Classifier Plan

What changes in how Clara decides whether to reply, before and after, with numbers from 639 real moments. For Gera and Fede.

2026-10-27 · v2 · after Fable, Astra, the voice memos and the turn study · nothing built yet except the rent fix and Resident Services

Targets are goals to measure against, not results yet.

What the study found

• Thread state decides most cases. Teammate active in the last few messages → hold or silent 84% of the time. A person took the call → silent 44 of 45.
• Kinds of message (9): a real ask 47%, an answer to our question 16%, thanks or bye 12%, an FYI 7%, ok 6%, unclear 5%, automated 3%, greeting 1%.
• Topics (10): leasing, maintenance, general contact, Resident Services, not known yet, collections, vendor, billing, move-out, renewal. 9% of messages carry two topics.
• Risks (9): money trouble, privacy, legal or dispute, safety, fair housing, hardship, waiting on a staff decision, distress, confusion. Plus "wants a person".
• "Not enough context yet" happens on 12% of messages, mostly the first turn of a call. It clears by the next message 65% of the time.
• Soft replies ("hi", "thanks"): only 20% were clean. The rest came with staff mid-thread, a risk nearby, or a person on the call.
• Reviewer agreement: who's writing 0.85 and topic 0.82 are reliable. Risk is weak at 0.39, so it gets a keyword floor that can't be argued with.

Where it sits in Fede's machine

• Gather context adds who's writing and the thread state, both from records (no model), plus the topic and the keyword risk floor.
• Should Clara speak stays thin: unknown sender, someone already on it, or a module off → stop before any draft.
• Write (Opus) returns the draft plus its own read of kind, topic, risk and value. That matches Fede's rule that judgment lives inside the writing call.
• The check on the draft (Opus, one configurable layer, every check must pass) → send. Otherwise hold, ask the team privately, or stay silent.
• Record keeps every decision and which check stopped it. That's where the counts come from.

What Fable and Astra changed

• A correct label isn't permission. The check reads the draft, not just the topic.
• Asking the same model twice doesn't catch its repeated mistakes. Instead: a keyword floor, different reviewers, and a check on the draft.
• A whitelist: a topic that maps to no module holds.
• The answer key has to be bigger and checked by people: about 1,000 messages, 300+ of them about money.

Order of work

1. In flight: stop the rent leak (#22879), and make Resident Services a real module with its own switch.
2. You and Fede: vote on the 40 disagreement cases on the scenario board. Those votes become the written rules.
3. Build the answer key from the study's 639 moments, plus more money cases, settled by people.
4. The reply check learns who's writing, thread state and the risk floor. Measured on the key before it goes live.
5. The Opus check on the draft, running silently next to today's logic for a week, then one customer at a time.
6. A count of blocks per check, visible to you.

Decisions still open

• Where "what's my balance?" (Resident Services) ends and collections begins.
• Whether an Off module tells the tenant "we don't handle that here", or stays quiet.
• The Friday-night case: is one "the office will reply Monday" ever allowed?
• Fede: OK for the check on the draft to sit between his writer and the send.

Try it: Clara Reply Scenarios (flip modules, vote on 40 real cases) · Research: Classifier: Fail Closed · Fede's engine: Clara, the new hire