = untaggedImages.map { image ->
                         async {
                             semaphore.withPermit {
-                                processImage(image, detector, faceNetModel, modelCentroids, trainingCount, baseThreshold, personId, faceModel.id, batchMatches, batchUpdateMutex, completed, facesFound, startTime, totalToScan, person.name)
+                                processImage(image, detector, faceNetModel, modelCentroids, otherModelCentroids, trainingCount, baseThreshold, groupPhotoThreshold, distributionMin, person.isChild, personId, faceModel.id, batchMatches, batchUpdateMutex, completed, facesFound, startTime, totalToScan, person.name)
                             }
                         }
                     }
@@ -200,7 +215,10 @@ class PersonInventoryViewModel @Inject constructor(
 
     private suspend fun processImage(
         image: ImageEntity, detector: com.google.mlkit.vision.face.FaceDetector, faceNetModel: FaceNetModel,
-        modelCentroids: List<FloatArray>, trainingCount: Int, baseThreshold: Float, personId: String, faceModelId: String,
+        modelCentroids: List<FloatArray>, otherModelCentroids: List<Pair<String, List<FloatArray>>>,
+        trainingCount: Int, baseThreshold: Float, groupPhotoThreshold: Float,
+        distributionMin: Float, isChildTarget: Boolean,
+        personId: String, faceModelId: String,
         batchMatches: MutableList<Triple<String, String, Float>>, batchUpdateMutex: Mutex,
         completed: AtomicInteger, facesFound: AtomicInteger, startTime: Long, totalToScan: Int, personName: String
     ) {
@@ -225,9 +243,13 @@ class PersonInventoryViewModel @Inject constructor(
             val scaleX = sizeOpts.outWidth.toFloat() / detectionBitmap.width
             val scaleY = sizeOpts.outHeight.toFloat() / detectionBitmap.height
 
-            val imageQuality = ThresholdStrategy.estimateImageQuality(sizeOpts.outWidth, sizeOpts.outHeight)
-            val detectionContext = ThresholdStrategy.estimateDetectionContext(faces.size)
-            val threshold = ThresholdStrategy.getOptimalThreshold(trainingCount, imageQuality, detectionContext).coerceAtMost(baseThreshold)
+            // CRITICAL: Use higher threshold for group photos (more likely false positives)
+            val isGroupPhoto = faces.size > 1
+            val effectiveThreshold = if (isGroupPhoto) groupPhotoThreshold else baseThreshold
+
+            // Track best match in this image (only tag ONE face per image)
+            var bestMatchSimilarity = 0f
+            var foundMatch = false
 
             for (face in faces) {
                 val scaledBounds = android.graphics.Rect(
@@ -237,27 +259,70 @@ class PersonInventoryViewModel @Inject constructor(
                     (face.boundingBox.bottom * scaleY).toInt()
                 )
 
+                // Skip very small faces (less reliable)
+                val faceArea = scaledBounds.width() * scaledBounds.height()
+                val imageArea = sizeOpts.outWidth * sizeOpts.outHeight
+                val faceRatio = faceArea.toFloat() / imageArea
+                if (faceRatio < 0.02f) continue  // Face must be at least 2% of image
+
+                // SIGNAL 2: Age plausibility check (if target is a child)
+                if (isChildTarget) {
+                    val ageGroup = FaceQualityFilter.estimateAgeGroup(face, detectionBitmap.width, detectionBitmap.height)
+                    if (ageGroup == FaceQualityFilter.AgeGroup.ADULT) {
+                        continue  // Reject clearly adult faces when searching for a child
+                    }
+                }
+
                 // CRITICAL: Add padding to face crop (same as training)
                 val faceBitmap = loadFaceRegionWithPadding(uri, scaledBounds, sizeOpts.outWidth, sizeOpts.outHeight) ?: continue
                 val faceEmbedding = faceNetModel.generateEmbedding(faceBitmap)
                 faceBitmap.recycle()
 
-                // Match against ALL centroids, use best match
-                val bestSimilarity = modelCentroids.maxOfOrNull { centroid ->
+                // Match against target person's centroids
+                val targetSimilarity = modelCentroids.maxOfOrNull { centroid ->
                     faceNetModel.calculateSimilarity(faceEmbedding, centroid)
                 } ?: 0f
 
-                if (bestSimilarity >= threshold) {
-                    batchUpdateMutex.withLock {
-                        batchMatches.add(Triple(personId, image.imageId, bestSimilarity))
-                        facesFound.incrementAndGet()
-                        if (batchMatches.size >= BATCH_DB_SIZE) {
-                            saveBatchMatches(batchMatches.toList(), faceModelId)
-                            batchMatches.clear()
-                        }
+                // SIGNAL 1: Distribution-based rejection
+                // If similarity is below (mean - 2*stdDev) or (min - 0.05), it's a statistical outlier
+                if (targetSimilarity < distributionMin) {
+                    continue  // Too far below training distribution
+                }
+
+                // SIGNAL 3: Basic threshold check
+                if (targetSimilarity < effectiveThreshold) {
+                    continue
+                }
+
+                // SIGNAL 4: "Best match wins" - check if any OTHER model scores higher
+                // This prevents tagging siblings/similar people incorrectly
+                val bestOtherSimilarity = otherModelCentroids.maxOfOrNull { (_, centroids) ->
+                    centroids.maxOfOrNull { centroid ->
+                        faceNetModel.calculateSimilarity(faceEmbedding, centroid)
+                    } ?: 0f
+                } ?: 0f
+
+                val isTargetBestMatch = targetSimilarity > bestOtherSimilarity
+
+                // All signals must pass
+                if (isTargetBestMatch && targetSimilarity > bestMatchSimilarity) {
+                    bestMatchSimilarity = targetSimilarity
+                    foundMatch = true
+                }
+            }
+
+            // Only add ONE tag per image (the best match)
+            if (foundMatch) {
+                batchUpdateMutex.withLock {
+                    batchMatches.add(Triple(personId, image.imageId, bestMatchSimilarity))
+                    facesFound.incrementAndGet()
+                    if (batchMatches.size >= BATCH_DB_SIZE) {
+                        saveBatchMatches(batchMatches.toList(), faceModelId)
+                        batchMatches.clear()
                     }
                 }
             }
+
             detectionBitmap.recycle()
         } catch (e: Exception) {
         } finally {

app/src/main/java/com/placeholder/sherpai2/ui/rollingscan/Rollingscanscreen.kt

@@ -2,7 +2,9 @@ package com.placeholder.sherpai2.ui.rollingscan
 
 import android.net.Uri
 import androidx.compose.foundation.BorderStroke
+import androidx.compose.foundation.ExperimentalFoundationApi
 import androidx.compose.foundation.clickable
+import androidx.compose.foundation.combinedClickable
 import androidx.compose.foundation.layout.*
 import androidx.compose.foundation.lazy.grid.GridCells
 import androidx.compose.foundation.lazy.grid.GridItemSpan
@@ -37,7 +39,7 @@ import com.placeholder.sherpai2.domain.similarity.FaceSimilarityScorer
  * - Quick action buttons (Select Top N)
  * - Submit button with validation
  */
-@OptIn(ExperimentalMaterial3Api::class)
+@OptIn(ExperimentalMaterial3Api::class, ExperimentalFoundationApi::class)
 @Composable
 fun RollingScanScreen(
     seedImageIds: List<String>,
@@ -48,6 +50,7 @@ fun RollingScanScreen(
 ) {
     val uiState by viewModel.uiState.collectAsState()
     val selectedImageIds by viewModel.selectedImageIds.collectAsState()
+    val negativeImageIds by viewModel.negativeImageIds.collectAsState()
     val rankedPhotos by viewModel.rankedPhotos.collectAsState()
     val isScanning by viewModel.isScanning.collectAsState()
 
@@ -70,6 +73,7 @@ fun RollingScanScreen(
                 isReadyForTraining = viewModel.isReadyForTraining(),
                 validationMessage = viewModel.getValidationMessage(),
                 onSelectTopN = { count -> viewModel.selectTopN(count) },
+                onSelectAboveThreshold = { threshold -> viewModel.selectAllAboveThreshold(threshold) },
                 onSubmit = {
                     val uris = viewModel.getSelectedImageUris()
                     onSubmitForTraining(uris)
@@ -93,8 +97,10 @@ fun RollingScanScreen(
                 RollingScanPhotoGrid(
                     rankedPhotos = rankedPhotos,
                     selectedImageIds = selectedImageIds,
+                    negativeImageIds = negativeImageIds,
                     isScanning = isScanning,
                     onToggleSelection = { imageId -> viewModel.toggleSelection(imageId) },
+                    onToggleNegative = { imageId -> viewModel.toggleNegative(imageId) },
                     modifier = Modifier.padding(padding)
                 )
             }
@@ -159,19 +165,26 @@ private fun RollingScanTopBar(
 }
 
 // ═══════════════════════════════════════════════════════════
-// PHOTO GRID
+// PHOTO GRID - Similarity-based bucketing
 // ═══════════════════════════════════════════════════════════
 
+@OptIn(ExperimentalFoundationApi::class)
 @Composable
 private fun RollingScanPhotoGrid(
     rankedPhotos: List<FaceSimilarityScorer.ScoredPhoto>,
     selectedImageIds: Set<String>,
+    negativeImageIds: Set<String>,
     isScanning: Boolean,
     onToggleSelection: (String) -> Unit,
+    onToggleNegative: (String) -> Unit,
     modifier: Modifier = Modifier
 ) {
-    Column(modifier = modifier.fillMaxSize()) {
+    // Bucket by similarity score
+    val veryLikely = rankedPhotos.filter { it.finalScore >= 0.60f }
+    val probably = rankedPhotos.filter { it.finalScore in 0.45f..0.599f }
+    val maybe = rankedPhotos.filter { it.finalScore < 0.45f }
 
+    Column(modifier = modifier.fillMaxSize()) {
         // Scanning indicator
         if (isScanning) {
             LinearProgressIndicator(
@@ -180,69 +193,78 @@ private fun RollingScanPhotoGrid(
             )
         }
 
+        // Hint for negative marking
+        Text(
+            text = "Tap to select • Long-press to mark as NOT this person",
+            style = MaterialTheme.typography.bodySmall,
+            color = MaterialTheme.colorScheme.onSurfaceVariant,
+            modifier = Modifier.padding(horizontal = 12.dp, vertical = 4.dp)
+        )
+
         LazyVerticalGrid(
             columns = GridCells.Fixed(3),
             contentPadding = PaddingValues(8.dp),
             horizontalArrangement = Arrangement.spacedBy(8.dp),
             verticalArrangement = Arrangement.spacedBy(8.dp)
         ) {
-            // Section: Most Similar (top 10)
-            val topMatches = rankedPhotos.take(10)
-            if (topMatches.isNotEmpty()) {
+            // Section: Very Likely (>60%)
+            if (veryLikely.isNotEmpty()) {
                 item(span = { GridItemSpan(3) }) {
                     SectionHeader(
                         icon = Icons.Default.Whatshot,
-                        text = "🔥 Most Similar (${topMatches.size})",
-                        color = MaterialTheme.colorScheme.primary
+                        text = "🟢 Very Likely (${veryLikely.size})",
+                        color = Color(0xFF4CAF50)
                     )
                 }
-
-                items(topMatches, key = { it.imageId }) { photo ->
+                items(veryLikely, key = { it.imageId }) { photo ->
                     PhotoCard(
                         photo = photo,
                         isSelected = photo.imageId in selectedImageIds,
+                        isNegative = photo.imageId in negativeImageIds,
                         onToggle = { onToggleSelection(photo.imageId) },
+                        onLongPress = { onToggleNegative(photo.imageId) },
                         showSimilarityBadge = true
                     )
                 }
             }
 
-            // Section: Good Matches (11-30)
-            val goodMatches = rankedPhotos.drop(10).take(20)
-            if (goodMatches.isNotEmpty()) {
+            // Section: Probably (45-60%)
+            if (probably.isNotEmpty()) {
                 item(span = { GridItemSpan(3) }) {
                     SectionHeader(
                         icon = Icons.Default.CheckCircle,
-                        text = "📊 Good Matches (${goodMatches.size})",
-                        color = MaterialTheme.colorScheme.tertiary
+                        text = "🟡 Probably (${probably.size})",
+                        color = Color(0xFFFFC107)
                     )
                 }
-
-                items(goodMatches, key = { it.imageId }) { photo ->
+                items(probably, key = { it.imageId }) { photo ->
                     PhotoCard(
                         photo = photo,
                         isSelected = photo.imageId in selectedImageIds,
-                        onToggle = { onToggleSelection(photo.imageId) }
+                        isNegative = photo.imageId in negativeImageIds,
+                        onToggle = { onToggleSelection(photo.imageId) },
+                        onLongPress = { onToggleNegative(photo.imageId) },
+                        showSimilarityBadge = true
                     )
                 }
             }
 
-            // Section: Other Photos
-            val otherPhotos = rankedPhotos.drop(30)
-            if (otherPhotos.isNotEmpty()) {
+            // Section: Maybe (<45%)
+            if (maybe.isNotEmpty()) {
                 item(span = { GridItemSpan(3) }) {
                     SectionHeader(
                         icon = Icons.Default.Photo,
-                        text = "📷 Other Photos (${otherPhotos.size})",
-                        color = MaterialTheme.colorScheme.onSurfaceVariant
+                        text = "🟠 Maybe (${maybe.size})",
+                        color = Color(0xFFFF9800)
                     )
                 }
-
-                items(otherPhotos, key = { it.imageId }) { photo ->
+                items(maybe, key = { it.imageId }) { photo ->
                     PhotoCard(
                         photo = photo,
                         isSelected = photo.imageId in selectedImageIds,
-                        onToggle = { onToggleSelection(photo.imageId) }
+                        isNegative = photo.imageId in negativeImageIds,
+                        onToggle = { onToggleSelection(photo.imageId) },
+                        onLongPress = { onToggleNegative(photo.imageId) }
                     )
                 }
             }
@@ -258,24 +280,34 @@ private fun RollingScanPhotoGrid(
 }
 
 // ═══════════════════════════════════════════════════════════
-// PHOTO CARD
+// PHOTO CARD - with long-press for negative marking
 // ═══════════════════════════════════════════════════════════
 
+@OptIn(ExperimentalFoundationApi::class)
 @Composable
 private fun PhotoCard(
     photo: FaceSimilarityScorer.ScoredPhoto,
     isSelected: Boolean,
+    isNegative: Boolean = false,
     onToggle: () -> Unit,
+    onLongPress: () -> Unit = {},
     showSimilarityBadge: Boolean = false
 ) {
+    val borderColor = when {
+        isNegative -> Color(0xFFE53935)  // Red for negative
+        isSelected -> MaterialTheme.colorScheme.primary
+        else -> MaterialTheme.colorScheme.outline.copy(alpha = 0.3f)
+    }
+    val borderWidth = if (isSelected || isNegative) 3.dp else 1.dp
+
     Card(
         modifier = Modifier
             .aspectRatio(1f)
-            .clickable(onClick = onToggle),
-        border = if (isSelected)
-            BorderStroke(3.dp, MaterialTheme.colorScheme.primary)
-        else
-            BorderStroke(1.dp, MaterialTheme.colorScheme.outline.copy(alpha = 0.3f)),
+            .combinedClickable(
+                onClick = onToggle,
+                onLongClick = onLongPress
+            ),
+        border = BorderStroke(borderWidth, borderColor),
         elevation = CardDefaults.cardElevation(
             defaultElevation = if (isSelected) 4.dp else 1.dp
         )
@@ -289,22 +321,47 @@ private fun PhotoCard(
                 contentScale = ContentScale.Crop
             )
 
-            // Similarity badge (top-left) - Only for top matches
-            if (showSimilarityBadge) {
+            // Dim overlay for negatives
+            if (isNegative) {
+                Box(
+                    modifier = Modifier
+                        .fillMaxSize()
+                        .padding(0.dp),
+                    contentAlignment = Alignment.Center
+                ) {
+                    Surface(
+                        modifier = Modifier.fillMaxSize(),
+                        color = Color.Black.copy(alpha = 0.5f)
+                    ) {}
+                    Icon(
+                        Icons.Default.Close,
+                        contentDescription = "Not this person",
+                        tint = Color.White,
+                        modifier = Modifier.size(32.dp)
+                    )
+                }
+            }
+
+            // Similarity badge (top-left)
+            if (showSimilarityBadge && !isNegative) {
                 Surface(
                     modifier = Modifier
                         .align(Alignment.TopStart)
                         .padding(6.dp),
                     shape = RoundedCornerShape(8.dp),
-                    color = MaterialTheme.colorScheme.primary,
+                    color = when {
+                        photo.finalScore >= 0.60f -> Color(0xFF4CAF50)
+                        photo.finalScore >= 0.45f -> Color(0xFFFFC107)
+                        else -> Color(0xFFFF9800)
+                    },
                     shadowElevation = 4.dp
                 ) {
                     Text(
-                        text = "${(photo.similarityScore * 100).toInt()}%",
+                        text = "${(photo.finalScore * 100).toInt()}%",
                         modifier = Modifier.padding(horizontal = 8.dp, vertical = 4.dp),
                         style = MaterialTheme.typography.labelSmall,
                         fontWeight = FontWeight.Bold,
-                        color = MaterialTheme.colorScheme.onPrimary
+                        color = Color.White
                     )
                 }
             }
@@ -332,7 +389,7 @@ private fun PhotoCard(
             }
 
             // Face count badge (bottom-right)
-            if (photo.faceCount > 1) {
+            if (photo.faceCount > 1 && !isNegative) {
                 Surface(
                     modifier = Modifier
                         .align(Alignment.BottomEnd)
@@ -395,6 +452,7 @@ private fun RollingScanBottomBar(
     isReadyForTraining: Boolean,
     validationMessage: String?,
     onSelectTopN: (Int) -> Unit,
+    onSelectAboveThreshold: (Float) -> Unit,
     onSubmit: () -> Unit
 ) {
     Surface(
@@ -416,39 +474,49 @@ private fun RollingScanBottomBar(
                 )
             }
 
+            // First row: threshold selection
             Row(
                 modifier = Modifier.fillMaxWidth(),
-                horizontalArrangement = Arrangement.spacedBy(8.dp)
+                horizontalArrangement = Arrangement.spacedBy(6.dp)
             ) {
-                // Quick select buttons
                 OutlinedButton(
-                    onClick = { onSelectTopN(10) },
-                    modifier = Modifier.weight(1f)
+                    onClick = { onSelectAboveThreshold(0.60f) },
+                    modifier = Modifier.weight(1f),
+                    contentPadding = PaddingValues(horizontal = 8.dp, vertical = 4.dp)
                 ) {
-                    Text("Top 10")
+                    Text(">60%", style = MaterialTheme.typography.labelSmall)
                 }
-
                 OutlinedButton(
-                    onClick = { onSelectTopN(20) },
-                    modifier = Modifier.weight(1f)
+                    onClick = { onSelectAboveThreshold(0.50f) },
+                    modifier = Modifier.weight(1f),
+                    contentPadding = PaddingValues(horizontal = 8.dp, vertical = 4.dp)
                 ) {
-                    Text("Top 20")
+                    Text(">50%", style = MaterialTheme.typography.labelSmall)
                 }
+                OutlinedButton(
+                    onClick = { onSelectTopN(15) },
+                    modifier = Modifier.weight(1f),
+                    contentPadding = PaddingValues(horizontal = 8.dp, vertical = 4.dp)
+                ) {
+                    Text("Top 15", style = MaterialTheme.typography.labelSmall)
+                }
+            }
 
-                // Submit button
-                Button(
-                    onClick = onSubmit,
-                    enabled = isReadyForTraining,
-                    modifier = Modifier.weight(1.5f)
-                ) {
-                    Icon(
-                        Icons.Default.Done,
-                        contentDescription = null,
-                        modifier = Modifier.size(18.dp)
-                    )
-                    Spacer(Modifier.width(8.dp))
-                    Text("Train ($selectedCount)")
-                }
+            Spacer(Modifier.height(8.dp))
+
+            // Second row: submit
+            Button(
+                onClick = onSubmit,
+                enabled = isReadyForTraining,
+                modifier = Modifier.fillMaxWidth()
+            ) {
+                Icon(
+                    Icons.Default.Done,
+                    contentDescription = null,
+                    modifier = Modifier.size(18.dp)
+                )
+                Spacer(Modifier.width(8.dp))
+                Text("Train Model ($selectedCount photos)")
             }
         }
     }

app/src/main/java/com/placeholder/sherpai2/ui/rollingscan/Rollingscanviewmodel.kt

@@ -44,6 +44,11 @@ class RollingScanViewModel @Inject constructor(
         private const val TAG = "RollingScanVM"
         private const val DEBOUNCE_DELAY_MS = 300L
         private const val MIN_PHOTOS_FOR_TRAINING = 15
+
+        // Progressive thresholds based on selection count
+        private const val FLOOR_FEW_SEEDS = 0.30f      // 1-3 seeds
+        private const val FLOOR_MEDIUM_SEEDS = 0.40f   // 4-10 seeds
+        private const val FLOOR_MANY_SEEDS = 0.50f     // 10+ seeds
     }
 
     // ═══════════════════════════════════════════════════════════
@@ -71,6 +76,11 @@ class RollingScanViewModel @Inject constructor(
     // Cache of selected embeddings
     private val selectedEmbeddings = mutableListOf<FloatArray>()
 
+    // Negative embeddings (marked as "not this person")
+    private val _negativeImageIds = MutableStateFlow<Set<String>>(emptySet())
+    val negativeImageIds: StateFlow<Set<String>> = _negativeImageIds.asStateFlow()
+    private val negativeEmbeddings = mutableListOf<FloatArray>()
+
     // All available image IDs
     private var allImageIds: List<String> = emptyList()
 
@@ -156,24 +166,55 @@ class RollingScanViewModel @Inject constructor(
             current.remove(imageId)
 
             viewModelScope.launch {
-                // Remove embedding from cache
                 val cached = faceCacheDao.getEmbeddingByImageId(imageId)
                 cached?.getEmbedding()?.let { selectedEmbeddings.remove(it) }
             }
         } else {
-            // Select
+            // Select (and remove from negatives if present)
             current.add(imageId)
+            if (imageId in _negativeImageIds.value) {
+                toggleNegative(imageId)
+            }
 
             viewModelScope.launch {
-                // Add embedding to cache
                 val cached = faceCacheDao.getEmbeddingByImageId(imageId)
                 cached?.getEmbedding()?.let { selectedEmbeddings.add(it) }
             }
         }
 
-        _selectedImageIds.value = current
+        _selectedImageIds.value = current.toSet()  // Immutable copy
+
+        scanDebouncer.debounce {
+            triggerRollingScan()
+        }
+    }
+
+    /**
+     * Toggle negative marking ("Not this person")
+     */
+    fun toggleNegative(imageId: String) {
+        val current = _negativeImageIds.value.toMutableSet()
+
+        if (imageId in current) {
+            current.remove(imageId)
+            viewModelScope.launch {
+                val cached = faceCacheDao.getEmbeddingByImageId(imageId)
+                cached?.getEmbedding()?.let { negativeEmbeddings.remove(it) }
+            }
+        } else {
+            current.add(imageId)
+            // Remove from selected if present
+            if (imageId in _selectedImageIds.value) {
+                toggleSelection(imageId)
+            }
+            viewModelScope.launch {
+                val cached = faceCacheDao.getEmbeddingByImageId(imageId)
+                cached?.getEmbedding()?.let { negativeEmbeddings.add(it) }
+            }
+        }
+
+        _negativeImageIds.value = current.toSet()  // Immutable copy
 
-        // Debounced rescan
         scanDebouncer.debounce {
             triggerRollingScan()
         }
@@ -190,13 +231,33 @@ class RollingScanViewModel @Inject constructor(
 
         val current = _selectedImageIds.value.toMutableSet()
         current.addAll(topPhotos)
-        _selectedImageIds.value = current
+        _selectedImageIds.value = current.toSet()  // Immutable copy
 
         viewModelScope.launch {
-            // Add embeddings
             val embeddings = faceCacheDao.getEmbeddingsForImages(topPhotos.toList())
             selectedEmbeddings.addAll(embeddings.mapNotNull { it.getEmbedding() })
+            triggerRollingScan()
+        }
+    }
 
+    /**
+     * Select all photos above a similarity threshold
+     */
+    fun selectAllAboveThreshold(threshold: Float) {
+        val photosAbove = _rankedPhotos.value
+            .filter { it.finalScore >= threshold }
+            .map { it.imageId }
+
+        val current = _selectedImageIds.value.toMutableSet()
+        current.addAll(photosAbove)
+        _selectedImageIds.value = current.toSet()  // Immutable copy
+
+        viewModelScope.launch {
+            val newIds = photosAbove.filter { it !in _selectedImageIds.value }
+            if (newIds.isNotEmpty()) {
+                val embeddings = faceCacheDao.getEmbeddingsForImages(newIds)
+                selectedEmbeddings.addAll(embeddings.mapNotNull { it.getEmbedding() })
+            }
             triggerRollingScan()
         }
     }
@@ -207,17 +268,24 @@ class RollingScanViewModel @Inject constructor(
     fun clearSelection() {
         _selectedImageIds.value = emptySet()
         selectedEmbeddings.clear()
-
-        // Reset ranking
         _rankedPhotos.value = emptyList()
     }
 
+    /**
+     * Clear negative markings
+     */
+    fun clearNegatives() {
+        _negativeImageIds.value = emptySet()
+        negativeEmbeddings.clear()
+        scanDebouncer.debounce { triggerRollingScan() }
+    }
+
     // ═══════════════════════════════════════════════════════════
     // ROLLING SCAN LOGIC
     // ═══════════════════════════════════════════════════════════
 
     /**
-     * CORE: Trigger rolling similarity scan
+     * CORE: Trigger rolling similarity scan with progressive filtering
      */
     private suspend fun triggerRollingScan() {
         if (selectedEmbeddings.isEmpty()) {
@@ -228,7 +296,15 @@ class RollingScanViewModel @Inject constructor(
         try {
             _isScanning.value = true
 
-            Log.d(TAG, "Starting scan with ${selectedEmbeddings.size} selected embeddings")
+            val selectionCount = selectedEmbeddings.size
+            Log.d(TAG, "Starting scan with $selectionCount selected, ${negativeEmbeddings.size} negative")
+
+            // Progressive threshold based on selection count
+            val similarityFloor = when {
+                selectionCount <= 3 -> FLOOR_FEW_SEEDS
+                selectionCount <= 10 -> FLOOR_MEDIUM_SEEDS
+                else -> FLOOR_MANY_SEEDS
+            }
 
             // Calculate centroid from selected embeddings
             val centroid = faceSimilarityScorer.calculateCentroid(selectedEmbeddings)
@@ -240,17 +316,38 @@ class RollingScanViewModel @Inject constructor(
                 centroid = centroid
             )
 
-            // Update image URIs in scored photos
-            val photosWithUris = scoredPhotos.map { photo ->
-                photo.copy(
-                    imageUri = imageUriCache[photo.imageId] ?: photo.imageId
-                )
-            }
+            // Apply negative penalty, quality boost, and floor filter
+            val filteredPhotos = scoredPhotos
+                .map { photo ->
+                    // Calculate max similarity to any negative embedding
+                    val negativePenalty = if (negativeEmbeddings.isNotEmpty()) {
+                        negativeEmbeddings.maxOfOrNull { neg ->
+                            cosineSimilarity(photo.cachedEmbedding, neg)
+                        } ?: 0f
+                    } else 0f
 
-            Log.d(TAG, "Scan complete. Scored ${photosWithUris.size} photos")
+                    // Quality multiplier: solo face, large face, good quality
+                    val qualityMultiplier = 1f +
+                        (if (photo.faceCount == 1) 0.15f else 0f) +
+                        (if (photo.faceAreaRatio > 0.15f) 0.10f else 0f) +
+                        (if (photo.qualityScore > 0.7f) 0.10f else 0f)
 
-            // Update ranked list
-            _rankedPhotos.value = photosWithUris
+                    // Final score = (similarity - negativePenalty) * qualityMultiplier
+                    val adjustedScore = ((photo.similarityScore - negativePenalty * 0.5f) * qualityMultiplier)
+                        .coerceIn(0f, 1f)
+
+                    photo.copy(
+                        imageUri = imageUriCache[photo.imageId] ?: photo.imageId,
+                        finalScore = adjustedScore
+                    )
+                }
+                .filter { it.finalScore >= similarityFloor }  // Apply floor
+                .filter { it.imageId !in _negativeImageIds.value }  // Hide negatives
+                .sortedByDescending { it.finalScore }
+
+            Log.d(TAG, "Scan complete. ${filteredPhotos.size} photos above floor $similarityFloor")
+
+            _rankedPhotos.value = filteredPhotos
 
         } catch (e: Exception) {
             Log.e(TAG, "Scan failed", e)
@@ -259,6 +356,19 @@ class RollingScanViewModel @Inject constructor(
         }
     }
 
+    private fun cosineSimilarity(a: FloatArray, b: FloatArray): Float {
+        if (a.size != b.size) return 0f
+        var dot = 0f
+        var normA = 0f
+        var normB = 0f
+        for (i in a.indices) {
+            dot += a[i] * b[i]
+            normA += a[i] * a[i]
+            normB += b[i] * b[i]
+        }
+        return if (normA > 0 && normB > 0) dot / (kotlin.math.sqrt(normA) * kotlin.math.sqrt(normB)) else 0f
+    }
+
     // ═══════════════════════════════════════════════════════════
     // SUBMISSION
     // ═══════════════════════════════════════════════════════════
@@ -299,9 +409,11 @@ class RollingScanViewModel @Inject constructor(
     fun reset() {
         _uiState.value = RollingScanState.Idle
         _selectedImageIds.value = emptySet()
+        _negativeImageIds.value = emptySet()
         _rankedPhotos.value = emptyList()
         _isScanning.value = false
         selectedEmbeddings.clear()
+        negativeEmbeddings.clear()
         allImageIds = emptyList()
         imageUriCache = emptyMap()
         scanDebouncer.cancel()

app/src/main/java/com/placeholder/sherpai2/ui/trainingprep/Facedetectionhelper.kt

@@ -67,13 +67,14 @@ class FaceDetectionHelper(private val context: Context) {
                 val inputImage = InputImage.fromBitmap(bitmap, 0)
                 val faces = detector.process(inputImage).await()
 
-                // Filter to quality faces only
+                // Filter to quality faces - use lenient scanning filter
+                // (Discovery filter was too strict, rejecting faces from rolling scan)
                 val qualityFaces = faces.filter { face ->
-                    FaceQualityFilter.validateForDiscovery(
+                    FaceQualityFilter.validateForScanning(
                         face = face,
                         imageWidth = bitmap.width,
                         imageHeight = bitmap.height
-                    ).isValid
+                    )
                 }
 
                 // Sort by face size (area) to get the largest quality face

app/src/main/java/com/placeholder/sherpai2/ui/trainingprep/TrainViewModel.kt

@@ -192,11 +192,10 @@ class TrainViewModel @Inject constructor(
                     .first()
 
                 if (backgroundTaggingEnabled) {
-                    // Lower threshold (0.55) since we use multi-centroid matching
+                    // Use default threshold (0.62 solo, 0.68 group)
                     val scanRequest = LibraryScanWorker.createWorkRequest(
                         personId = personId,
-                        personName = personName,
-                        threshold = 0.55f
+                        personName = personName
                     )
                     workManager.enqueue(scanRequest)
                 }
@@ -382,7 +381,7 @@ class TrainViewModel @Inject constructor(
             faceDetectionResults = updatedFaceResults,
             validationErrors = updatedErrors,
             validImagesWithFaces = updatedValidImages,
-            excludedImages = excludedImages
+            excludedImages = excludedImages.toSet()  // Immutable copy for Compose state detection
         )
     }
 

app/src/main/java/com/placeholder/sherpai2/workers/Libraryscanworker.kt

@@ -9,6 +9,7 @@ import com.google.mlkit.vision.common.InputImage
 import com.google.mlkit.vision.face.FaceDetection
 import com.google.mlkit.vision.face.FaceDetectorOptions
 import com.placeholder.sherpai2.data.local.dao.FaceModelDao
+import com.placeholder.sherpai2.data.local.dao.PersonDao
 import com.placeholder.sherpai2.domain.clustering.FaceQualityFilter
 import com.placeholder.sherpai2.ml.FaceNormalizer
 import com.placeholder.sherpai2.data.local.dao.ImageDao
@@ -54,7 +55,8 @@ class LibraryScanWorker @AssistedInject constructor(
     @Assisted workerParams: WorkerParameters,
     private val imageDao: ImageDao,
     private val faceModelDao: FaceModelDao,
-    private val photoFaceTagDao: PhotoFaceTagDao
+    private val photoFaceTagDao: PhotoFaceTagDao,
+    private val personDao: PersonDao
 ) : CoroutineWorker(context, workerParams) {
 
     companion object {
@@ -67,7 +69,8 @@ class LibraryScanWorker @AssistedInject constructor(
         const val KEY_MATCHES_FOUND = "matches_found"
         const val KEY_PHOTOS_SCANNED = "photos_scanned"
 
-        private const val DEFAULT_THRESHOLD = 0.70f  // Slightly looser than validation
+        private const val DEFAULT_THRESHOLD = 0.62f  // Solo photos
+        private const val GROUP_THRESHOLD = 0.68f   // Group photos (stricter)
         private const val BATCH_SIZE = 20
         private const val MAX_RETRIES = 3
 
@@ -139,21 +142,40 @@ class LibraryScanWorker @AssistedInject constructor(
                 )
             }
 
+            // Step 2.5: Load person to check isChild flag
+            val person = withContext(Dispatchers.IO) {
+                personDao.getPersonById(personId)
+            }
+            val isChildTarget = person?.isChild ?: false
+
             // Step 3: Initialize ML components
             val faceNetModel = FaceNetModel(context)
             val detector = FaceDetection.getClient(
                 FaceDetectorOptions.Builder()
                     .setPerformanceMode(FaceDetectorOptions.PERFORMANCE_MODE_ACCURATE)
+                    .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_ALL)  // Needed for age estimation
                     .setMinFaceSize(0.15f)
                     .build()
             )
 
+            // Distribution-based minimum threshold (self-calibrating)
+            val distributionMin = (faceModel.averageConfidence - 2 * faceModel.similarityStdDev)
+                .coerceAtLeast(faceModel.similarityMin - 0.05f)
+                .coerceAtLeast(0.50f)  // Never go below 0.50 absolute floor
+
             // Get ALL centroids for multi-centroid matching (critical for children)
             val modelCentroids = faceModel.getCentroids().map { it.getEmbeddingArray() }
             if (modelCentroids.isEmpty()) {
                 return@withContext Result.failure(workDataOf("error" to "No centroids in model"))
             }
 
+            // Load ALL other models for "best match wins" comparison
+            // This prevents tagging siblings incorrectly
+            val allModels = withContext(Dispatchers.IO) { faceModelDao.getAllActiveFaceModels() }
+            val otherModelCentroids = allModels
+                .filter { it.id != faceModel.id }
+                .map { model -> model.id to model.getCentroids().map { it.getEmbeddingArray() } }
+
             var matchesFound = 0
             var photosScanned = 0
 
@@ -172,9 +194,12 @@ class LibraryScanWorker @AssistedInject constructor(
                                 personId = personId,
                                 faceModelId = faceModel.id,
                                 modelCentroids = modelCentroids,
+                                otherModelCentroids = otherModelCentroids,
                                 faceNetModel = faceNetModel,
                                 detector = detector,
-                                threshold = threshold
+                                threshold = threshold,
+                                distributionMin = distributionMin,
+                                isChildTarget = isChildTarget
                             )
 
                             if (tags.isNotEmpty()) {
@@ -236,9 +261,12 @@ class LibraryScanWorker @AssistedInject constructor(
         personId: String,
         faceModelId: String,
         modelCentroids: List<FloatArray>,
+        otherModelCentroids: List<Pair<String, List<FloatArray>>>,
         faceNetModel: FaceNetModel,
         detector: com.google.mlkit.vision.face.FaceDetector,
-        threshold: Float
+        threshold: Float,
+        distributionMin: Float,
+        isChildTarget: Boolean
     ): List<PhotoFaceTagEntity> = withContext(Dispatchers.IO) {
 
         try {
@@ -250,45 +278,94 @@ class LibraryScanWorker @AssistedInject constructor(
             val inputImage = InputImage.fromBitmap(bitmap, 0)
             val faces = detector.process(inputImage).await()
 
+            if (faces.isEmpty()) {
+                bitmap.recycle()
+                return@withContext emptyList()
+            }
+
+            // Use higher threshold for group photos
+            val isGroupPhoto = faces.size > 1
+            val effectiveThreshold = if (isGroupPhoto) GROUP_THRESHOLD else threshold
+
+            // Track best match (only tag ONE face per image to avoid false positives)
+            var bestMatch: PhotoFaceTagEntity? = null
+            var bestSimilarity = 0f
+
             // Check each face (filter by quality first)
-            val tags = faces.mapNotNull { face ->
+            for (face in faces) {
                 // Quality check
                 if (!FaceQualityFilter.validateForScanning(face, bitmap.width, bitmap.height)) {
-                    return@mapNotNull null
+                    continue
+                }
+
+                // Skip very small faces
+                val faceArea = face.boundingBox.width() * face.boundingBox.height()
+                val imageArea = bitmap.width * bitmap.height
+                if (faceArea.toFloat() / imageArea < 0.02f) continue
+
+                // SIGNAL 2: Age plausibility check (if target is a child)
+                if (isChildTarget) {
+                    val ageGroup = FaceQualityFilter.estimateAgeGroup(face, bitmap.width, bitmap.height)
+                    if (ageGroup == FaceQualityFilter.AgeGroup.ADULT) {
+                        continue  // Reject clearly adult faces when searching for a child
+                    }
                 }
 
                 try {
                     // Crop and normalize face for best recognition
                     val faceBitmap = FaceNormalizer.cropAndNormalize(bitmap, face)
-                        ?: return@mapNotNull null
+                        ?: continue
 
                     // Generate embedding
                     val faceEmbedding = faceNetModel.generateEmbedding(faceBitmap)
                     faceBitmap.recycle()
 
-                    // Match against ALL centroids, use best match (critical for children)
-                    val similarity = modelCentroids.maxOfOrNull { centroid ->
+                    // Match against target person's centroids
+                    val targetSimilarity = modelCentroids.maxOfOrNull { centroid ->
                         faceNetModel.calculateSimilarity(faceEmbedding, centroid)
                     } ?: 0f
 
-                    if (similarity >= threshold) {
-                        PhotoFaceTagEntity.create(
+                    // SIGNAL 1: Distribution-based rejection
+                    // If similarity is below (mean - 2*stdDev) or (min - 0.05), it's a statistical outlier
+                    if (targetSimilarity < distributionMin) {
+                        continue  // Too far below training distribution
+                    }
+
+                    // SIGNAL 3: Basic threshold check
+                    if (targetSimilarity < effectiveThreshold) {
+                        continue
+                    }
+
+                    // SIGNAL 4: "Best match wins" - check if any OTHER model scores higher
+                    // This prevents tagging siblings incorrectly
+                    val bestOtherSimilarity = otherModelCentroids.maxOfOrNull { (_, centroids) ->
+                        centroids.maxOfOrNull { centroid ->
+                            faceNetModel.calculateSimilarity(faceEmbedding, centroid)
+                        } ?: 0f
+                    } ?: 0f
+
+                    val isTargetBestMatch = targetSimilarity > bestOtherSimilarity
+
+                    // All signals must pass
+                    if (isTargetBestMatch && targetSimilarity > bestSimilarity) {
+                        bestSimilarity = targetSimilarity
+                        bestMatch = PhotoFaceTagEntity.create(
                             imageId = photo.imageId,
                             faceModelId = faceModelId,
                             boundingBox = face.boundingBox,
-                            confidence = similarity,
+                            confidence = targetSimilarity,
                             faceEmbedding = faceEmbedding
                         )
-                    } else {
-                        null
                     }
                 } catch (e: Exception) {
-                    null
+                    // Skip this face
                 }
             }
 
             bitmap.recycle()
-            tags
+
+            // Return only the best match (or empty)
+            if (bestMatch != null) listOf(bestMatch) else emptyList()
 
         } catch (e: Exception) {
             emptyList()