Answers are grounded in this paper and cite their sources.
Performance numbers can reproduce without establishing the comparison attached to them. We audit published language-model inference assertions through six source-specific reconstructions and a fixed probability sample. The workstation case reproduces 576 saved statistics and 320 displayed entries, yet nine of twelve comparator elapsed-time variability values contradict a stated threshold even after display rounding. Other cases quantify a disclosed timing-window sensitivity, support a memory estimator despite a component increase, reconcile a latency headline with saved outputs, restrict a placement observation to its tested models, and identify inconsistent model attribution. The separate 150-family cohort contains 59 eligible, 67 excluded and 24 unresolved families after documented corrections; 54 assessed assertions comprise 23 supported within scope, one source conflict and 30 unresolved. These are single-author source assessments, not independently validated error rates. We supply calculation inputs, source locators and a correction record. The contribution is a reproducible account of what particular evidence supports, not physical replication, a field-wide error estimate or a new benchmark. Preprint submitted for peer review.
No takes yet. Share an insight, caveat, or question.
Pranesh S (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: