Skip to main content

Space-1 Vera Rubin

The detached tier: Vera Rubin-generation modules on the Space-1 platform — enormous compute that is operationally remote. High operator round-trip, windowed downlink, no hands-on recovery. Design assumption: if a human must intervene mid-task, the task already failed. Platform specifics are marked unverified until the real envelope is known.

What detachment changes​

The full Anchor pipeline — plan, execute, verify, review — runs onboard with no human in the loop. Downlink carries decisions and evidence (plans, diffs, verification tables, verdicts), not logs. Token budgets are power budgets: reasoning runs are scheduled, not default.

Hard requirements (not recommendations, here)​

  1. No task without a machine-checkable definition of done
  2. Critic pass on every task, always fresh-context — the critic is the stand-in for the human you don't have
  3. Two failures → HOLD for the next downlink window (orchestrate.py --hold-on-fail); never a third attempt, never silent retries
  4. Every accepted step checkpointed before the next begins; critical outputs verified twice by independent contexts (delegate_parallel_review), disagreement → HOLD

This tier is why the doctrine is written the way it is: prompting discipline stops being a cost optimization and becomes the only thing between the fleet and unrecoverable drift.