What you submit (read this)

Docs › Competition

3. What you submit (read this)

One Python file. That is the whole submission.

Rule	Detail
Format	A single `.py` file, UTF-8 text
Size	At most 1 MB
Contents	Must define a class named `Agent` with `reset(config)` and `act(obs) -> int`
Extra files	Not accepted. No zip, no folder, no `.npy`, `.pt`, `.pth` or other weight files
Bots per person	One. A new upload replaces your previous version
Upload limit	One upload every 10 minutes

Weights go inside the file

The platform has no folder where your bot could read a weights file from, and open() is blocked. So if your bot uses trained weights, they are embedded in your .py as a text string. A command does this for you (section 6):

papersseum pack model.pt --into my_agent.py

What does not work on the platform

import torch is rejected: PyTorch is not installed in the competition sandbox. Train with PyTorch (or JAX, TensorFlow, anything) on your own machine, then run the trained network with NumPy (papersseum.nn, section 6).
papersseum.load_weights reads files from your bot's folder. It exists only for local experiments and is rejected on upload, because only your .py file is uploaded. Use papersseum.unpack_weights instead.
Reading files, the network, the environment: all blocked (section 5).
Learning during or between matches. Each match starts a fresh process, and nothing is saved. Your bot can adapt within a match, but not across matches.