Community · E2 · artifact verified

Steer a hospital delivery robot past four hundred people

A hospital delivery robot judged about four times a second: code samples twelve candidate paths and removes predicted collisions, Jev picks one among the survivors, and a deterministic safety brake guarantees no contact - built at the TypeSafe JEVATHON.

01 · Role in the system

What Jev does here

The browser floor holds wards, doors, four hundred moving people and thirty carts or beds, with the robot heading from Pharmacy to ICU bed 4 on a STAT delivery. Code samples twelve candidate three-second paths and discards any predicted to collide; Jev then picks one from the survivors with a probability per option, about four times a second. A deterministic safety brake - code, not a model - guarantees no contact. The author's measured run on seed 42: delivery in 84 seconds against a DeepSeek baseline's 115, with about 126 millisecond p50 decisions at roughly six hundredths of a cent each; with 400 people the LLM was still forty meters short when the recording ended, and the validator rejected its answers that named unoffered paths. A split-screen Brain Race replays both engines on the same floor.

Code filters the paths, the model picks among survivors

The safety architecture runs the right direction: code samples twelve candidate three-second paths and discards any predicted to collide, Jev chooses one among the survivors about four times a second with a probability per option, and a deterministic safety brake — code, not a model — guarantees no contact regardless of what any engine answers. The model's freedom ends where collision begins.

The measured run makes the comparison concrete: on seed 42 with 40 people, delivery in 84 seconds against a DeepSeek baseline's 115, at about 126-millisecond median decisions costing hundredths of a cent. The baseline detail worth reading: with 400 people the LLM was still forty meters short when the recording ended, and the validator rejected its answers that named unoffered paths — a failure mode the choice-based engine cannot have.

A hackathon build that cites its own scale

The record names its nature: a one-day JEVATHON build in a simulated browser hospital — no real hospital, robot hardware, or deployment — with headline numbers author-measured on one seed against one author-chosen baseline. It extends JevPilot, borrowing its motion and path-sampling machinery.

Within that frame, the transferable idea is the candidate-filter-then-choose shape: enumeration and collision filtering in code shrink the model's decision to exactly the judgment code cannot make, which path among safe candidates — and every engine, judged or generative, answers inside the same validator and the same brake.

02 · Control boundary

Where Jev sits

Candidate paths enumerated and collision-filtered by code; one choice per tick among survivors; the safety brake and validator stay deterministic regardless of engine.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • A one-day hackathon build in a simulated browser hospital; no real hospital, robot hardware, or deployment is involved.
  • Headline numbers are author-measured on one seed against one LLM baseline chosen by the author.
  • Extends JevPilot; the decision layer assumes its motion and path-sampling machinery.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.