Bench: retest sweep of every VERIFY cell
Last verified: 2026-08-19
The method below is fixed and published. Timers read Not yet measured until the run completes.
A VERIFY flag is a promise, not a hedge. This entry closes the open flags on the scoreboard and records every score change it causes. Desk-drafted from the bench brief. Method is fixed and published below. Numbers land here after the run, and until then every timer on this page reads Not yet measured.
Which cells are flagged VERIFY today?
One: Buzz model flexibility at 6.5. The evidence note reads that agent model configuration in preview docs is thin and that supported providers need repository review before the score moves.
What else sits in the verification queue?
The Buzz license file name, the BAND corporate entity and any security attestations, and the two time-to-real-work timers. Attestations that a vendor does not claim stay unscored, not assumed absent and not assumed present.
How do score changes ship?
Same day as the entry that caused them, with the before and after values, the reason, and a changelog note. Corrections are made in place with a dated line, never silently.
Related
- Time to first agent reply, Buzz (Block-hosted)
- Time to first multi-agent room, BAND (Free tier)
- The 24-hour test: what survives in a BAND Free room
- Buzz vs BAND verdict
- How we test
Method fixed 2026-08-19 · DESK REVIEW · PROVISIONAL