Table 10 reports the full zero-shot evaluation row for Hybrid H3-130M.

openai/gpt-5.6-luna-pro

613f4b3397

The contracted paper Markdown file was absent; the fixed PDF was used for Table 10 on physical page 52 and Appendix D.3 on physical page 51. · The available corresponding public release is named H3-125M, whereas the paper row is labeled Hybrid H3-130M; the exact original paper checkpoint was not separately available, so exact-weight equivalence is not established. · The evaluation ran on CPU rather than a reference accelerator; no hardware-specific performance claim is made. · The paper labels the table columns acc and does not specify acc_norm. The protocol therefore selected ordinary acc,none while retaining neighboring acc_norm,none values; comparisons using a different metric convention would not be equivalent. · The exact original Hybrid H3-130M checkpoint was not separately available; the run used danfu09/H3-125M at the recorded revision and checkpoint hash. · The declared workload contains 29 model-row units, but raw execution evidence demonstrates only the H3 row and its nine assigned measurements. · The source Markdown path was absent, so the fixed PDF was used. The supplied PDF excerpt does show the H3-130M Table 10 row and its reported values. · The paper labels the accuracy columns as acc, while the harness also emitted acc_norm fields; ordinary acc,none was selected, but the source does not provide a more detailed metric-convention specification. · The evidence context does not retain the complete revised evaluator adapter and CPU fallback implementation: the main script is truncated and several helper modules are omitted, limiting independent inspection of every implementation path, although the captured command and raw output establish the executed configuration. · The captured execution demonstrates a complete fresh zero-shot evaluation of the seven Table 10 tasks for one H3 release, using GPT-2 tokenization, lm-eval 0.4.2, ordinary acc,none fields, and a recomputed seven-task average. However, the tested artifact is danfu09/H3-125M, not the exact Hybrid H3-130M checkpoint used for the paper row; the original checkpoint was not available. This is a material scientific difference, so neither the close matches nor the large discrepancies can establish or refute the H3-130M row. The experiment establishes only the reported values for the substituted H3-125M release under CPU execution. In addition, the declared workload lists 29 model-row units, while the captured run covers only the assigned H3 row.