Add bulk scan results — 35 fixed-size actionable targets identified

medal-solvers/BULK_SCAN.md

@@ -0,0 +1,74 @@
+# Bulk Scan Results — All 400 Tasks Profiled
+
+## Summary
+- **400 tasks scanned** — grid sizes, base model nodes/params/memory, score estimates
+- **204 fixed-size** tasks, **196 variable-size**
+- **35 actionable fixed-size targets** (base < 13.5, not already done, not known non-viable)
+- **10 variable-size targets** with base < 11 (harder but huge gains)
+
+## TOP PRIORITY TARGETS (fixed-size, ordered by potential gain)
+
+| Task | Grid | Nodes | Params | Base Score | Potential |
+|------|------|-------|--------|-----------|-----------|
+| **219** | 15×10 | 532 | 3,448 | 8.42 | **+6.1** |
+| **157** | 10×15 | 433 | 202,281 | 9.14 | **+5.4** |
+| **182** | 20×20 | 522 | 946 | 10.63 | **+3.9** |
+| 110 | 29×29 | 163 | 1,185 | 12.11 | +2.4 |
+| 54 | 30×30 | 118 | 161 | 12.15 | +2.4 |
+| **398** | 1×5 | 44 | 770 | 12.36 | **+2.1** |
+| **34** | 9×9 | 87 | 2,847 | 12.39 | **+2.1** |
+| 71 | 16×16 | 49 | 930 | 12.53 | +2.0 |
+| 5 | 21×21 | 106 | 336 | 12.62 | +1.9 |
+| **384** | 9×9 | 39 | 135 | 12.70 | **+1.8** |
+| 216 | 20×20 | 286 | 100 | 12.76 | +1.7 |
+| 128 | 15×15 | 28 | 102 | 12.78 | +1.7 |
+| 208 | 21×21 | 196 | 522 | 12.82 | +1.7 |
+| **217** | 9×9 | 34 | 1,935 | 12.82 | **+1.7** |
+| 201 | 13×13 | 135 | 932 | 12.90 | +1.6 |
+| 365 | 10×10 | 242 | 71 | 12.91 | +1.6 |
+| 165 | 20×20 | 25 | 994 | 13.02 | +1.5 |
+| 36 | 30×30 | 56 | 159 | 13.03 | +1.5 |
+| **57** | 8×8 | 18 | 11 | 13.08 | **+1.4** |
+| 70 | 17×17 | 71 | 78 | 13.20 | +1.3 |
+| 351 | 16×16 | 25 | 40 | 13.20 | +1.3 |
+| 400 | 24×24 | 25 | 39 | 13.20 | +1.3 |
+| 22 | 11×11 | 363 | 121 | 13.21 | +1.3 |
+| 11 | 11×11 | 52 | 2,331 | 13.22 | +1.3 |
+| 162 | 20×20 | 19 | 43 | 13.22 | +1.3 |
+| 33 | 17×17 | 16 | 49 | 13.22 | +1.3 |
+| 107 | 5×5 | 52 | 13,681 | 13.24 | +1.3 |
+| 17 | 21×21 | 67 | 623 | 13.26 | +1.2 |
+| 265 | 18×18 | 32 | 131 | 13.30 | +1.2 |
+| 156 | 10×10 | 65 | 967 | 13.31 | +1.2 |
+| 222 | 16×16 | 21 | 87 | 13.34 | +1.2 |
+| 175 | 21×21 | 22 | 2,751 | 13.35 | +1.2 |
+| 119 | 12×12 | 165 | 1,912 | 13.37 | +1.1 |
+| 221 | 3×3 | 28 | 5,028 | 13.43 | +1.1 |
+| 374 | 10×10 | 144 | 633 | 13.45 | +1.0 |
+
+## QUICK WINS — Small grids with few base nodes (likely simple rules)
+
+These have small grids AND few nodes, suggesting the rule might be simple:
+- **Task 398** (1×5, 44 nodes) — absurdly small grid, probably a very simple rule
+- **Task 57** (8×8, 18 nodes) — very few nodes, maybe identity-like
+- **Task 384** (9×9, 39 nodes) — small grid, few nodes
+- **Task 217** (9×9, 34 nodes) — small grid, few nodes
+- **Task 34** (9×9, 87 nodes) — small grid
+
+## HIGH-VALUE TARGETS — Low base score, many nodes (base model is doing something complex)
+
+These have the biggest gains because the base model is very inefficient:
+- **Task 219** (15×10, 532 nodes, base 8.42) — enormous potential, similar to task 285
+- **Task 157** (10×15, 433 nodes, base 9.14) — rule partially understood (5→1 reflection)
+- **Task 182** (20×20, 522 nodes, base 10.63) — high node count = learned Conv model
+
+## APPROACH FOR NEXT SESSION
+
+1. Start with **quick wins** (398, 57, 384, 217, 34) — small grids, likely crackable fast
+2. Then hit **high-value** (219, 157, 182) — bigger gains per task
+3. For each: look at 2-3 train examples, crack rule, verify solver, build ONNX if viable
+
+## NOTE: These are STATIC estimates
+
+Static scoring can be off by 50%. Always profile with `score_model.py` before submitting.
+The profiled scores from LEARNING.md show that actual Kaggle scores can differ significantly from static estimates.