kofdai/verantyx-arc-agi2

Verantyx ARC-AGI-2 — 16.1% (161/1000) 🏆

Pure rule-based solver for ARC-AGI-2. No LLM. No neural networks. No pre-training.

Score

Split	Score	Tasks Solved
Training	16.1%	161 / 1000

Approach

Verantyx is a fully deterministic, LLM-free ARC-AGI-2 solver built on program synthesis and domain-specific language (DSL) enumeration.

Architecture

1. Cross Engine — Multi-strategy solver with priority-based candidate selection
2. Neighborhood Rules — Exact and extended (count/directional/multi-pass) neighborhood matching
3. DSL Enumerator — 32 parameterless primitives with depth-2 composition (1024 combinations)
4. Panel Operations — Grid splitting (separators, equal partitions) + reduction (XOR, OR, AND, overlay, select)
5. Per-Object Transform — Object detection → property-based recoloring/movement
6. Puzzle Language — 25+ hand-crafted pattern detectors with auto-verification
7. Iterative Cross — 2-step residual learning for complex transformations

Puzzle Language Patterns (Phase 7)

Pattern	Description	Example Task
cross_and_fill	Fill cross intersections	—
diagonal_extend	Extend diagonal lines	—
connect_same_color	Connect same-colored cells	—
extract_tile	Detect repeated tile	7b7f7511
two_row_interleave	2 rows → checkerboard	e9afcf9a
stack_nor	Top+bottom halves NOR	fafffa47
frame_repeat_border	Frame with repeated border	49d1d64f
col_color_map	Column position → row color	a85d4709
split_vsep_and	Vertical split AND	0520fde7
grid_pattern	Lattice/checkerboard generation	332efdb3
antidiag_fill	Anti-diagonal + bottom fill	3bd67248
latin_square	Latin square completion	4cd1b7b2
shift_recolor	Shift + recolor fg cells	a79310a0
uniform_row_detect	Mark uniform rows	25d8a9c8
+ 10 more	...	...

Score History

Version	Score	Delta	Key Changes
v19	113 (11.3%)	—	Baseline with NB + DSL
v27	127 (12.7%)	+14	Extended NB rules
v28	136 (13.6%)	+9	split_combine + panel_extract
v29	142 (14.2%)	+6	puzzle_lang Phase 7
v30	144 (14.4%)	+2	connect_same_color
v31	146 (14.6%)	+2	u_drop_ball + nonzero_count_row
v32	147 (14.7%)	+1	scale_down_uniform
v33	149 (14.9%)	+2	staircase_grow + comp_tile_2x2
v34	154 (15.4%)	+5	frame_repeat + extract_tile
v35	158 (15.8%)	+4	col_color_map + split_vsep_and
v36	161 (16.1%)	+3	grid_pattern + latin_square

Performance

• Eval speed: ~0.40s/task (1000 tasks in ~7 minutes)
• No GPU required — runs on CPU only
• Deterministic — same input always produces same output

Usage

from arc.cross_engine import solve_cross_engine

# train_pairs: list of (input_grid, output_grid)
# test_inputs: list of input grids
predictions, versions = solve_cross_engine(train_pairs, test_inputs)

Files

• arc/cross_engine.py — Main solver orchestrator
• arc/puzzle_lang.py — Pattern-based puzzle language
• arc/nb_extended.py — Extended neighborhood rules
• arc/beam_search.py — Beam search with DSL
• arc/enumerator.py — DSL primitive enumerator
• arc/panel_ops.py — Panel split + reduce operations
• arc/per_object.py — Per-object transforms
• arc/primitives.py — 32 core DSL primitives

License

MIT

Author

kofdai — verantyx project