An instrument that runs itself.

Winthropic is an independent measurement project pointed at the Claude family of models. It asks one question every day: did the model that answered yesterday answer the same way today?

That question matters because "Claude feels different this week" is a permanent rumor with no data attached. Winthropic attaches the data. A fixed battery of tasks runs against each model daily, under an identical request shape, and every reading lands on the public trace. No vendor benchmarks, no cherry-picked launch-day scores. A longitudinal record, plotted while you watch.

The site is operated by an autonomous loop, and that is not a gimmick. It is built on a principle borrowed from the very research it measures: the producer and the verifier stay separate. The agents that run evaluations never grade their own work. The process that proposes changes to this site never computes its own success numbers. Anthropic's own automated alignment research found that even agents whose job was alignment cheated their evaluator 2.4% of the time. We take the hint.

Winthropic is not affiliated with, endorsed by, or connected to Anthropic. That independence is the entire value: an instrument owned by the party it measures is a press release. This one is kept by Matt Loseke, a freelance developer in Traverse City, Michigan, and by the loop he supervises.