Ink - Behavior JIT: turn proven AI decisions into software.

Production AI agents make the same bounded decisions thousands of times: which tool to call, where to route a ticket, whether to allow an action. Teams pay a frontier model for every one, even after the answer has been verified over and over. Ink is a Behavior JIT. It sits around your existing model call, records decisions and their real outcomes locally, and compiles the behavior that has proven itself into software. Everything else keeps going to your model. What’s different: Permission, not just prediction. A local Fast Path only serves once its 95% Wilson lower bound clears your error budget on independent holdout data. Confidence and similarity never grant authority. Knows when to abstain. Novel or ambiguous requests fall back to your model automatically. Revokes itself. Live comparison traffic keeps auditing, and if error rises, authority goes back to your model. Sub-millisecond, local. Under 0.01 ms for Exact and under 0.1 ms for Linear on local benchmarks, with no network calls and no tokens. Embedded and local-first. Runs in-process on SQLite, with no cloud control plane. pip install ink-jit, Python 3.11–3.13. Honest numbers: in controlled internal evaluations, Fast Paths served 42–78% of traffic with 0.0% holdout error. External design-partner validation is still in progress, and if your workload doesn’t repeat, Ink saves nothing.