Table 10 reports the full zero-shot evaluation row for OPT-1.3B.

openai/gpt-5.6-luna-pro

523870770c

The evaluation used an NVIDIA L4, CUDA float16, batch size 8, and lm-evaluation-harness 0.4.2; the paper does not fully specify the hardware, precision, batching, or exact evaluator revision for this row. · The paper's table labels the multiple-choice columns as acc, while the evaluator also emits acc_norm for several tasks. This run reports the literal ordinary acc field and retains the normalized alternatives in the raw output; the paper does not explicitly document its metric convention, so the comparison is conditional on that interpretation. · The official released checkpoint bytes and tokenizer identity were verified, but equivalence to the authors' historical training state and data order cannot be established from final-weight evaluation. · Only the assigned OPT-1.3B row was evaluated. Other Table 10 rows, training claims, and the paper's separate speed and MQAR claims were not tested in this run. · The full-evaluation invocation and exit evidence are omitted or truncated in the supplied audit records. · The paper does not explicitly define whether its generic multiple-choice acc columns use ordinary acc or acc_norm, and the raw evaluator exposes both for several tasks. · The declared workload contains 29 model-row units, but retained raw coverage demonstrates only the OPT-1.3B row. · The paper's hardware, precision, batching, exact evaluator revision, and historical checkpoint state are not fully specified or established as identical. · Execution evidence provenance is unknown: The raw output and evaluator configuration are retained, but the direct runner-captured full-evaluation command and exit output are not present in the supplied command records; item_51 is truncated and primarily shows a normalization event, while items_56 and 57 inspect previously written files. The full-run provenance is therefore not independently established.