Key Takeaways

  • Current AI focuses on recall, but the next frontier is provable accuracy, especially in domains where a wrong answer is catastrophic, like tax, legal, and healthcare.
  • AI hallucinations are already causing real-world harm; Ron John points to the Australian government suing a "Big Four" firm over an AI-induced error.
  • Pmana Labs, co-founded by John, converts complex English-expressed knowledge, like the US tax code, into Lean, a mathematical proving language.
  • This approach yields answers that come with verifiable mathematical proofs of correctness, moving beyond the probabilistic nature of most AI models.
  • To build truly trustworthy AI for high-stakes applications, founders must integrate a rigorous methodology like Pmana Labs' Four-Layer Autoformalization Process.

The Pmana Labs' Four-Layer Autoformalization Process for Provably Accurate AI

Ron John and Vinod Khosla discuss the urgent need to move past probabilistic AI. John outlines Pmana Labs' process for achieving provable accuracy, which shifts AI from a guessing machine to a precise tool capable of verifiable outputs.

Layer 1: Offline Formalization

We are formalizing the US tax code... converting that into lean which is a proving language. So when we so uh when we do that we get it ratified by experts.

Layer 2: Constraint Conversion

whenever a question pops up we convert it into a series of constraints. ... So we ensure that we ask the right set of questions so that ultimately we can give you a reliable answer.

Layer 3: Solver and Prover

Post that we have a solver and prover working in tandem to ultimately give you an answer along with a proof of correctness.

Layer 4: Proof of Correctness

The proof is something which a mathematician trusts. That's supposed to give you more uh comfort in a way. And like you suggested uh you're right in saying that even like best humans might make mistakes. uh but if you actually have a mathematical proof encoded in lean you can choose to trust it and that's the frontier which we are aiming for

When This Works (and When It Doesn't)

This method truly shines “in very specific mission critical domains like tax, legal, healthcare and governance where a wrong answer could be catastrophic.” John emphasizes the stakes: “a wrong answer could be catastrophic.” Think about automated legal advice, medical diagnostics, or financial compliance where even a slight error can lead to millions in fines, lawsuits, or worse. The process isn't meant for general-purpose tasks or creative AI, where ambiguity or probabilistic responses are acceptable, or even desired. If you're building a content generator or a customer service chatbot, this level of formalization is overkill. But if your AI's output can directly impact someone's freedom, health, or fortune, the Pmana Labs approach becomes essential.

What to Do With This

If you're building an AI product that operates in any high-stakes domain – say, automated contract review for a law firm – you need to move beyond typical probabilistic large language models. This week, pull your core domain knowledge. For contract review, that might be a specific type of contract (e.g., SaaS agreements) and relevant legal precedents. Apply Pmana Labs' process: first, actively formalize these legal rules and contract clauses into a mathematical proving language like Lean (Layer 1). This is hard, expert-driven work. Second, design your query interface to convert user questions (e.g., "Is clause X compliant with regulation Y?") into precise, formal constraints (Layer 2). Third, integrate a solver and prover that can process these constraints and yield a definitive answer (Layer 3). Finally, demand that every answer includes a verifiable mathematical proof of correctness (Layer 4), a critical audit trail that establishes trust and protects against costly errors. Your users need certainty, not just a high probability.