# Reverse Engineering 5743 Submission — Complete Analysis

> **Base**: commit 0b46512 (V58, LB 1430.24, 105 tasks)
> **Target**: Implement solutions for Tasks 33, 39, 65, 74, 126, 146 + extend two-layer Conv to 50+ tasks
> **Critical Rule**: NEVER use string replacement for registry entries — download, manually edit the full file, verify with regex, then upload

---

## Part 1: Implemented Solvers (in wave6.py)

| Task | Solver | Pattern | Validated |
|------|--------|---------|-----------|
| 033 | `s_periodic_tile_fill` | Period-6 tiling + bg-fill from last-row color | ✅ exact match |
| 039 | `s_bbox_crop_3x3` | Bounding box crop to 3×3 via ArgMax+Gather | ✅ exact match |
| 065 | `s_centroid_zoom` | Single-pixel centroid → zoom/pan 15×15 grid | ✅ exact match |
| 074 | `s_eightfold_symmetry` | D4 symmetry: Max of 8 flip/shift/transpose | ✅ exact match |
| 126 | `s_ushape_fill_bottom` | 5-dir shift multiply → U-detect → fill bottom with color 4 | ✅ exact match |
| 146 | `s_select_asymmetric_block` | 3 stacked 3×3 blocks → select non-diagonal-symmetric | ✅ exact match |

---

## Part 2: The Two-Layer Conv Opportunity (50+ tasks)

### Architecture Pattern

5743 uses this learned architecture for Tasks 4, 23, 37, 90, 93, 369:

```
Cast(fp16) → Conv(10→H, K×K, pad=K//2) + bias → ReLU → Conv(H→10, 1×1) + bias → output
```

Task 4: K=3, H=10, total 4 nodes (smallest possible), fp16 weights ~2KB

### Why This Matters

Our current `conv.py` does **single-layer linear regression** (lstsq):
- Can only learn LINEAR mappings from patches to output
- Fails on any task needing AND/OR/threshold logic
- Ceiling: ~25 tasks

The two-layer approach adds:
- **Non-linearity** (ReLU) → learns threshold-based rules
- **Hidden representation** → combines multiple local patterns
- **Larger receptive field** → K=7,9,11 sees more context

### Pattern Distribution in 5743 (146 Conv-using tasks total)

| Pattern | Task Count | Description |
|---------|-----------|-------------|
| Neighbor counting (all-ones 3×3) | 29 | Sum 3×3 neighborhood → threshold → action |
| Shift Conv (directional shifts) | 50 | Move content 1px in direction → detect adjacency |
| Two-layer trained (ReLU) | 6 | Full learned ConvNet |
| Channel select + spatial | 31 | Conv1x1(filter) + Conv(spatial) + threshold |
| Other Conv patterns | 30+ | Various compositions |

### Implementation Plan

```python
def solve_conv_twolayer(td, path, providers, time_budget=60.0):
    """Train 2-layer Conv with ReLU via PyTorch, export to ONNX.
    
    Sweep: kernel_size ∈ [3,5,7,9,11,13,15], hidden ∈ [4,6,8,10,16,20]
    Training: Adam lr=0.01, CrossEntropyLoss, early stop at 100% accuracy
    Validation: must match ALL arc-gen examples
    Export: opset 17, fp16 weights, static shapes
    """
```

Conservative estimate: 30-50 new tasks solved
Optimistic: 50-80 new tasks solved

---

## Part 3: Registry Update Protocol

**CRITICAL: NEVER use string replacement for registry entries!**

### Correct Protocol:
1. Download full `solver_registry.py` via `hf_repo_files(operation='read')`
2. Parse entire content into memory
3. Add new imports at the exact right location
4. Add entries to ANALYTICAL_SOLVERS list at the desired position
5. Verify with regex: count entries, check no duplicates, validate syntax
6. Upload the COMPLETE edited file via `hf_repo_files(operation='upload')`

### What to Add:

**Imports** (after the wave5 import block):
```python
from .wave6 import (s_select_asymmetric_block, s_bbox_crop_3x3, s_periodic_tile_fill,
                    s_eightfold_symmetry, s_ushape_fill_bottom, s_centroid_zoom)
```

**Registry entries** (add to ANALYTICAL_SOLVERS list):
```python
    # Wave 6 solvers (reverse-engineered from 5743)
    ('select_asymmetric_block', s_select_asymmetric_block),
    ('bbox_crop_3x3', s_bbox_crop_3x3),
    ('periodic_tile_fill', s_periodic_tile_fill),
    ('eightfold_symmetry', s_eightfold_symmetry),
    ('ushape_fill_bottom', s_ushape_fill_bottom),
    ('centroid_zoom', s_centroid_zoom),
```

---

## Part 4: Key Findings from 5743 Analysis

### Model Statistics
- Total files: 400 ONNX models
- Trivial (≤3 nodes): 43 tasks (identity, constant, simple permutation)
- Simple (4-10 nodes): 37 tasks
- Medium (11-20 nodes): 69 tasks
- Complex (>20 nodes): 251 tasks

### Opset Distribution
- Opset 10: 131 tasks (most common)
- Opset 11: 73 tasks
- Opset 13: 60 tasks
- Opset 17: 56 tasks
- Opset 12: 36 tasks
- Opset 16: 43 tasks

### File Sizes (target tasks)
All target models are tiny (1-2 KB), scoring ~19-21 points each.

### What 5743 Has That We Lack
1. **Shift-Conv propagation** (50 tasks): Uses Conv kernels like `[0,1,0]` to shift content
2. **Neighbor counting** (29 tasks): All-ones 3×3 Conv for neighborhood sums
3. **ArgMax scalar grid** (40+ tasks): Works on decoded color grid, not one-hot
4. **fp16 + uint8 mixed precision**: Reduces file size and MACs
5. **Two-layer learned ConvNet** (6 tasks): PyTorch-trained, ONNX-exported