Task 319 analysis: rule partially cracked (192/267)

medal-solvers/TASK319_ANALYSIS.md

@@ -0,0 +1,38 @@
+# Task 319 — Pattern Matching Analysis
+
+## Rule (Partially Cracked: 192/267 = 72%)
+
+### Input Structure
+- Background (most common color) + 3 colored objects
+- Input grid: variable size (15-20 × 15-20), output: one-hot [1,10,30,30]
+
+### Pattern
+1. One object is a "template" (has repeated row/col structure)
+2. Template has groups of identical consecutive rows and columns
+3. Downsample template by taking one representative from each row/col group
+4. The downsampled pattern appears as a sub-pattern of one of the other two objects
+5. Output = that matching object (its full binary pattern with its color)
+
+### Template Identification
+- Template is NOT always the largest object by area
+- Template is the object whose downsampled version matches a sub-region of another object
+- Template must have at least some row or column group > 1 (compression)
+
+### Known Issues (75 failures)
+- Some templates have non-uniform row/col groups (e.g., [1,2,1,1] or [2,4,2,2])
+- The "group identical consecutive rows" approach fails when:
+  - Rows are similar but not identical
+  - The scaling is based on a different criterion than identical rows
+  - The template matching requires considering transposition/rotation
+  
+### Current Score
+- 2709 nodes, 26M memory, 8943 params
+- Score: 7.92 points
+- Need: ~12.45 points (memory+params < 282K)
+
+### ONNX Rebuild Approach (if rule is fully cracked)
+- Small model: just need object segmentation + pattern comparison
+- Key ops: ReduceSum for object detection, comparison for matching
+- If rule were simple: ~50-100 nodes possible → score 13+
+- But complex rule (3 candidates × variable downsampling × sub-pattern matching) 
+  makes it hard to fit in fixed ONNX without loops