track()

Replies, opens (where consented), bounces, and downstream meeting bookings. Wired back into the eval set.

Inputs and outputs

In. Inbound mail through your IMAP. Calendar events through your calendar integration. GA4 events from per-lead PDF views.

Out. Per-lead funnel: sent → opened → replied → meeting booked. Per-angle reply-rate updates. Eval-set entries for any negative outcome.

Current version

v2.0.0 Forkable through the prompt library.

Model defaults

Reply classification is a small hosted classifier; we use it to label replies as positive, negative, neutral, out-of-office, or wrong-person. Anything ambiguous goes to the human. You can override per stage in your routing config.

How it fails

Threading. Reply detection across forwarded chains and aliases is hard. We get the easy 92% right; the rest get flagged for manual labelling.

Evals

Reply classifier agreement with human labellers on 3,000 replies: 94.1%. The full eval set ships with the prompt and is editable. Eval format.