# The RL trainer, ONE INSTANCE PER BOT — ingest that bot's own episode log, run
# its own gym self-play, rewrite its own checkpoints. Every bot therefore has an
# independent brain: the six diverge from a common ancestor and can be compared
# on the hub's `brains` line and in the leaderboards.
#
# Gym self-play is the dominant data source (séance 2000, neon-grid 400,
# valley-blocks 200 episodes per run) and each instance runs its own gym at full
# rate, so N brains do not learn N times slower — only the live-match ingest is
# partitioned.
#
#   install: cp deploy/ardegazu-trainer@.{service,timer} /etc/systemd/system/
#            systemctl enable --now ardegazu-trainer@duhul.timer
[Unit]
Description=ardegazu RL trainer for %i (own gym + own episodes + own checkpoints)

[Service]
Type=oneshot
User=ardegazu-bot
Group=ardegazu-bot
ExecStart=/usr/bin/node /opt/ardegazu-bot/stack-a/dist/rl/train.js --config /etc/ardegazu-bot/trainer-%i.json
StateDirectory=ardegazu-bot
Environment=NODE_ENV=production
Nice=10
MemoryMax=1G
NoNewPrivileges=yes
ProtectSystem=strict
ProtectHome=yes