Operations Copilot: Shift Handoff and Campaign Context

Summarizing machine state, open items, and outstanding risks between crews with citations, so context is not lost across a handoff.

The handoff problem

Continuity across shifts is a real operational risk: a crew inherits a machine whose recent history it did not witness. The Operations Copilot generates a structured handoff briefing — current twin state, what changed during the shift, open anomaly events, pending maintenance, and any degraded operating limits — each item cited to the underlying telemetry or record.

What the briefing contains

• Current operating point and distance to the envelope boundary
• Events since last handoff: shots run, anomalies triaged, actions taken
• Open items: unresolved anomalies, deferred maintenance, degraded limits
• Active advisories from the Plasma and Engineering copilots with confidence

The briefing is assembled by querying the L2 fabric for the shift window, pulling the triaged anomaly events from the Engineering Copilot, and summarizing with citations. Nothing is asserted without a source; a claim like "coil 3 temperature margin reduced" links to the exact telemetry and the subsystem analysis that concluded it.

handoff(shift_window):
  events = L2.query(all, shift_window)
  anomalies = engineering_copilot.triaged(shift_window)
  changes = diff(state_start, state_end)
  brief = summarize(events, anomalies, changes)
  assert every claim in brief has a citation
  return brief   # for human review, not auto-filed

Honesty in summarization

Summarization is where a copilot is most tempted to smooth over uncertainty. The handoff generator is evaluated specifically for faithfulness: it must not drop an open risk to make the summary tidy, and it must carry uncertainty forward. Items it could not verify are labeled as unverified rather than omitted. The evaluation harness includes handoff-faithfulness tests — see grounding evaluation.

The briefing is a decision aid for the incoming shift lead, who confirms and owns it. Because both machines are pre-operational, handoff generation is being validated today against simulated campaign runs on the twin so the format and faithfulness are proven before FOAK. Handoffs are retained for audit and provenance.