Pramaana wants AI to prove it, not guess
Pramaana Labs came out with a 27 million dollar seed round led by Khosla Ventures, with Accel, Nexus Venture Partners, Premji Invest, BoldCap and Unbound joining. The San Francisco startup is building a verification layer for AI, and the pitch is one line: in industries where being wrong has real consequences, AI shouldn't guess, it should prove its work.
The mechanism is the interesting part. Pramaana pairs a normal LLM with deterministic formal verification, using LEAN, the proof language mathematicians use to machine-check theorems. So instead of trusting the model's confident-sounding answer, the system tries to produce an actual proof that the answer is correct. Proof replaces probability. They're aiming it at law, drug discovery and tax, with former IRS commissioner Danny Werfel advising on the tax side and professors from IIT and Berkeley on the rest.
This lands right on top of the week's biggest agent anxiety. Agents' Last Exam showed the best agents fail three out of four real economic tasks. Everyone agrees models are smart and still can't be trusted to be right. Pramaana's bet is that the fix isn't a bigger model, it's a verification layer bolted on top that refuses to ship an unproven answer.
The honest caveat: formal verification works beautifully for things you can formalize, like math and certain legal logic, and gets hard fast for fuzzy real-world judgment. But the fact that Khosla is funding a proof-based reliability layer at seed tells you the market now sees correctness, not capability, as the thing worth paying for.
← Back to all articles
The mechanism is the interesting part. Pramaana pairs a normal LLM with deterministic formal verification, using LEAN, the proof language mathematicians use to machine-check theorems. So instead of trusting the model's confident-sounding answer, the system tries to produce an actual proof that the answer is correct. Proof replaces probability. They're aiming it at law, drug discovery and tax, with former IRS commissioner Danny Werfel advising on the tax side and professors from IIT and Berkeley on the rest.
This lands right on top of the week's biggest agent anxiety. Agents' Last Exam showed the best agents fail three out of four real economic tasks. Everyone agrees models are smart and still can't be trusted to be right. Pramaana's bet is that the fix isn't a bigger model, it's a verification layer bolted on top that refuses to ship an unproven answer.
The honest caveat: formal verification works beautifully for things you can formalize, like math and certain legal logic, and gets hard fast for fuzzy real-world judgment. But the fact that Khosla is funding a proof-based reliability layer at seed tells you the market now sees correctness, not capability, as the thing worth paying for.
Comments