Exeter, UK · Founder, Arbiter Security

Laurence Avent

I build AI-native offensive security tools in Rust and study how proof assistants can check the reasoning produced by language models.

W(3,3) in Lean 4 ✓ verified
theorem vdw33_characterization (n : Int) :
  (∀ c, IsColoring n c → ¬ APfree n c)
    ↔ 27 ≤ n
382,964 proof steps See the proof

What I build

Venture

Arbiter Security

AI-native offensive security tools, built in Rust. Two products exposed as MCP servers: Arbiter (web-application vulnerability detection across 52 classes) and Aletheia (binary analysis across 14 CWE classes).

Civic data

Tightrope Tracker

An open-source live dashboard of UK economic constraint, drawn from official OBR and ONS sources, with public methodology and codebase.


Research

Reasoning that carries its own proof

A working system in which untrusted generators, language models among them, propose reasoning steps, a small kernel enforces the rules, and every accepted chain is re-checked byte-for-byte and re-proved in Lean 4. Its first case study is a determination of the van der Waerden number W(3,3); the papers and artifacts are on the research page.

Framework

Avent's Razor

Judge a reasoning system by the constraints it can enforce during search: enforcement prunes the space and, unlike insight, can be checked by machine.

Systems

Verified Chain-of-Thought

An 18-rule kernel, two independent byte-level checkers held to exact agreement, and a Lean 4 oracle that re-proves every step. Language models operate the loop through a typed gate.

Case study

W(3,3) = 27

Both directions machine-checked: a 382,964-step certificate covering the entire search, compiled to Lean and reduced to a single biconditional over colourings.


Contact

Get in touch

Open to conversations about security engineering, coordinated disclosure, and verified reasoning. Email is the fastest route.