One Jev call, laid out as a town, with the fields shown
Decision Depot is one System One call drawn as a town. A van carries a support ticket from App Yard through the System One Gate, State Bay, and the Question Forge. At the forge, Choice, Score, and Noul are filled together. The van then visits each booth so you can read the card, crosses Wire Yard, and either turns into Review Bay or drives straight to Act Gate. The lamps at the end are your if statements.
The bars on the van are the vector it is carrying: keyword counts before the call, then the distribution of the booth it is reading. The crate is the state string. The three cards in the bed appear at the same stop, because one response returned all three.
Mandated shape, computed live in your browser: the request is model, state, and questions. The three questions share that state and are scored inside one systemOne() call, not three. Choice returns choice, probabilities (they sum to 1), and confidence. Score returns score, probabilities, legend, and confidence, and the score is the probability-weighted level index, so it can fall between rungs. Level count is kept inside 2–10. Noul returns noul in [0, 1] and no confidence field. The response is model, answers keyed by your names, and usage. applyPolicy() is ordinary code: it thresholds noul and choice confidence. The model does not return an action. All of that lives in js/model.js.
Scaled down: four Choice options instead of the API maximum of 255. One string of state instead of an object or array, both of which the API accepts. Five score levels unless you drag the slider. One call, not a batch.
Assumed: the four toy tickets; the footer the Trim toggle appends; a review cutoff of 0.50 and an escalate cutoff of 0.70 (the same figures appear as examples in the TypeSafe docs, and they are still your policy, not Jev’s); the word lists and the logit weights; token counts of characters divided by 4 on the way in and 12 per question on the way out. Those token numbers are marked “est.” on the panel.
Pedagogue fake: real Jev is a trained model at POST https://api.typesafe.ai/v1/systemone. This page never calls it, and it has no API key. Logits are a weighted sum of keyword counts, then a softmax or a sigmoid. Choice and Score confidence is one minus normalized Shannon entropy. TypeSafe derives confidence from the distribution and does not publish the statistic, so do not quote these confidences as Jev’s. The three booths are visited in series so each card can be read; the call itself is parallel, which is why the three cards appear together at the forge. Treat the probabilities as a worked example of the wire. Treat the shared state, the typed answers, and the code-owned branch as the lesson.
Scenery: the sheds, the trees, the road paint. The post heights, the ramp bead, the dial needle, the crate, and the lamps are the numbers. The rest is the town those numbers needed somewhere to happen.
The first time the van reaches a station it stops for long enough to read that station’s write-up — between 9 and 26 seconds — and the bar under the text shows how much of the stop is left. Once every district on the trip has been explained, later runs go faster. Space holds a stop, S steps one station, and Speed scales everything, reading stops included. Reset (⟲) replays the slow tour. Run keeps what you have already read.
Built from the isometric-explainer skill. Wire shapes follow the Jev brief and docs.typesafe.ai. The stub is original.