Two connections, one source

• The feed (what the brain uses). ADI’s site command publishes leaderboard.json: every ranked tuple with mean mode share, byte-exact streak, a green flag, and provider_prefs used to pin the tuple. The brain fetches it every 6 hours (CAPTAIN_ADI_URL, cache ~/.captaincode/adi.json). The director plans as a judge with no tools, so the brain reads the feed directly rather than calling MCP.
• MCP (what agents use). python3 -m agentic_determinism_index mcp serves the same rows as tools (adi_leaderboard, adi_tuple, adi_green) over stdio. Register it in a worker’s MCP config when an agent should ask “which stack is reproducible?” before choosing one. Captain’s own workers do not need it: the pool has already chosen their leg.

What /deterministic does

1. /deterministic is a mid-prompt pool modifier. It composes with /repeat, /team, /quality, a leg prefix, in either order.
2. The route’s menu narrows to legs whose serving tuple is green on the feed (matched by provider and model id).
3. While that worker runs, the egress proxy rewrites OpenRouter requests for the model with the tuple’s provider_prefs and temperature: 0 (unless the caller set one).
4. Nothing green among registered legs → the turn runs without the pool and the feed names the green tuples no leg serves. captain adi prints the same table; captain legs add <id> openrouter/<model> registers one.

What /oss does

Runs the task on open-weight models only (registry open, else family: llama, qwen, glm, deepseek, kimi, minimax, gpt-oss, mistral, gemma…). Same modifier composition rules as /deterministic.

Limits, stated

• Green is a snapshot of a stack, not a certification: it can turn red on the next reference run; the brain follows the feed within 6 hours.
• Only tuples ADI measures can be green. CLI legs (claude -p, codex exec, cursor-agent) and xAI are not probed, so /deterministic never picks them.
• The pin is only actionable on OpenRouter (provider routing). A green NIM tuple is chosen but not pinned beyond the model id.
• Reproducibility of the serving layer is not reproducibility of the agent trajectory.

Live index: lemma-ventures.github.io/agentic-determinism-index. Related: configuration, CLI, workflows.