120 visual puzzles from the ARC-AGI-2 public evaluation set. Each task shows 2 to 6 example pairs of coloured grids (rows of the digits 0 to 9, up to 30 by 30) and asks for the output grid of a new input; 45 tasks have two or three test grids, so the run has 167 Cases, one per test grid. The prompt and the reply parser are the ARC Prize harness's own (a \boxed{} grid, else the last JSON grid in the reply), and a grid counts only if every cell matches. The Candidate gets two Attempts per grid and the grid is solved if either matches. The Headline Score is the ARC Prize's: the mean over tasks of grids solved / grids in the task, so a task with two grids and one solved scores 0.5. A failed Case (both Attempts failed) counts 0 in its task; one failed Attempt (a reply cut at the token cap) scores 0 for that Attempt and the other still counts, as the harness credits it; a task whose every grid failed drops out of the mean, where the harness counts it 0. Named Deviation: when a reply holds no readable grid the harness asks the model again to extract one (OpenAI-style and Anthropic models) or re-asks the question; here that Attempt scores 0. So a run whose Report shows no failed Case and no "no grid found" Attempt is scored exactly as the harness scores the same replies; answers still vary from run to run, and the Candidate's settings (temperature, reasoning effort, token cap) must match to compare. Answers run to 2,760 characters before any reasoning, so a low token cap fails Attempts. The Human Baseline, 0.75, is per task: the share of human attempts that solved a task, averaged over the final tasks, where a task counts only if all its grids are right (66% per test grid). The Frontier Score is from the ARC Prize's semi-private set, a different 120 tasks calibrated to the same human difficulty, so it sits beside our public-eval number rather than on the same scale. The public set has been on GitHub since March 2025 and may be in a model's training data. Answer format: write the output grid as a list of rows, as in '[[1, 2], [3, 4]]', and make it the last bracketed text in the reply (a grid in \boxed{} is read first): 'Output: [[1, 2], [3, 4]]' counts, but 'Output: [[1, 2], [3, 4]], last row [3, 4]' ends on a single row, is read as no grid, and scores 0.