By: ValentinHirschi <valentin.hirschi@gmail.com>
Re-measure the binary64 masters and time three more top-loop diagrams Independent re-measurement of the 2026-10-17 numbers on another pinned CPU, with the OneLOop drivers rebuilt from the same sources, two interleaved rounds and the tensor fix in place. The gg -> HHH pentagon is 1.15-1.20x OneLOop per point (1.21-1.24x before), and the mixed scalar fixtures and tensor rows reproduce the earlier ratios within a few per cent. tests/generate_bench_rows.py gains hgg, aaaa and gghh: Ninja's requests for the H -> gg top triangle and the gamma gamma -> gamma gamma and gg -> HH top boxes. H -> gg (0.58-0.62x) and gg -> HH (0.98-1.05x) pass, but gamma gamma -> gamma gamma is 1.60-1.72x per diagram: its massless-leg boxes D0(0,0,0,0,s,t) take 2.0-2.15x OneLOop's time and its triangles C0(0,0,s) 1.8x. Record the tables in performance/2026-10-17-claude/review and correct the status line in the README. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>