Jev, Robotics, and the Edge: The Smallest Category With the Longest Fuse

Housekeeping first, because this page will not pretend otherwise: robotics and devices is the smallest category in the directory, six builds at the time of writing against the hundreds elsewhere. We're writing about it anyway, because small-and-early is exactly when honest documentation matters most, and because the physics of the argument is strong even while the catalog is thin. Consider this the "watch this space" page, with the watching done rigorously.

Why a decision model points at hardware at all

The game builds established, publicly and unfakeably, that Jev sustains roughly ten structured verdicts a second inside a live loop. A game loop and a control loop are cousins: observe state, decide, act, repeat, under a deadline. The moment a general model fits inside any real-time budget, the question "could it sit in a supervisory control seat?" stops being science fiction and becomes an engineering negotiation about which loops.

The honest negotiation: not the fast ones. Motor control, balance, collision reflexes run at hundreds of hertz on local firmware and always will; a network-hop verdict has no business there, and the limitations page applies with safety multipliers. The plausible seat is the deliberate layer above reflexes, the one currently occupied by brittle state machines: which task next, is this object the target, is the workspace state safe to proceed, does this sensor pattern warrant stopping. Ten-per-second is luxurious at that layer, and "small closed-set judgment" is precisely its native format.

What the early builds actually show

The cataloged experiments cluster into three honest signals. Supervisory verdicts on real devices: task-selection and safe-to-proceed rulings driving simple hardware, state machines with the brittleness swapped for judge questions. Sensor-stream triage: classifying event streams from devices (anomaly or normal, which alert tier), which is support triage with accelerometers, and probably the most immediately deployable pattern here. The text-only perception trick: the screenshot-free computer-use build, where local OCR turns a screen into text so the model never needs vision, matters enormously for this category, because "cheap local perception feeding a fast text-verdict layer" is a template that generalizes from screens to sensor summaries. Numbers and behaviors as reported by build authors, per house rules, and with six builds we resist extrapolating curves from a scatter plot.

The seatbelt paragraph (non-negotiable)

Everything the agent-verification page says about irreversible actions applies here with mass and momentum attached. Verdicts advising a human or gating within hard-coded safety envelopes: reasonable today. Verdicts as the last word before physical motion: no, and not because the model is slow, but because nondeterminism plus kinetic energy is a product-liability seminar. The builders in this category so far have been admirably conservative, which we suspect is why they're still in this category.

What would change our assessment

We'll upgrade this page from "long fuse" to "burning" when the catalog shows: sustained deployments past demo length, verdict layers surviving real sensor noise, and anyone publishing failure data with the same candor as success reels. Submit exactly those, please, the unglamorous ones especially. Six builds is a scatter plot; sixty with receipts is a field.

Frequently asked questions

Can Jev control a robot?

It can plausibly occupy the supervisory decision layer (task selection, safety-state verdicts, sensor triage) above local firmware reflexes; it cannot and should not run low-level real-time control, which stays on-device by physics and by sense.

Is anyone actually using it with hardware?

A handful of cataloged early experiments: supervisory verdicts on simple devices and sensor-stream classification. Real, interesting, and too few to generalize from; that's the honest count.

Does Jev run on-device?

Access runs through TypeSafe's platform per docs.typesafe.ai; the cataloged pattern pairs local perception (OCR, sensor preprocessing) with remote verdicts. On-device inference would change this category's ceiling entirely; we'll report it when the receipts exist.

What's the realistic first industrial use?

Sensor and event-stream triage: high-volume, text-representable, tolerant of occasional escalation, and identical in shape to the triage patterns already proven elsewhere. Boring first, ambulatory later.