Waits are measured over the same jobs each racer has handled so far. The clock stops while you choose, so your reaction time doesn't count.
Inside the experiment
Each job has to go into one of six lanes, and the lanes clear at the same speed. You can't see how long a job will take until it's done, the same as an LLM call. You only see how many jobs are queued in each lane and how long the current one has been running.
Random picks any lane. Round-robin takes turns. Best of two is offered two random lanes and takes the one with fewer jobs. Shortest checks all six and picks the shortest. In Two doors mode you're offered the same two lanes the best-of-two bot gets.
The surprise is how much the second look is worth. Random and round-robin spread the jobs evenly, but they ignore which lanes are stuck behind a long job. Comparing just two lanes avoids most of those. Checking every lane is better again, but only while the board is up to date.
Set the board to 12 s stale (about two average jobs) and every dispatcher reads the same old snapshot. They all send work to whichever lane looked empty, so that lane fills up and the shortest-lane rule falls apart. Best of two holds up better, because each dispatcher is offered a different pair of lanes and they don't all pile into the same one.
One more twist: once the board is stale, plain round-robin wins with steady jobs and ties best of two with long-tail jobs. It reads no board at all, so out-of-date numbers can't mislead it. Stale information can be worse than none.
One round of 40 jobs is noisy, so here are averages over 600 rounds of this exact game (six lanes, average wait in seconds, lower is better):
| Turns | Best of 2 | Shortest |
|---|
| Steady, live | 3.9 | 3.5 | 1.8 |
| Long-tail, live | 6.0 | 3.8 | 1.6 |
| Steady, 12 s stale | 3.9 | 5.7 | 8.0 |
| Long-tail, 12 s stale | 6.0 | 6.0 | 6.7 |
And a bigger run: twenty lanes, 60,000 jobs, five seeds, long-tail jobs at 90% load, mean time in the system in units of one average job:
| Random | Turns | Best of 2 | Shortest |
|---|
| Long-tail 90% | 42.8 | 35.2 | 6.3 | 2.3 |
| …board 2 jobs stale | 42.8 | 35.2 | 7.1 | 9.1 |
| …board 8 jobs stale | 42.8 | 35.2 | 8.9 | 26.9 |
With twenty lanes, round-robin stays far behind even when best of two is working from a stale board (35.2 against 8.9), so the round-robin twist depends on scale and job mix. Best of two stays usable throughout. The same idea shows up in hash tables (put each item in the emptier of two buckets) and in load balancers such as Envoy's least-request mode.
Measured: both tables (seeded simulations, this page's own engine for the first) and the per-round bot scores, which are computed exactly from the jobs you were dealt. Not measured: real LLM traffic. Jobs here are random, lengths are capped at 15 average jobs, and every lane runs at the same speed.