ARC-AGI-2 HONEST BASELINE

PURE LLM INFERENCE · NO OBJECT DECOMPOSITION · NO DSL · qwen2.5 on Lounge RTX 4090
TOPOLOGICAL BLINDNESS — PRE-OPTIC-NERVE
0%
AGI-2 EVAL SCORE
8.9%
EVAL CELL MATCH
5%
AGI-2 TRAINING PEAK
017c7c7b
ONLY TASK SOLVED
ALL RUNS — COMPLETE RECORD honest · unfiltered · exact telemetry
SET MODEL TASKS CORRECT SCORE AVG CELL MATCH NOTES
V8AGI-1 training ×5 runs qwen2.5:14b 54 each 14–19 25.9–35.2% AGI-1 bench · highly repetitive grids · pattern matching works
AGI-1Training 20-task qwen2.5:7b 20 0 0% ~0% Smaller model · AGI-1 · 0/20
AGI-1Training ×8 editions qwen2.5:7b 20 0 0% 0.0% Memorisation collapse confirmed · rotated grids → 0%
AGI-2Training qwen2.5:7b 20 0 0% 21.6% 2 "close" · 30×30 wall begins
AGI-2Training qwen2.5:14b 20 1 5% 32.8% 4 "close" · ONLY SOLVE: 017c7c7b (tiling+recolor — guessable)
AGI-2 EVALBlind set qwen2.5:14b 10 0 0% 8.9% 1 "close" · THE WALL · TOTAL COGNITIVE COLLAPSE

AUTOPSY — WHY THE WALL IS REAL

The AGI-1 mirage: 25-35% on v8 looked like progress. It was memorisation. Smaller grids, repetitive patterns, transformer pattern-matching working on pixel noise. The model never learned rules — it learned textures.

AGI-2 is different: 30×30 grids. Novel topologies. No two tasks share a pattern. The model receives 900 integers and must extract the spatial rule. At 8.9% cell match on the eval set, it is not even close to the right shape — it is hallucinating pixel values with no geometric grounding.

Rick's Law holds: You cannot transform an object you cannot perceive. 017c7c7b was solved because it was a simple tiling + recolor — the model could guess the answer from statistical patterns. The other 9 eval tasks required actual spatial reasoning. Flatline.

The fix is not a bigger model. The fix is the Optic Nerve.

EPOCH 3 — WHAT CHANGES THE NUMBER

✓ extract_objects_impl — 0 sorry · all 4 ARC_Object axioms proven
✓ BFS_ConnInv + BFS_MaxInv — connected · maximal · monochrome
✓ adjacent4_cells_spec — 2D space formally defined

OVERSEER will receive: [Object_1(Red, 4px, bbox(2,4,1,2)), Object_2(Blue, 6px, bbox(0,3,4,6))]
NOT: [[[0,0,1,0,...900 integers...]]]

Next burns:
· translate_object — shift_cell injectivity · IRF-ARC-DSL-002-OBJ
· ExtractInv — outer loop COVER + DISJOINT · global partition
· tau_select_mem — foldl invariant · target ∈ objs
· Run 4 — qwen2.5:14b + object decomposition payload · break the wall