# System Architecture

## Piece Recognition Pipeline
1. Board Detection
   - VNDetectRectanglesRequest for board location
   - Aspect ratio and size validation
   - Square extraction with equal dimensions

2. Image Preprocessing
   - Contrast and brightness adjustment
   - Unsharp mask for edge enhancement
   - Consistent image scaling

3. ML Classification
   - CoreML model prediction
   - Confidence score analysis
   - Position-based adjustments

## Recognition Patterns

### Square Classification
1. Empty Square Detection
   - Exact empty_dark/empty_light matching
   - High confidence threshold (>0.9)
   - Early detection and return

2. Piece Recognition
   - Strict label format validation
   - Position-based confidence adjustment
   - Piece count tracking

3. Error Prevention
   - Empty square validation
   - Label format checking
   - Position rule enforcement

### Classification Flow
1. Input Validation
   - Image dimensions check
   - Model availability check
   - Configuration setup

2. Square Analysis
   - Empty square check first
   - Piece classification second
   - Position validation last

3. Confidence Checks
   - Empty squares: >0.9
   - Pieces: >0.98 with >5.0 ratio
   - Position adjustments

4. Error Handling
   - Clear error messages
   - Detailed logging
   - Safe fallbacks

## Validation Patterns
1. Piece Count Rules
   - Maximum 1: king, queen
   - Maximum 2: rooks, bishops, knights
   - Maximum 8: pawns
   - Track by color and type

2. Position Rules
   - Kings: not on opponent's back rank
   - Pawns: no backward movement
   - All pieces: within board bounds
   - All pieces: valid movement patterns

3. Piece Tracking
   - Maximum piece counts
   - Color-specific tracking
   - Total position validation
   - Captured piece limits

4. Recognition Flow
   - Empty square detection first
   - Piece classification second
   - Position validation last
   - Clear error reporting