01
Asynchronous schedulers
Work-stealing queues, task runtimes, and completion loops. We prove the scheduler cannot deadlock, and that a task which is ready runs under the assumptions you state.
Prove the concurrent code.
Mathematical proof that asynchronous schedulers, trading algorithms, and multi-threaded environments are free of deadlocks and race conditions.
Discuss the systemWhat we prove
02
Trading algorithms
Order matching, cancels, and the book. Two events at once must not drop an order or leave a price the rules forbid. The proof covers the interleavings, not one busy morning.
03
Multi-threaded code
Locks, atomics, and shared state. We prove the locking order cannot deadlock, and that no write races a read of the same location.
Why tests miss it
A test runs one schedule.
Production runs another. A deadlock or a race that needs a particular interleaving will not appear because someone thought to test for it.
The prover checks the rule against every interleaving the specification allows. If a deadlock or a data race is possible, the proof does not go through, and the code does not ship.
How an engagement works
-
Specify
The events, the shared state, and what must never happen: a deadlock, a data race, or a book left in a state the rules forbid.
-
Model
The scheduler, the algorithm, or the threads, including the locks, the wake-ups, and the order in which events may arrive.
-
Prove
AI proposes the proof. A theorem prover checks every interleaving the specification allows, not the schedule a test happened to run.
-
Hand over
The proof, the assumptions it rests on, and scripts you can re-run when the code changes.
The proof covers the scheduler, algorithm, or threaded code in the specification, and it records the assumptions it rests on. Code outside that specification is outside the proof. When the fault is a retry or a cancellation across several systems, that work is security and access correctness. Trading systems sit with the fintech sector. The method is vericoding, and LeanIn is the concurrency runtime whose model we have machine-checked.
Discuss the concurrent system
Tell us the scheduler, the algorithm, or the threaded code, and the failure you cannot afford.