mirror of
https://github.com/kevinveenbirkenbach/computer-playbook.git
synced 2026-11-11 13:17:07 +00:00
kevinveenbirkenbachandClaude Opus 5
478d3ac0b5
A cold LibreTranslate answers its opening requests well above the rate it
settles at. ac5e7d54c4 measured each pair once, so the bonus landed entirely on
the first cell and it won every sweep: the grid reported batch=10 lanes=10 at
8.45 entries/s and called it ahead of the field, while measuring that same pair
again right afterwards gave 5.7, 6.4 and 6.0.
A discarded warm-up run was not enough. Repeating a pair back to back keeps all
of its runs inside the same opening window, so the median could not see past
it. The grid is now walked once per pass and the passes repeat, which puts the
bonus on one run of one pair: the same cell reported 7.7 in pass 1 and 5.9 in
pass 3.
Each cell records the distance between its best and worst run, and the winner
counts as a real difference only when it beats the field's median by more than
the widest of those. The rebuilt sweep answers 5.91 entries/s, inside the
noise, which agrees with the independent re-measurement and is the honest
verdict for this host: the client settings are not what limits it.
The sample halves to 200 entries so three passes cost what one used to.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>