Global evidence dashboard

Qualified benchmark statistics, once a genuine public source is authorized. Today the dashboard shows its structure and honest empty states — nothing estimated, nothing filled in.

Source: unavailable.

Summary

Charts

Reading this dashboard

Every statistic belongs to one exact cohort (model, build, runtime, workload, configuration all equal). Cohorts are never averaged together.
Throughput means tokens per second of measured on-device generation. Whiskers span observed min–max, not error bars.
n counts qualified runs; i counts contributing installations in that cohort. Installation counts cannot be honestly summed across models.
Below-quorum cohorts show counts without statistics. Synthetic test data is excluded from every figure on this page.
This dashboard is not live: it refreshes only on request, and only from an explicitly authorized public endpoint.