Escalation is the core loop, not an error path. A twin that guesses when it doesn’t know is worse than useless — it’s confidently wrong in your name. So when confidence is low, the question comes to you.

​
How it works

  1. Someone asks the twin something and the answer’s confidence lands below 0.6.
  2. The twin consults its lowConfidenceBehavior policy (below) and, by default, escalates — telling the asker it’s checking with you rather than leaving them hanging.
  3. You get the question, the draft answer, and the sources it pulled from.
  4. You approve, edit, answer privately, or say you don’t know.
  5. Your answer goes back to the asker and is learned into the knowledge base.
Next time someone asks something similar, the twin answers directly.

​
The policy

Each twin has a lowConfidenceBehavior setting: Change it in plain language from iMessage or a Slack DM:
“ask me first when you’re unsure”
The burden is deliberately on the bot, not the asker. A teammate asking a question doesn’t know who owns the twin — the bot does, so the bot routes it.

​
Where escalations reach you

  • Slack — a DM with the draft, sources, and the option menu
  • iMessage — escalations send even when other proactive texts are gated
  • macOS app — the escalations view
  • WhatsApp — questions from team-twin members escalate the same way

​
Answering

Reply with your answer. Prefixes control how widely it’s saved: You can also work the queue conversationally: “show pending escalations”, then “approve the trademark one”.

​
Finding the gaps

Shows what escalated, how often, and where the documentation gaps are — plus the pending queue. Read it as a to-do list. A question that escalates repeatedly is a gap you can close permanently by answering it once.

trace & analytics

Why an answer came out the way it did.

Review queues

The other queue — memories waiting on your sign-off.