Reflexr decides from facts. Reflexr Vision decides from pixels. Both are small formulas found by Ravun, and both run on your device.
Reads numbers you already have: a table row, an API response, a game's state.
Each option is described by a few facts; one formula scores every option the same way, and the best score is the decision.
Reads only the screen, as a person does. No API, no game state.
The model gets only the frame's raw pixels. It picks out the few that matter, here the ball and the paddle, and one formula turns them into a button press.
Breakout, one decision, native C on one core. The language model is a published median answer time over an API.
To scale: the Reflexr's bar is 0.0014% of Jev's. Jev at its observed 0.33 s median a call, one after another; parallel calls shorten the wait, but every line still pays a network round trip. The Reflexr: single-threaded, one Apple M4 Max.
Ravun learns a short formula from your examples and tests it on cases it hasn't seen. It then compiles the formula into a tiny file you can run anywhere.
Thirteen public datasets, 100 KB to 20 MB, decisions in microseconds.
Each Reflexr learned from the dataset's labelled examples and was scored once, on held-out items. Every score runs from 0 to 1, percentages included; hover one for what it measures. Speed is the time of one decision as an application pays it, on an Apple M4 Max.
Reflexrs are almost ready. Be the first to know when you can build your own.