DRACO Design of Resilient Architectures for Computing

Dark Logic

Dark Logic research poster, from the project's earlier phase on AI-generated malicious RTL.
Dark Logic research poster, from the project’s earlier phase on AI-generated malicious RTL.

Formal verification is sound, complete, and largely unused on designs of production scale. The reason is not skepticism — it is state explosion. Model checking a block is routine; model checking the system that block sits in is frequently not attempted, because the proof does not finish. Everything downstream of that fact, including the industry’s reliance on simulation coverage as a proxy for correctness, follows from it.

Dark Logic attacks the cost directly. The work is on the techniques that shrink what a checker or a prover must actually establish:

The project reached this question from the other side. It began by asking whether generative models could insert malicious RTL that survived formal verification — and the recurring obstacle was that verification frequently did not complete at all on designs of interesting size. A defense that does not scale is not much of a defense, so the scaling problem became the project.

Open questions

Previous project
CHIRP
Next project
Dark Matter