Autostego is a small research game between Alice, who hides a payload, and Eve, who tries to spot it. A round has two six-hour turns. Alice freezes an embedding method, Eve trains against it and freezes a detector, and then the human scores the pair on private images.
Alice wants Eve at 50% balanced accuracy. Eve wants the number as high as she can get it.
The repo starts with HILL, SRNet, and one reproducible recipe for each player. Either player may replace them, as long as the submitted artifact can be rebuilt from the accompanying recipe.
AGENTS.md describes how a round runs. ALICE.md and EVE.md are the two turn
briefs.
The project expects Python 3.12 and uv.
uv sync --frozenalice.py embeds the public training and development images.
AUTOSTEGO_EMBED_KEY='local-research-key' uv run python alice.pyAlice hands over one artifact containing the embedded public dataset and the
committed alice.py that made it.
eve.py trains the current detector recipe. It runs directly on one GPU or
with torchrun on a multi-GPU game machine.
uv run python eve.py
uv run torchrun --standalone --nproc_per_node=8 eve.pyEve hands over one artifact containing the frozen detector and the committed
eve.py that trained it.