Question 1

How fast is a full tournament and what does it cost?

Accepted Answer

The README's measured run of 2,000 animals decided 1,999 fights in about 16 seconds with 64 API calls, about 7.6 ms per decision and roughly US$0.01 total. A single hand-picked duel returns in about 220 ms.

Question 2

Does the tournament handle position bias in Jev's answers?

Accepted Answer

Yes. The README reports a mild bias where the first-listed option gains about 0.1 probability, so fight order is shuffled per fight and single-fight mode asks both option orders and averages them.

Question 3

What is speculative fan-out, and how does it affect the call count?

Accepted Answer

One request asks the current champion against each of the next K challengers at once. Code walks the answers in order and, the moment the champion falls, throws the rest away because they were about the old champion, which is why a 1,999-fight tournament costs only 64 API calls. K grows with the champion's streak, since a fresh champion tends to fall soon.

Question 4

How is the key handled, and why is there a local server?

Accepted Answer

A tiny server bound to 127.0.0.1 holds the key and talks to Jev, because the TypeSafe API does not accept cross-origin browser calls and a key must not live in a web page. Beyond the calls to TypeSafe itself, the key never leaves your machine. The UI is at http://127.0.0.1:8777.

Question 5

What did the authors learn about how Jev reads a question?

Accepted Answer

It reads literally. Avión zapador is a bird, the sand martin, but avión also means aeroplane, and Jev scored it 0.56 against a rhino; presenting every animal with its class, as Avión zapador (ave; animal volador), raised that to 0.75. Arena rules go in state and genuinely change the answer: lion beats great white shark 1.00 on land and loses 1.00 in water.

Question 6

Do the probabilities behave like probabilities in practice?

Accepted Answer

They did in the authors' check. Replaying one 300-animal draw with the with-chance rule produced 23 upsets where the probabilities predicted 20.7, and with favourite-wins there were none. Reusing a warmed keep-alive connection also mattered: a fresh TLS handshake per call, about 0.3 s, had been hiding Jev's speed, taking a 6-decision request from roughly 500 ms to 200 ms.

Question 7

Can I run the tests without a key, and what license applies?

Accepted Answer

Yes. python3 -m unittest discover -s tests -v replaces Jev with a deterministic fake referee and runs offline, failing if the winner-stays-on chain breaks, if speculative answers stop being discarded, if per-type quotas or seeding change, or if the single fight stops cancelling position bias. The project is MIT licensed, needs Python 3.10+ with no dependencies, and its UI, animal list and prompts are in Spanish.

Jev Torneo de Animales

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community