Add V37 analysis: full scoring audit, task 77 partial solve (211/266), strategy recommendations

medal-solvers/ANALYSIS_V37.md

@@ -0,0 +1,93 @@
+# Analysis for V37+ — Finding +38.45 pts to reach 6100
+
+## Current: V36 = 6061.55 | Target: 6100 | Gap: +38.45
+
+## Complete Task Score Audit (base submission-6043)
+
+### Top Priority Targets (score < 12, not yet optimized)
+
+| Task | Score | Nodes | Params | Memory | Difficulty | Rule Summary |
+|------|-------|-------|--------|--------|------------|--------------|
+| 366 | 7.08 | 12512 | 51,979 | 60.4M | VERY HIGH | Shape placement at dot positions (backtracking) |
+| 219 | 8.42 | 532 | 3,448 | 15.8M | HIGH | Staircase continuation |
+| 133 | 8.95 | 368 | 228,894 | 9.1M | HIGH | Unknown (v90 failed arc-gen) |
+| 157 | 9.14 | 433 | 202,281 | 7.5M | HIGH | Reflect shapes + pattern matching |
+| 158 | 9.30 | 822 | 765 | 6.6M | HIGH | Shape copy + bridges |
+| 233 | 9.60 | 1372 | 142,960 | 4.8M | HIGH | Extract subgrid + fill holes |
+| 209 | 10.11 | 4286 | 138,262 | 2.8M | HIGH | Template-based block placement |
+| 44 | 10.32 | 1054 | 1,415 | 2.4M | HIGH | Fill rect holes with patches |
+| 200 | 10.33 | 91 | 198,010 | 2.2M | MEDIUM | Stripe pattern from seed position |
+| 76 | 10.34 | 220 | 229,498 | 2.1M | HIGH | Complex pattern extension |
+| 118 | 10.42 | 1163 | 1,887 | 2.2M | HIGH | Mark cells between color-2 markers |
+| 367 | 10.58 | 694 | 59,173 | 1.8M | MEDIUM | Conv-based pattern |
+
+### Medium Priority (score 11-13, simpler rules possible)
+
+| Task | Score | Nodes | % Changed | Key Property |
+|------|-------|-------|-----------|--------------|
+| 77 | 12.83 | 72 | 5.1% | **211/266 SOLVED** - bbox fill around 8-conn 2-clusters |
+| 173 | 12.89 | 69 | 3.6% | Very few changes, tiny model |
+| 54 | 12.89 | 118 | 8.6% | |
+| 182 | 11.45 | 522 | 2.9% | Very few changes, Conv-based |
+| 191 | 11.53 | 1037 | 5.8% | Only 165 params! Pure compute |
+| 18 | 11.32 | 219 | 8.1% | |
+
+## Task 77 — Partially Solved (211/266)
+
+**Rule**: For each 8-connected cluster of color-2 cells (≥2 cells), compute bounding box. Fill all fg-color cells within bbox with color 4.
+
+**Failing cases** (55/266): When 2-cells are separated by fg cells but should be treated as one group. The correct rule likely involves:
+- Option A: Find 2-clusters connected through exactly 1 fg cell (bridge detection)
+- Option B: Use a different connectivity criterion for grouping
+- Option C: Iterative expansion of 2-cluster bboxes
+
+**If fully solved**: Score ≈ 16+ pts (only need small model with 8-conn MaxPool + bbox ops). Gain: ~+3 pts.
+
+## Scoring Formula Reference
+
+```
+score = max(1.0, 25.0 - ln(memory + params))
+```
+
+| Target Score | Max (memory+params) | Reduction from 10pts base |
+|-------------|--------------------|----|
+| 15 pts | 22,026 | 2.7× |
+| 16 pts | 8,103 | 7.4× |
+| 17 pts | 2,981 | 20× |
+| 18 pts | 1,097 | 55× |
+| 20 pts | 148 | 405× |
+
+## Key Technical Insights
+
+### What reduces memory (from onnxruntime-perf.txt analysis)
+1. **Fewer MaxPool iterations**: Each adds [1,1,30,30]×4 bytes = 3600 bytes to memory
+2. **Slice to active area first**: 10×10 grid → 9× less memory per tensor vs 30×30
+3. **Float16 for intermediates**: 2 bytes/element vs 4 (max gain = ln(2) = 0.69 pts)
+4. **Integer types (uint8)**: 1 byte/element (max gain = ln(4) = 1.39 pts)
+5. **Avoid broadcast**: [1,C,H,W] operations create huge intermediates
+
+### What doesn't help
+- Graph optimization (Kaggle uses ORT_DISABLE_ALL)
+- Initializer dtype (params counts elements, not bytes)
+- Removing value_info (scorer runs shape_inference)
+
+## Recommended Strategy
+
+### Path 1: Complete 6-8 new tasks at +5 avg each = +30-40 pts
+Focus on tasks where Python solver exists (in own-solver/) and can be converted to compact ONNX.
+
+### Path 2: Improve existing + a few new tasks
+- Improve task 285 from 8.82→10+ (reduce MaxPool iterations): +1-2 pts
+- Complete task 77 (if 266/266): +3 pts  
+- Build 5-6 new compact solvers: +25-30 pts
+
+### Path 3: Systematic wave22 conversion
+The wave22.py file has 266/266 solvers for many tasks. These are already working Python implementations. Converting the simplest ones to compact ONNX could be highly productive.
+
+## Files Modified
+- None (analysis only)
+
+## Next Steps
+1. Complete Task 77 rule (test bridge connectivity between 2-clusters)
+2. Look at wave22 solvers for tasks in our target list
+3. Profile existing optimized models to verify scores