Nov 8 5 days ago 08:35 correct the qev media c=1 cells after re-measurement and simplify the warm-up wording Nov 6 7 days ago 10:27 name the A100 in the same shape as the RTX 4090 10:25 27b tier: state the measurement condition first, then the weights-vs-graph-pools memory reason 10:19 state the memory cause of the 27b eager choice 10:03 9b/27b performance and accuracy on A100 80G vs qev Oct 25 19 days ago 07:07 Report qev c=64 media cells as n/a: zero in-window completions is no measurement, not a measured zero; recovery-run numbers stay as a footnote reference 03:56 27b re-measured with the same duration-window protocol (A100 80 GB); qev c64 media cells completed zero in-window, drain numbers footnoted 02:21 Duration-window numbers for 0.8b/4b/9b (10s warm-matrix protocol); 27b kept pending A100 re-measure Oct 24 20 days ago 08:24 Review pass: latency ratio columns, files-less snapshot layout, license wording 08:14 OneJev 1.6 family: four variants (0.8b/4b/9b/27b), manifests only -- checkpoints fetch from the official OmniJev repos on first serve 
