# medal-solvers

## ⚠️ NEW AGENT: Read these files IN ORDER before doing anything:
1. **LEARNING.md** — scoring rules, what works, what doesn't
2. **TODO.md** — THE ordered task list with base scores
3. **NEXT_AGENT.md** — current state, instructions, priorities
4. **BULK_SCAN.md** — all 400 tasks profiled with base scores

---

## What This Is

Hand-crafted ONNX models that solve ARC-like tasks more efficiently than the competition's base models. Each model that scores higher than the base adds points to our Kaggle leaderboard score.

## Current Score: 6072.59 (v66, rank ~150) | Target: 6500 (silver)

## How It Works

1. Each task has a rule (input grid → output grid)
2. The competition provides a base ONNX model per task (usually a learned Conv net)
3. We hand-craft smaller ONNX models that implement the rule in fewer nodes/params
4. Run **onnxsim** on every model (proven +0.01 to +0.77 per model)
5. Smaller model = lower memory = higher score (see scoring formula in LEARNING.md)
6. **ALWAYS verify score > base before uploading**

## Quick Start

```bash
pip install onnx onnxruntime onnxsim numpy

# Score a model vs base:
python score_model.py --model optimized/task200.onnx --task-num 200 --task-data-dir task-data/
python score_model.py --model ../base-models/task200.onnx --task-num 200 --task-data-dir task-data/

# Build submission:
python build_submission.py --base ../submission-6043.zip --optimized-dir optimized --output submission.zip
```

## Current Models (15 tasks, ~+32 pts)

| Task | Gain | Method |
|------|------|--------|
| 025 | +1.86 | Hand-craft + onnxsim |
| 028 | +1.48 | Hand-craft + onnxsim |
| 062 | +0.51 | Hand-craft + onnxsim |
| 084 | +2.69 | Hand-craft + onnxsim |
| 153 | +4.19 | Hand-craft + onnxsim |
| 200 | +4.32 | Hand-craft + onnxsim |
| 209 | +0.35 | onnxsim on base |
| 217 | +0.90 | Hand-craft + onnxsim |
| 255 | +4.21 | Hand-craft + onnxsim |
| 277 | +1.59 | Hand-craft + onnxsim |
| 285 | +2.85 | Hand-craft + onnxsim |
| 319 | +3.96 | Hand-craft + onnxsim |
| 323 | +0.01 | onnxsim on base |
| 367 | +0.10 | onnxsim on base |
| 387 | +0.70 | Hand-craft + onnxsim |

## Key Rules

```python
# ALWAYS before saving any model:
from onnxsim import simplify
model_simp, ok = simplify(model)
del model_simp.graph.value_info[:]
model_simp = onnx.shape_inference.infer_shapes(model_simp, strict_mode=True)
onnx.save(model_simp, output_path)

# ALWAYS score vs base before uploading:
my_score = score_model(my_path, task_num, task_data_dir)
base_score = score_model(base_path, task_num, task_data_dir)
assert my_score > base_score, "Don't upload - worse than base!"
```

## Resources

- `michaelhodel-arc-dsl.txt` — ARC Domain Specific Language (primitives for cracking rules)
- `onnx-neural-compressor.txt` — Quantization docs (doesn't work with scorer, don't use)
- `submission-6043.zip` — All 400 base models for comparison
- `own-solver/neurogolf-2026.zip` — All 400 task JSON files