Recur

About this project

Recur explores what happens when repeated interactions with an AI system can be turned into reusable computation. Instead of sending the same type of request to a language model indefinitely, Recur collects demonstrations and looks for recurring task patterns that can be expressed as small programs.

Its compiler searches a typed dataflow language composed of readers, transforms, and emitters. Candidate programs are executed against the original demonstrations, allowing the system to test whether a proposed computation actually explains the observed examples. Programs are then subjected to multiple verification stages, including consistency checks, leave-one-out recompilation, generated in-domain inputs, explicit decline behaviour, and a lightweight neural acceptance head.

Once a program has been accepted, it is stored in a registry together with its demonstrations, verification results, and runtime information. Future matching requests can then be executed locally through Recur's deterministic virtual machine without another model call. The system keeps language generation separate from computation, using model output as evidence for synthesis while requiring computed values to come from the verified program itself.