Probably, a startup backed by a16z and Accel with $9M in seed funding, is tackling AI hallucinations by wrapping weaker, cheaper models in a deterministic validation 'harness' rather than scaling up model size. Its first product is a local verifiable data agent built on DuckDB that runs each answer through a validator, checks it against actual data, and only ships results with citations and audit trails. The approach targets cost-conscious enterprises by running on desktop hardware instead of data centers, while keeping raw data local for privacy. The product is in public preview at v0.1, and the claimed 99.99% accuracy is still a goal rather than a proven result.
Table of contents
A harness, not a bigger brainAimed at the token-cost backlashA provocative claim, and the catch168 Impressions