← back to the simulator

A32X with Jev at the flare

A landing comes down to when the flare begins. In the A32X flight simulator its scripted crew begins it when radio altimeter 1 reads below 33 ft. Here TypeSafe's Jev, a fast judging model, decides instead which of the aircraft's four height readings to believe, and the flare begins when that one goes below 33 ft. The simulator is the A32X 0.1.6 release, unchanged; this page drives it from beside it, through the debugging hooks the release exposes for exactly this experiment.

Who decides what

What was measured

An A/B on the simulator's own judge (a stabilised approach from 1,000 ft, then touchdown under 360 fpm and 2.6 g, 300 to 900 m past the threshold, on the centreline, no bounce): the ILS to Dallas Fort Worth 17C, ten winds per case, the trigger the only difference. Native builds, landings flown in lockstep, latency modelled from each call's measured round trip.

case (10 winds each)today's formulaJeva code voter
All instruments working8/108/108/10
Radio altimeter 1 fails8/108/108/10
Radio altimeter 1 reads 15 ft high0/108/108/10
Radio altimeter 1 reads 10 ft low0/106/106/10
Both radio altimeters read 15 ft high0/107/107/10
Both radio altimeters read 10 ft low0/104/104/10
Both radio altimeters fail0/108/109/10
Ground 20 ft above the runway before the threshold0/108/108/10

STABLE landings out of ten on the simulator's judge, the trigger-height law flown after the trigger. The two misses on ordinary approaches are the strongest headwinds, where no flare begun at 33 ft lands STABLE. On the simulator's default law every case where a radio altimeter reads wrong was 0/10 whoever triggered. Measured 2026-11-02; 31,717 Jev calls, $1.65.

Where Jev does not help: on ordinary approaches it lands exactly as the formula does, the same feet; a twenty-line code voter that groups the readings that agree and trusts the group with the most kinds of instrument did as well as Jev in every case; and asked from the camera alone, Jev could not judge the height precisely enough to time the flare. Where it does: when a radio altimeter lies or fails, today's formula has nothing else to go by, and Jev, given every reading, picks the right one.

Cost, limits, data

A landing costs about $0.004 of Jev (about 90 calls of about 1,150 tokens). This site's Jev spend is capped at $2 a UTC day; at the cap the page flies by today's formula and says so. Sessions are limited per address and per day. The calls carry numbers from the simulator, nothing about the visitor; TypeSafe may keep request data for telemetry and fraud monitoring, not for training.

Sources and licences

The simulator under /sim/ is the A32X 0.1.6 release's web/ files, byte for byte, with its licence and notices (A32NX, JSBSim, three.js, Emscripten and the rest beside them in /sim/notices/): free software under the GNU GPL version 3; its Corresponding Source is the release's source archive. The page beside it (driver.js, fuse.js, packs.js) is served here as written. It belongs to the Jev Park showcase, which is not affiliated with or endorsed by TypeSafe. The simulator is not affiliated with Airbus, FlyByWire Simulations or Aerowinx, and is not for type training.