Table 10 reports the full zero-shot evaluation row for Hybrid H3-360M.

openai/gpt-5.6-luna-pro

8bd4894c21

The paper's hardware conditions were not available; all measurements were run on an NVIDIA L4. This supports numerical results under the observed device, not any hardware-specific performance claim. · The evaluator used unfused PyTorch compatibility modules because the repository's FlashAttention dependencies were unavailable, and the official H3 implementation used its slow Vandermonde fallback because pykeops was unavailable. · Batch size was 1, with maximum sequence length 2048 and one evaluation repetition per task; no variance across repeated evaluations was estimated. · The available released artifact is H3-355M, evaluated as the checkpoint corresponding to the paper's Hybrid H3-360M Table 10 row label; the artifact-label distinction is retained rather than silently treating the names as identical. · The paper's Table 10 values were read before execution, so the comparison was not blind. · The evaluated artifact was H3-355M, while the claim names Hybrid H3-360M; no supplied evidence verifies that these are the same checkpoint or training stage. · The declared workload contains 29 complete model rows, but captured raw coverage demonstrates only the seven tasks for the H3 row; the other 28 rows lack execution evidence. · The metric-binding revision occurred after results were observed. The supplied evidence does not include the cited harness task YAMLs, metrics implementation, official H3 README, or full official source files needed to independently verify the revised semantics. · Execution used an NVIDIA L4, unfused PyTorch compatibility modules, and a Vandermonde fallback; the paper hardware and exact implementation conditions were not established. The L4 difference alone does not invalidate accuracy, but it remains a recorded protocol difference. · Only one evaluation repetition per task was run, and the paper values were available before execution. · Captured records establish a fresh seven-task evaluation of the released H3-355M checkpoint using GPT2 tokenization on an NVIDIA L4, with raw task fields and a recomputed average. The values are numerically close to the displayed Hybrid H3-360M row. However, equivalence between the H3-355M artifact and the paper's H3-360M checkpoint is not established. In addition, the revised plan changed several metrics from ordinary acc to acc_norm after observing the results, and the supplied evidence does not independently establish that convention from the paper or the omitted task-definition files. The declared workload covers 29 complete model rows, whereas raw evidence demonstrates only the H3 row. Therefore the observed pattern is consistent with the claim but cannot verify the claimed full H3-360M row under comparable, fully evidenced conditions.