Add TRM solver README

trm_solver/README.md

@@ -0,0 +1,84 @@
+# TRM Solver — NeuroGolf 2026
+
+LLM-driven neural network executor for ARC-AGI image transformations.
+
+## Architecture
+
+```
+ARC Grid → Kilo/DeepSeek (headless CLI, free tier)
+         → Markdown: transform name + parameters
+         → Tiny NN executor (frozen weights)
+         → ONNX export (per-task)
+```
+
+## Why this works for NeuroGolf
+
+- **LLM does the heavy reasoning** (identifying transforms, shapes, patterns)
+- **NN is tiny** — just executes the identified transform with frozen weights
+- **ONNX is absurdly small** — most transforms have 0-1000 params
+- **Free tier LLM** — Kilo headless CLI uses DeepSeek free tier, no API cost
+
+## File structure
+
+- `executor.py` — 14 transform NN implementations + ONNX export
+- `kilo_bridge.py` — Kilo/DeepSeek interface + full pipeline
+- `requirements.txt` — dependencies
+
+## Quick start
+
+### Dry run (no LLM needed)
+```bash
+python trm_solver/kilo_bridge.py --dry-run --output-dir onnx_models
+```
+
+### Single task
+```bash
+# Render task and call Kilo
+python trm_solver/kilo_bridge.py --task data/training/007bbfb7.json --output-dir onnx_models
+
+# Pre-rendered image
+python trm_solver/kilo_bridge.py --task rendered_task.png --no-render --output-dir onnx_models
+```
+
+### Batch mode
+```bash
+python trm_solver/kilo_bridge.py --data-dir data/training --output-dir onnx_models --max-tasks 10
+```
+
+### Via SDK tunnel (Kaggle → local server)
+```bash
+python trm_solver/kilo_bridge.py --data-dir data/training --use-sdk --server-url http://localhost:8765
+```
+
+## Transform types (14)
+
+| Transform | Params | Description |
+|-----------|--------|-------------|
+| identity | 0 | Output equals input |
+| color_map | 100 | Per-pixel color remapping |
+| flip | 0 | Horizontal/vertical flip |
+| transpose | 0 | Matrix transpose |
+| rotate | 0 | 90/180/270 rotation |
+| upscale | <100 | Nearest-neighbor upscale |
+| kron_self_similar | 0 | Kronecker self-similar |
+| tile_repeat | 0 | Tile input n×m |
+| concat_patterns | 0 | Concat transformed copies |
+| pos_color_lut | H×W | Position-based color LUT |
+| spatial_gather | H×W×2 | Pixel rearrangement |
+| onehot_conv | K²×100 | One-hot convolution (most powerful) |
+| onehot_linear | 100 | One-hot linear transform |
+
+## Dependencies
+
+```
+torch
+numpy
+Pillow
+# For Kilo: kilo-cli or kilo-sdk
+```
+
+## Notes
+
+- Free tier Kilo has latency (~10-60s per task). Batch overnight for 400 tasks.
+- If Kilo misidentifies a transform, edit the `_spec.json` and re-export manually.
+- The onehot_conv transform solved 220 tasks in the previous approach — it's the fallback.