Copilot Conversation Surface
The window through which L5 copilots explain, recommend, and answer — always advisory, always with citations and confidence, never issuing commands on their own.
Where the copilot meets the operator
The L5 copilots reason over the machine, the documentation, and the shot database; L6 is where their output reaches a human. The conversation surface lets an operator ask 'why is the disruption-risk index rising?' or 'what changed since the last good shot?' and get a grounded answer. Crucially, the copilot advises and explains — it never actuates. Any action it recommends flows through the same trust-and-override path as any other AI recommendation.
Grounded, cited, and honest about uncertainty
Every copilot answer is grounded in retrievable sources — specific diagnostics, twin estimates, documentation pages, prior shots — and shows its citations so the operator can verify rather than trust blindly. Answers carry the copilot's confidence, and the surface distinguishes what is measured from what is inferred. When the copilot is uncertain or the question falls outside what it can ground, it says so plainly instead of fabricating, because a confident wrong answer in a control room is worse than an honest 'I don't know.'
- Explain: why an alert fired, what precursors drive it, what changed
- Recommend: a response, with rationale, confidence, and citations — advisory only
- Retrieve: relevant procedures, prior similar shots, documentation
- Never: issue a command directly; recommendations go through human approval
Integrated, not a separate app
The conversation surface is woven into the operator's context. Asking about the current alert automatically scopes the copilot to that alert's data; a recommendation links to the exact provenance and envelope margins involved. During an upset the copilot's role is to compress the situation into a clear rationale for the decision surface, not to add another thing to read — its output is filtered through the same attention budget as any notification.
The copilot surface depends on the honesty machinery around it — see provenance display and trust and override — and its recommendations are captured, like all actions, for audit.