Table 10 reports the full zero-shot evaluation row for Mamba-790M.

openai/gpt-5.6-luna-pro

cb87475c88

The paper Markdown source named by the contract was absent; comparison targets were read from fixed PDF Table 10, physical page 52, and are not fresh evidence. · The evaluation ran on an NVIDIA L4 rather than the paper's reference hardware; these results do not establish any hardware-specific speed claim. · The installed evaluator did not establish the paper's exact held-out Pile validation split, so Pile validation perplexity remains unresolved and no Pile component score was substituted. · The evaluator emitted both raw acc and acc_norm for several tasks. The declared and retained binding is raw acc,none because Table 10 labels the metric acc; neighboring acc_norm values are disclosed in table10-raw-results.json but are not substituted based on numerical proximity. · Unresolved metric interpretation: m-t10-m790-pile: The paper defines this as Pile validation perplexity, but the installed lm-eval 0.4.2 Pile task exposes component training data rather than the paper's exact held-out validation split. I will not substitute a training-component score. · Unresolved measurements: m-t10-m790-pile: The preserved parser evidence did not contain a trustworthy metric: JSON 指针没有指向有限数值 · The exact held-out Pile validation split was not evaluated. · The execution covered the Mamba-790M seven-task evaluation only, not all 29 model rows declared in the workload. · The paper Markdown source was absent; the fixed PDF Table 10 was used as the comparison source. · The source table does not specify whether each multiple-choice accuracy uses raw acc or acc_norm, and the current evaluator exposes both for several tasks. · The run used an NVIDIA L4 rather than the paper's hardware; this is not material for interpreting these numerical accuracy values, but exact paper-side evaluator and dataset-version provenance remains incomplete. · The captured execution demonstrates a complete seven-task evaluation of the released Mamba-790M checkpoint and produces fresh raw outputs, but it does not establish the exact Pile validation measurement and does not cover the wider 29-row workload declared in the protocol. LAMBADA perplexity approximately matches the paper, while LAMBADA and WinoGrande accuracies differ. For the multiple-choice tasks, lm-eval emitted both acc and acc_norm, and the paper provides only an acc label; therefore the source-to-parser metric mapping is not fully established. The close acc_norm values for HellaSwag, ARC-C, and OpenBookQA make raw-acc comparisons non-discriminating. The full-row claim consequently remains inconclusive rather than supported or contradicted. Some original claim measurements remain without evidence; available measurements are assessed individually.