Update TODO.md

TODO.md

@@ -0,0 +1,29 @@
+Here's what I need from you and what the situation looks like:
+
+Current Status: 306/400 ✅ (was 294)
+What the model_runner.py code IS
+That's the 2024 ARC Prize winning LLM infrastructure — it fine-tunes a language model to generate ARC solutions as text. It's a completely different 
+competition format (text answers, 
+not ONNX models). It won't help us directly.
+
+What WOULD help — things you can search for:
+1. ARC-GEN additional training data The ARC-GEN paper generates extra input/output examples for each task. More examples → our lstsq conv solver gets more
+
+2. data → better fitting → more tasks solved. Search Kaggle/GitHub for "arc-gen" data files or the generator code.
+
+3. NeuroGolf / Code Golf Championship notebooks Search Kaggle for neurogolf 2026 or code golf championship ARC ONNX. Other competitors' approaches would be gold.
+
+4. CompressARC code (paper) This 76K param approach trains a tiny network PER TASK via gradient descent. If we could adapt it to produce opset-10 ONNX models,
+
+5. it could solve many of our 94 remaining tasks. Search GitHub for "CompressARC" or the authors' repos.
+
+What's actually blocking us (94 unsolved tasks)
+Category	Count	Example	Why unsolved
+Variable diff-shape (output smaller)	~60	Extract subregion from grid	Output shape depends on input content — can't build static ONNX
+Variable diff-shape (output larger)	~17	Tile/upscale by variable factor	Same problem
+Same-shape but complex	~10	Multi-step reasoning, flood fill	Conv can't learn non-local/algorithmic patterns
+Fixed diff-shape (output larger)	~7	Input-driven block placement	Output depends on input VALUES, not just positions
+The fundamental challenge: 94 tasks require reasoning that depends on input content (not just a fixed pixel remapping or local conv pattern). 
+Our current ONNX opset 10 toolkit (Conv, Gather, ArgMax, etc.) can only express fixed mappings. 
+We'd need to find tasks where the mapping IS fixed but our solver just hasn't found it yet — likely by adding more training examples
+(ARC-GEN) or trying bigger conv kernels with more time budget.