mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
Statistics, evidence and the manifest are typed values in hanzo-bench (`board`), computed from the raw samples alone. The ledger, a rounded view of the evidence, is gone with the Python scripts; the site reads the evidence. The measuring side builds its model with ServerBuilder, so a run measures what a server gets and can draft with --dflash or --mtp-model, and it exits before the ROCm runtime's teardown, which aborted after the samples were written. Co-authored-by: Hanzo Dev <[email protected]>