LLMs + Proof Irrelevance: A Fix for Formal Verification?
28 Jul 2026
The proof-automation bottleneck
Dependently-typed programming languages—the kind that let you encode correctness properties directly into your type system—have stayed extremely niche for one core reason: the proof effort overhead is brutal. According to the report, the seL4 project, a well-known formally verified microkernel, spent about 10 times as much time proving as it did designing and implementing the system. It also ended up with more than 20 times as many lines of proof code as C code. That ratio has historically scared most teams away from dependent types entirely.
A new claim: LLMs change the math
The report's central argument, published 26 Jul 2026, is that large language models combined with a technique called proof irrelevance could become "an extremely capable form of proof automation for dependently-typed languages." If true, this would directly attack the cost structure that has made dependent types impractical for most production use—automating the very proof-writing labor that made seL4-style projects so expensive.
As evidence, the author describes building a Zstandard decompressor in Lean and shares a personal take on the process: "Doing proofs is actually quite fun: it's challenging, interactive, and there's a clear goal." That's a subjective anecdote, but it's offered as a data point suggesting proof work is well-suited to the kind of interactive, goal-directed problem solving LLMs are increasingly good at.
Why founders should care
- Formal verification could become cheaper. If LLM-assisted proof automation matures, the historically prohibitive proving-to-coding time ratio (roughly 10:1 per the seL4 data) may shrink—though the report gives no specific projection for how much.
- Faster shipping cycles are plausible, not guaranteed. The report suggests this shift "may indicate faster shipping cycles for teams adopting the new tools," but this is a directional signal, not a confirmed outcome.
- This is a single-source claim. The report explicitly flags that it is awaiting corroboration—there's no second source, competing data set, or independent benchmark cited here. Treat the automation potential as an early, unverified thesis rather than an established trend.
- Niche today, optionality tomorrow. For founders building in security-critical, safety-critical, or correctness-critical domains (infra, crypto, compilers, embedded systems), cheaper formal verification could eventually widen the set of viable tools beyond current niche adopters—if the automation claim holds up under scrutiny.
The bottom line
The report frames this as an opportunity worth watching rather than a settled fact: LLMs paired with proof irrelevance could dramatically lower the barrier to dependently-typed programming, historically gated by punishing proof-to-code ratios like seL4's 10x time and 20x line-count overhead. No conflicting sources or risk factors are cited, but with only one source behind the claim, founders evaluating formal-verification tooling should treat this as an early signal to monitor rather than a proven inflection point.