Reply router
Nine categories, three buckets, and a confidence gate that overrides every one of them.
This is the decision layer of a reply classifier that runs once a day against a cold email campaign. A reply comes back, Claude assigns it one of nine categories with a confidence score, and then the code decides what actually happens to the person who wrote it.
Pick a reply on the left. Then drag the confidence slider below 0.70 and watch a positive reply stop being positive. That gate is applied last and unconditionally, after the category mapping, because the expensive mistake here is registering someone for an event they never agreed to attend.
Floor is 0.7. Drag below it.
The routing, step by step
Nine categories, three buckets
Why it is built this way
Unsure is a destination, not a failure. The obvious build forces a binary: interested or not. That buries the ambiguous replies in Negative, and the ambiguous ones are exactly the replies most worth a human read. Four of the nine categories route to Unsure, including referral, because a human handoff deserves a human.
The push decision ignores the model's own answer. The classifier returns a push_to_goldcast boolean. The code discards it and re-derives the decision from the final bucket. One code path, and no way for a model quirk to route around the confidence floor.
An out-of-office that names a colleague is still an out-of-office. The prompt says so explicitly, because the tempting read is "they gave me a name, that's a referral". It is an auto-reply footer, and treating it as a human handoff fills the review queue with robots.
The mapping table, the 0.70 floor and the push rule on this page are a direct port of the production module. The example replies are written for this page. Real prospect replies are not published here.