The gap between AI marketing and AI reliability has become its own investment category. Pramaana Labs closed a $27 million seed round from Khosla Ventures to bring formal verification, the mathematical technique used to prove software correctness, to AI outputs in high-stakes domains: law, drug discovery, tax. The same week, Fast Company noted that AWS is aggressively marketing autonomous agents while quietly building extensive guardrail infrastructure. Both stories tell the same story from different angles: the industry knows the reliability problem is real and is now pricing it accordingly.
Formal Verification as a New Infrastructure Layer
Formal verification has existed in software engineering since the 1970s, used to prove the correctness of critical systems in aviation, finance, and semiconductor design. Applying it to probabilistic AI outputs is genuinely hard, which is why the $27M seed is notable. TurboFund's list of active AI seed investors shows how crowded this space has become, but Pramaana's bet is on a specific infrastructure layer: trust, not capability. That is a different and probably more durable market than raw performance benchmarks.
The Academic Stakes of AI Errors
The arXiv paper "Towards Auditing AI Systems in the Wild" by Vadlamani, Srinivasan, and Parthasarathy frames the problem precisely: AI system behavior in deployment is shaped by dynamic real-world conditions that benchmarks don't capture. Pramaana is betting that high-cost error domains, where a wrong answer in a legal brief or a drug interaction calculation has material consequences, will pay a premium for provable correctness. The G7 summit brought OpenAI and Anthropic to the table with world leaders this week. Governance is catching up to capability. Formal verification might be where engineering catches up to governance.