What happened
Probably, founded and led by CEO Peter Elias, raised $9 million in seed funding with Andreessen Horowitz as sole investor. The company aims to build AI that prevents hallucinations from reaching users, targeting 99.99% accuracy. Its first product is a data-science tool that produces answers with citations and audit trails.
The system checks an LLM's first-pass answers against a deterministic validator, bouncing any result that does not match the underlying dataset. Because the validator does the accuracy work, Probably says the setup can run on a model four classes weaker than frontier models, cheap enough for local hardware. It plans to extend the approach to accounting, medical and other precision domains.
Why it matters
Hallucination remains the core barrier to using LLMs in domains where a wrong answer carries real cost. Pairing a weaker model with a deterministic validator is a pragmatic way to trade raw model power for verifiable accuracy.
Running on cheaper hardware also makes reliable AI more accessible for smaller teams and on-premises deployments.
MintedBrain take
Deterministic validation only covers what it can check against a known dataset, so the guarantees are scoped, not absolute. Look closely at where the validator applies and where it does not before trusting the 99.99% figure, and treat citations and audit trails as the real value here.
Discussion
Sign in to comment. Your account must be at least 1 day old.