arlolisted

# arlo, a real, local operator arlo hands you the real command to run when the frontier model is down or rate-limited. You type your own words ("bounce the deriver", "reset a password"); arlo answers from your system's own ground truth, on an independent local path. **The invariant, narrow and sacred: arlo never invents a command that is not real ground truth.** At worst it surfaces the *wrong real* command — always with its confidence and the alternatives, so you can tell — never one that does not exist. (The failure it prevents: a flag table said `--parent` for months after the real flag became `--parent-id`; everyone who trusted it made orphans. arlo answers from the live surface.) ## The trust gradient Every answer is **labeled with the trust it earned**. Climb only as far as the answer needs: - **rung 0, retrieve**: return a card's command verbatim. Full grounding. - **rung 1, slot-fill**: bind a real template's hole from your words (`bounce the deriver` → `docker compose restart deriver`); skeleton verbatim, only the slot inferred and shown. - **rung 2, reason-rank**: when scores can't separate real cards, pick *which real card* and say why; return an id, never a command. Full grounding. - **rung 3, explain**: narrate the chosen command from its *own* `--help`, never memory; every cited flag must be a real token in that help. Grounded. - **rung 4, compose**: chain real cards into an ordered runbook; atoms real, order inferred → **less-trusted**. Return ids;