🏎️ Hairpin · two cars, one track, one memory

Two race cars lap the same track. At every section each driver makes one live call to the Jepela API: the same question and the same view of the road, with five ways to drive it. Some sections hide a surprise that a driver only learns by crashing. Car A has no memory, so it meets every surprise as if for the first time. After each of Car B's crashes the game writes one line to its Jepela memory, naming the way through; watch Car B's lap times fall while Car A crashes in the same places, lap after lap.

laps Starting…

Car A · no memory

Car B · Jepela memory

Race feed newest first

Press Race, or step one section at a time.

Lap board

Car A, no memoryCar B, Jepela memorylap time, seconds

Car B's memory lines and rules

The track incidents this race

Car B's last API call, word for word

No call yet.

Model jepela-english. Nothing is scripted: both drivers ask the same engine the same question with the same words, and only Car B names a memory. The track and what it hides are game code; every choice is the engine's. Car B's calls read only the lines that share words with the section's own description (the question's words are left out of the search, and a line needs at least half the best line's score), so each section brings back its own lesson. Measured on 2026-09-28: without memory the engine keeps full throttle at every section (33 to 50%), which crashes at the Hairpin (brake early 7%), the Esses (lift off 16%), the Crest (lift off 20%), the Chicane (inside line 18%) and, once the oil is down, the Back Straight (outside line 14%). With its five lessons in memory at once, Car B took the safe way every time: Hairpin 98%, Esses 92%, Crest 92%, Back Straight 85%, Chicane 67%. A five-lap race: Car A 89.1 s, then 95.5 s every lap; Car B 89.1 s, 68.0 s, then 62.0 s. In a second race, with its memory kept, Car B drove 62.0 s from lap 1.

Engine answering · API health check, 09:09 UTC