Research, perspectives, and product updates from the Meerkat team.
Today we deployed Meerkat's deep analysis engine to production. Five domain-adapted verification models running on GPU, catching the subtle errors that signal-based checks miss.
Geoffrey Hinton reframed AI hallucinations as confabulations. That reframing has profound implications for how we build verification systems.
A model that is 95% accurate produces 50 errors per 1,000 outputs. At scale, those errors have names, consequences, and liability.