Harness versus weights: the gap these two scoreboards support is 69.8 points, not 23.8

The mechanism NVIDIA puts forward has two parts. Persistent memory keeps prior implementations, evaluation results, compiler and profiler output and accumulated reasoning available, so the agent resumes from its current state instead of rebuilding the search; a supervisor watches the trajectory for stagnation and redirects the main agent when a line of attack stops paying ^[6]. On a benchmark where the agent is dropped into an unfamiliar environment with no instructions, no stated rules and no stated goals, that state-keeping is doing work no single context can do ^[7]. NVIDIA's evidence for generality is that the ARC-AGI-3 agent is the same agent it ran on GPU kernels, with only environment-specific tools and evaluation changed ^[5]. The kernel run is the better stamina demonstration: seven days of continuous operation, more than 500 optimization directions, 40 committed kernel versions ^[8].

The accounting is where it stops. The 100.00 is an RHAE score reported by NVIDIA on its own developer blog ^[1][11]; the 30.2% is a verified leaderboard entry published by ARC Prize ^[2][10]. Subtract them and you get 69.8 points ^[12]. No figure in either source, and no difference between figures in either source, comes to 23.8 ^[13].

Even 69.8 is not a wrapper number. The post never names the frontier model inside AVO ^[9], so nothing in the delta can be assigned to scaffolding rather than to weights. And the environment-specific tooling was built for this benchmark ^[5], which means a like-for-like test would still be measuring tools plus memory plus supervisor together. Meanwhile the comparison figure is soft in the other direction: ARC Prize says Opus 5 was run on ARC-AGI-3 only at High reasoning effort because the testing window was short, and it reports 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 at Max ^[3].