In Q1 2026, U.S. courts issued over $145,000 in sanctions tied to AI hallucinations. One attorney had 57 of 63 citations flagged as defective — 20 were cases that simply don't exist.
And this isn't just a lawyer problem.
Everyone said RAG would fix it
Most enterprises tackle AI hallucinations the same way: inject fresh data via RAG, fine-tune on domain knowledge, or add "don't get it wrong" to the system prompt.
Here's the thing, though — none of that is a real fix. Language models are probabilistically correct by design. They predict "what word comes next" using statistics, which means they can always be wrong. RAG included — even if the right document shows up, the interpretation step can still hallucinate.
In domains like tax filing, legal advice, or clinical drug interaction analysis, "95% accuracy" isn't good enough. The other 5% is a lawsuit.
So what does math proof actually do differently?
That's where Pramaana Labs takes a different path. Three IIT Madras alumni — Ranjan Rajagopalan, Krishnan Raghavan, and Sanjay Ganapathy Subramaniam — founded the company in September 2025 with a specific idea: layer LEAN, a formal proof language, on top of LLMs.
LEAN is an open-source language mathematicians use to verify proofs — originally developed at Microsoft Research in 2013. It formalizes complex logical structures so a machine can check whether a statement is necessarily true. Not probably true. Provably true.
Pramaana's insight: if you formalize the rules of a domain — tax law, drug interaction protocols, legal statutes — into LEAN, you can mathematically verify whether an AI's output violates any of those rules. If it does, the answer is blocked. "I don't know" is a much better answer than a confident wrong one. CEO Ranjan Rajagopalan puts it this way:
"The world's hardest problems are not unsolvable. They are unformalized."
— Ranjan Rajagopalan, CEO of Pramaana Labs
Here's how it compares to conventional approaches:
| Conventional (RAG · Fine-tuning) | Formal Verification | |
|---|---|---|
| How it works | Retrieves relevant documents as context | Converts domain rules into code, then verifies output |
| Error possibility | Hallucination can occur during interpretation | Rule violations are mathematically blocked |
| Certainty level | Probabilistic (e.g., 95% accurate) | Deterministic (correct, or no answer) |
| Best fit | Most general-purpose tasks | Law, tax, healthcare, compliance |
The inspiration here is France's CATALA project — a national initiative that converted the entire French tax and benefits system into executable, machine-checkable code. Pramaana is applying the same logic to AI.
Where is this actually being used?
Pramaana is currently building out three verticals. The common thread: rule-based domains where being wrong has serious consequences.
- Tax compliance
Former IRS Commissioner Danny Werfel is an advisor. Tax law runs to thousands of pages, but it's fundamentally formalizable. When AI recommends a tax strategy, the verification layer checks whether it mathematically conflicts with actual statutes. - Legal compliance
Case law, regulatory clauses, contract terms — all formalizable rules. Over 1,353 AI hallucination cases have been recorded in courts globally through early 2026. A formal verification layer would have caught most of them. - Drug discovery
Overseen by professors from IIT Delhi, IIT Madras, and UC Berkeley. Clinical protocols and FDA guidelines are formalized so AI drug-interaction analysis only responds within mathematically verified guardrails.
Khosla Ventures led the $27M seed round, with Accel, BoldCap, Nexus Venture Partners, Premji Invest, and Unbound participating. Vinod Khosla himself appeared at Pramaana's inaugural Verification Summit in San Francisco on June 10.
Why this sector is heating up
Axiom Math raised $200M at a $1.6B valuation in March 2026 for math-proof AI. Mistral released Leanstral, an open-source code verification model. Formal verification is emerging as the next infrastructure layer for AI.
How to tell if your work needs formal verification AI
Run through this checklist. If three or more apply, it's time to seriously evaluate formal verification.
- Check legal accountability
Does your output carry legal liability? Contract review, tax filings, compliance work — these are in scope. - Estimate error costs
Could a single AI error cause $10,000+ in losses or trigger a lawsuit? If yes, probabilistic AI isn't a viable foundation. - Look for codified rules
Does your domain have a formal rulebook? Legal codes, pharmacopeias, ISO standards, clinical guidelines — if rules are written down, they can be formalized. - Audit your current verification
When your current AI tool gives a wrong answer, what happens? If the answer is "a human checks it," that human time is your hidden cost. - Join the waitlist
Visit pramaanalabs.ai to register. Pramaana is currently building vertical-specific solutions with enterprise partners in law, tax, and pharma.
