gyrell

gyre's RL environment for ludic learning.

gyrell turns turn-based games into verifiable RL environments — and, if you like, into a live broadcast. Describe a game once as a turngames spec; the same spec drives scripted baselines, LLM-seated matches, verifiers environments, and an around-the-clock spectator show. The reward never needs a judge: you either won the game or you didn't. Commerce is not our goal here at gyrell; verifiable play is.

Two games ship as worked examples — codewords (Codenames-style hidden information) and taboo. The docs are a guide to adding yours.

See it in action

You can watch language models play codewords and taboo live on the LUDICA broadcast channel.

• src/turngames — the kernel: state machine, event log, per-seat visibility, word clocks, scripted and LLM actors.
• server/ — the broadcast channel: rotating lineups, an LLM host, archived replays, an Elo leaderboard. Pairs with the LUDICA spectator client (separate repository).
• environments/gyrell — the verifiers environment, published as lumpenspace/gyrell on the Environments Hub, with prime-rl configs in configs/.

Quickstart

uv venv && uv pip install -e ".[server]"
python scripts/play_codewords_demo.py     # scripted match in the terminal
uvicorn server.main:app --port 8000       # broadcast (replays need no keys;
                                          # live models need OPENROUTER_API_KEY)
pytest tests/                             # the house Voight-Kampff

Name

gyrell: gyre + RL + ludic learning. Any resemblance to corporations that manufactured beings more human than human is, of course, aspirational.