Update results: 47/400 (11.8%) — ITT solves 16, DSL solves 31, +7 new from ITT physics"

arc_results/RESULTS.md

@@ -1,117 +1,66 @@
 # PEMF Solver — ARC-AGI Training Set Evaluation
 
-## Results (v2 — object layer + greedy stacker)
+## Results (v3 — ITT physics + DSL beam)
 
-| Metric | v1 | **v2** | Change |
+| Metric | v1 | v2 | **v3** |
 |---|---|---|---|
-| **Tasks solved** | 31 (7.8%) | **40 (10.0%)** | **+9 (+29%)** |
-| ≥1 pair solved | 59 (14.8%) | — | — |
-| Pairs solved | 146 (11.2%) | — | — |
-| Transforms | 19 | **33** | +14 |
-| Total time | 17.1s | **50.8s** | +greedy stacker |
-| Regressions | — | **0** | — |
+| **Tasks solved** | 31 (7.8%) | 40 (10.0%) | **47 (11.8%)** |
+| via ITT | — | — | **16** |
+| via DSL | 31 | 40 | **31** |
+| Transforms (DSL) | 19 | 33 | 33 |
+| ITT rule types | — | — | 7 |
+| Total time | 17s | 51s | **36s** |
+| Regressions | — | 0 | **0** |
 
-## 9 Newly Solved Tasks
+## 7 Newly Solved Tasks (ITT-exclusive)
 
-| Task ID | Transform | Category |
+| Task ID | ITT Rule Type | What it does |
 |---|---|---|
-| 1f85a75f | ExtractLargestObject | **Object extraction** |
-| 22168020 | ConnectSameColorH | **Connect** |
-| 22eb0ac0 | ConnectSameColorH | **Connect** |
-| 23b5c85d | ExtractSmallestObject | **Object extraction** |
-| 4347f46a | DrawBorder | **Border** |
-| 746b3537 | tile_to_target | Tiling (new pair coverage) |
-| be94b721 | CropToContent | Crop (new pair coverage) |
-| ded97339 | overlay(ConnectSameColorV, ConnectSameColorH) | **Greedy stacker** |
-| eb5a1d5d | CompressGrid | **Compress** |
+| 0d3d703e | recolor | Direct color substitution |
+| 868de0fa | multi_region_fill | Frame size → fill color mapping |
+| 8d5021e8 | tile | Tile with learned reflection pattern |
+| b1948b0a | recolor | Color remap |
+| c0f76784 | multi_region_fill | Frame size → fill color with fallback |
+| c8f0f002 | recolor | Color substitution |
+| d511f180 | recolor | Color remap |
 
-**6 of the 9 new solves use new transforms.** 1 uses the greedy stacker overlay composition.
-
-## All 40 Solved Tasks
-
-| Task ID | Transform | Family |
-|---|---|---|
-| 007bbfb7 | KroneckerSelfSimilar | Self-similar |
-| 1190e5a7 | KroneckerSelfSimilarInv | Self-similar |
-| 1cf80156 | CropToContent | Crop |
-| 1e0a9b12 | GravityDown | Gravity |
-| 1f85a75f | ExtractLargestObject | Object |
-| 2013d3e2 | CropToContent | Crop |
-| 22168020 | ConnectSameColorH | Connect |
-| 22eb0ac0 | ConnectSameColorH | Connect |
-| 239be575 | tile_to_target | Tiling |
-| 23b5c85d | ExtractSmallestObject | Object |
-| 28bf18c6 | CropToContent | Crop |
-| 2dee498d | tile_to_target | Tiling |
-| 3906de3d | GravityUp | Gravity |
-| 3af2c5a8 | MirrorTileH | Mirror |
-| 3c9b0459 | Rotate_180 | Rotation |
-| 4347f46a | DrawBorder | Border |
-| 6150a2bd | Rotate_180 | Rotation |
-| 62c24649 | MirrorTile4Way | Mirror |
-| 67a3c6ac | Reflect_v | Reflection |
-| 67e8384a | MirrorTile4Way | Mirror |
-| 68b16354 | Reflect_h | Reflection |
-| 6d0aefbc | MirrorTileH | Mirror |
-| 6fa7a44f | MirrorTileV | Mirror |
-| 746b3537 | tile_to_target | Tiling |
-| 74dd1130 | Transpose | Transpose |
-| 7b7f7511 | tile_to_target | Tiling |
-| 8be77c9e | MirrorTileV | Mirror |
-| 9172f3a0 | Upscale_3x | Upscale |
-| 9dfd6313 | Transpose | Transpose |
-| a416b8f3 | tile_to_target | Tiling |
-| be94b721 | CropToContent | Crop |
-| c59eb873 | Upscale_2x | Upscale |
-| c9e6f938 | MirrorTileH | Mirror |
-| d10ecb37 | tile_to_target | Tiling |
-| d631b094 | ExtractLargestObject | Object |
-| d9fac9be | tile_to_target | Tiling |
-| de1cd16c | KeepSmallestObject | Object |
-| ded97339 | overlay(ConnectSameColorV, ConnectSameColorH) | Stacker |
-| eb5a1d5d | CompressGrid | Compress |
-| ed36ccf7 | Rotate_90 | Rotation |
+**None of these 7 tasks were solvable by the DSL beam search.** They require analyzing training pairs *together* to learn the rule — the ITT engine does this via `TransformationRule.learn()`.
 
 ## Architecture
 
-### 33 Atomic Transforms
-- **Tiling**: tile_to_target, ShiftedTile, FillEnclosedHarmonic
-- **Self-similar**: KroneckerSelfSimilar, KroneckerSelfSimilarInv
-- **Mirror**: MirrorTileH, MirrorTileV, MirrorTile4Way
-- **Upscale**: Upscale_2x, Upscale_3x
-- **Stack**: StackH_3, StackV_3
-- **Structural**: Transpose, CropToContent
-- **Object**: ExtractLargest/Smallest/Unique/MostCommon, KeepLargest/Smallest, SortBySize
-- **Fill/Connect**: FillInterior, ConnectSameColorH/V
-- **Compress**: CompressGrid, RemoveBlackLines
-- **Spatial**: ColorByProximity, DrawBorder
-- **Symmetry**: Rotate_90/180/270, Reflect_h/v
-- **Gravity**: GravityDown, GravityUp (optional)
-- **Color**: InvertColors (optional)
+### ITT-first, DSL-fallback
+1. **ITT solver** tries first: learns rule from training pairs using σ-analysis, ρ_q boundary charge, and field invariants
+2. If ITT achieves σ=0 on ALL training pairs → use it (16 tasks)
+3. Otherwise → **DSL beam search** with 33 transforms + greedy stacker (31 tasks)
 
-### Greedy Stacker
-After beam search, tries `overlay(T1(input), T2(input))` for the top-N depth-1 pieces. This allows composition of two independent transforms — critical for tasks where the output is a combination of two views.
+### ITT Physics Engine (`itt_engine.py`)
+- **PhiField**: dual-field Φ_q (int semantics) + Φ̃ (smooth float operators)
+- **ρ_q = |∇(∇²Φ̃)|**: boundary charge with physics-derived threshold (μ+1.5σ)
+- **SigmaResidue**: classifies transformations as fill/expansion/compression/recolor/erase
+- **Fan Signature**: 6-bit [Δ₁..Δ₆] task routing
+- **TransformationRule.learn()**: learns tile_pattern, size_to_color, frame_to_fill, color_map, shape_to_color from training pairs
+- **FieldInvariants**: enclosed mask, frame components, shape eigenspectrum, Fourier period detection
 
-### Object Layer (`object_layer.py`)
-Connected component extraction (4-conn, 8-conn), color splitting, list reducers (largest/smallest/most_common/unique), spatial queries (bbox, center, height, width), and composition (paint/underpaint/overlay/cover/canvas).
-
-## Parameters
-```json
-{
-  "beam_width": 8,
-  "max_depth": 2,
-  "lock_coeff": 0.0,
-  "max_fraction": 1.0,
-  "use_symmetry": true,
-  "use_gravity": true,
-  "use_color_ops": true,
-  "boundary_source": "target"
-}
-```
+### ITT Rule Types
+| Rule | Description |
+|---|---|
+| tile | Tile with learned per-block reflection/rotation pattern |
+| self_tile | Kronecker: input nonzero mask as meta-layout |
+| fill_enclosed | Fill enclosed regions via ρ_q boundary detection |
+| multi_region_fill | Frame size → fill color with fallback chain |
+| periodic_extension | Fourier period detection + color remap + tile |
+| shape_indicator | Laplacian eigenspectrum → output color |
+| recolor | Direct color substitution learned from pairs |
 
 ## How to reproduce
 ```bash
 git clone https://github.com/fchollet/ARC-AGI.git /tmp/arc
 cp -r /tmp/arc/data/training arc_data/training
-python scripts/run_all_arc.py
+python scripts/run_all_arc.py          # DSL only (40 tasks)
+python scripts/run_all_arc_v3.py       # ITT + DSL (47 tasks)
 ```
+
+## References
+- Original ITT solver: [Sensei-Intent-Tensor/0.0_ARC_AGI](https://github.com/Sensei-Intent-Tensor/0.0_ARC_AGI)
+- ITT physics: [Sensei-Intent-Tensor/0.0._Executable_Physics](https://github.com/Sensei-Intent-Tensor/0.0._Executable_Physics)
+- Zenodo: [https://zenodo.org/records/18077258](https://zenodo.org/records/18077258)