Upload medal-solvers/TODO.md with huggingface_hub

medal-solvers/TODO.md

@@ -2,81 +2,70 @@
 
 ## ⚠️ AGENT: DO THESE IN ORDER. Do NOT skip ahead or invent your own tasks.
 
-**Current**: 6070.53 (v50) | **Target**: 6500 (silver) | **Gap**: ~430 pts
+**Current**: 6070.53 + 2.62 = ~6073 (v51 pending) | **Target**: 6500 (silver) | **Gap**: ~427 pts
 **Strategy**: Crack rules for fixed-grid tasks, build ONNX, verify score > base, submit.
 
 ---
 
-## PHASE 1: QUICK WINS (small grids, likely simple rules)
-
-Crack these first — small grids = easy ONNX, fast iteration:
-
-| # | Task | Grid | Base Score | Base Nodes | Action |
+## DONE (14 models, +30.3 pts)
+
+| Task | Gain | Score | Build Script |
+|------|------|-------|-------------|
+| 025 | +1.83 | 10.54 | build_task025_onnx.py |
+| 028 | +1.47 | 14.30 | build_task028_onnx.py |
+| 057 | +1.09 | 14.17 | build_task057_onnx.py ⭐ NEW |
+| 062 | +0.50 | 14.74 | build_task062_onnx.py |
+| 084 | +2.67 | 12.39 | build_task084_onnx.py |
+| 153 | +4.10 | 14.91 | build_task153_onnx.py |
+| 200 | +4.32 | 14.64 | build_task200_onnx.py |
+| 217 | +1.06 | 13.88 | build_task217.py ⭐ NEW |
+| 255 | +4.16 | 11.57 | build_task255_onnx.py |
+| 277 | +1.55 | 14.85 | build_task277_onnx.py |
+| 285 | +2.83 | 8.82 | build_task285_scatter.py |
+| 319 | +3.74 | 11.65 | build_task319_onnx.py |
+| 384 | +0.47 | 13.17 | build_task384.py ⭐ NEW |
+| 387 | +0.50 | 10.96 | build_task387_onnx.py |
+
+## PHASE 1: QUICK WINS (remaining)
+
+| # | Task | Grid | Base Score | Base Nodes | Status |
 |---|------|------|-----------|-----------|--------|
-| 1 | **398** | 1×5 | 12.36 | 44 | Crack rule, build ONNX |
-| 2 | **57** | 8×8 | 13.08 | 18 | Crack rule (only 18 nodes!) |
-| 3 | **384** | 9×9 | 12.70 | 39 | Crack rule, build ONNX |
-| 4 | **217** | 9×9 | 12.82 | 34 | Crack rule, build ONNX |
-| 5 | **34** | 9×9 | 12.39 | 87 | Crack rule, build ONNX |
+| 1 | **398** | 1×5 | 12.36 | 44 | Script exists, score 10.05 (WORSE) — needs redesign |
+| 2 | **34** | 9×9 | 12.39 | 87 | Rule complex (2x2 → diagonal band), partially analyzed |
 
-## PHASE 2: HIGH-VALUE (big gains, more complex)
+## PHASE 2: HIGH-VALUE
 
-| # | Task | Grid | Base Score | Base Nodes | Potential | Action |
+| # | Task | Grid | Base Score | Base Nodes | Potential | Status |
 |---|------|------|-----------|-----------|-----------|--------|
-| 6 | **219** | 15×10 | 8.42 | 532 | +6.1 | Crack rule, build ONNX |
-| 7 | **157** | 10×15 | 9.14 | 433 | +5.4 | Rule partially known (5→1 reflection), finish cracking |
-| 8 | **182** | 20×20 | 10.63 | 522 | +3.9 | Crack rule, build ONNX |
-| 9 | **110** | 29×29 | 12.11 | 163 | +2.4 | Crack rule |
-| 10 | **54** | 30×30 | 12.15 | 118 | +2.4 | Crack rule |
+| 3 | **219** | 15×10 | 8.42 | 532 | +6.1 | Rule ~90% cracked (staircase completion), needs final fix |
+| 4 | **157** | 10×15 | 9.14 | 433 | +5.4 | Not attempted |
+| 5 | **182** | 20×20 | 10.63 | 522 | +3.9 | Not attempted |
+| 6 | **110** | 29×29 | 12.11 | 163 | +2.4 | Tile pattern completion, complex |
+| 7 | **54** | 30×30 | 12.15 | 118 | +2.4 | Not attempted |
 
 ## PHASE 3: MEDIUM GAINS
 
-| # | Task | Grid | Base Score | Base Nodes | Potential |
-|---|------|------|-----------|-----------|-----------|
-| 11 | 71 | 16×16 | 12.53 | 49 | +2.0 |
-| 12 | 5 | 21×21 | 12.62 | 106 | +1.9 |
-| 13 | 216 | 20×20 | 12.76 | 286 | +1.7 |
-| 14 | 128 | 15×15 | 12.78 | 28 | +1.7 |
-| 15 | 208 | 21×21 | 12.82 | 196 | +1.7 |
-| 16 | 201 | 13×13 | 12.90 | 135 | +1.6 |
-| 17 | 365 | 10×10 | 12.91 | 242 | +1.6 |
-| 18 | 165 | 20×20 | 13.02 | 25 | +1.5 |
-| 19 | 36 | 30×30 | 13.03 | 56 | +1.5 |
-| 20 | 107 | 5×5 | 13.24 | 52 | +1.3 |
+| # | Task | Grid | Base Score | Base Nodes | Potential | Status |
+|---|------|------|-----------|-----------|-----------|--------|
+| 8 | 71 | 16×16 | 12.53 | 49 | +2.0 | Enclosed-region filling (needs CC detection) |
+| 9 | 5 | 21×21 | 12.62 | 106 | +1.9 | Complex pattern completion |
+| 10 | 11 | 11×11 | 13.22 | 52 | +1.3 | Section-based voting, complex |
+| 11 | 162 | 20×20 | 13.22 | 19 | +1.3 | 3x3 block addition, needs CC |
+| 12 | 265 | 18×18 | 13.30 | 32 | +1.2 | 2x2 pattern fill, partially analyzed |
 
 ---
 
 ## WORKFLOW PER TASK
 
-```
-1. Load task JSON from task-data/taskNNN.json
-2. Look at train examples (input→output), identify the rule
-3. Write Python solver, verify 265+/265 accuracy
-4. Assess ONNX viability:
-   - Can the rule be expressed as coordinate arithmetic? (row/col grids, masks, modular ops)
-   - Does it need CC detection, flood fill, or dynamic matching? → NOT VIABLE
-   - Estimate node count. If >300 nodes AND grid >15×15 → probably not worth it
-5. Build ONNX using OnnxBuilder (see onnx_builder.py)
-6. Apply strict shape inference fix
-7. Score with score_model.py — verify score > base
-8. If gain confirmed: save to optimized/ directory, update this TODO
-```
-
-## DONE (11 models, +27.7 pts)
-
-| Task | Gain | Score |
-|------|------|-------|
-| 025 | +1.83 | 10.54 |
-| 028 | +1.47 | 14.30 |
-| 062 | +0.50 | 14.74 |
-| 084 | +2.67 | 12.39 |
-| 153 | +4.10 | 14.91 |
-| 200 | +4.32 | 14.64 |
-| 255 | +4.16 | 11.57 |
-| 277 | +1.55 | 14.85 |
-| 285 | +2.83 | 8.82 |
-| 319 | +3.74 | 11.65 |
-| 387 | +0.50 | 10.96 |
+
+
+## KEY LEARNINGS FROM THIS SESSION
+
+1. **Resize op with empty ROI array (dims=[0]) causes scoring failure** — use ConvTranspose instead for upscaling
+2. **Dynamic Slice with computed start/end causes unknown shapes** — use static slices + Gather for selection
+3. **Task 398 existing script generates 328 nodes** → too many intermediates, score worse than base
+4. **Task 219 rule** is: find reference staircase group, fill uncovered columns in partial groups with color 1. Almost cracked (228/265) but edge cases with per-column coverage remain
+5. **Good patterns for ONNX**: Extract→Transform (task 057), Fractal tile (task 217), 2x upscale with Gather (task 384)
 
 ## DO NOT TOUCH (base already excellent)