Reviewing a benchmark paper draft: run an accuracy reviewer and a referee reviewer separately

For a systems/format paper drafted from internal notes, two independent read-only review agents with different briefs catch different classes of error. (1) An accuracy reviewer gets the sources of truth (spec, code, measurement docs) and must mark every claim OK / WRONG / UNSUPPORTED / OVERSTATED with line numbers, and recompute every derived number (totals, ratios, geometric means). It finds things like a feature described as implemented that the spec calls proposed, a table mixing figures from two code versions, or 'corrupt files are rejected' when tests show some in-frame corruption decodes silently. (2) A referee reviewer gets the paper plus the paper it is modelled on and judges methodology: unstated baseline settings (a competitor format written with a non-default or weak compression mode can produce one outlier ratio that dominates a geometric mean), which average is quoted (geomean vs size-weighted total vs median), missing ablations, and internal inconsistencies (e.g. a scan whose per-change time is lower than the stated per-value decode cost, which usually means the decode figure was for much wider values). Also: benchmark numbers for commercial EDA tools may be restricted by licence (DeWitt-style clauses); the GTKWave FST paper omitted them for that reason, so check before publishing. Tell both reviewers what is deliberately out of scope so they don't recommend it.