Classifier: Fail Closed

How Clara should decide what a message is about, and when to stay quiet. For Gera and Fede, after the 10-27 call.

2026-10-27 · read-only audit of prod and code, plus a literature review · nothing changed yet

The collections bug is not the classifier. The check that decides whether Clara replies files "rent balance" under other, and other always sends. A bigger model would not stop it. The fix is to teach that check about money and make "unsure" mean "stay quiet".

Today vs. target

Today

Tenant: "I can't pay rent this month"
Reply check: 3 choices — renewal, maintenance, other
"Rent balance" is listed as other, which no module owns → sends
Timeout or error → sends
Topic label lands ~10s later. Nothing reads it to decide a reply.

Target

Tenant: "I can't pay rent this month"
Tripwire first: money trouble, hardship, dispute, legal, "talk to a person" → silent + human
Classifier: can pick more than one topic, each scored by 3–5 votes
Any possible topic whose module is off, or not sure → hold
Timeout or error → hold, never send

Next steps

Stop the leak: a money choice in the reply check, block when a module is off, timeouts hold.
Tripwire before every reply.
An answer key of 300+ checked messages.
Multiple topics with vote-based confidence.
Ask counsel about debt-collection rules; rent replies stay drafts until then.

Details

How good the labels are today

248 real tenant messages, judged by hand against the conversation:

Measure	Result
Overall, weighted by volume	87.7% strict · 94% counting defensible calls
Leasing · maintenance · billing	100% · 90% · 87%
General resident questions	75%
Collections-like messages labelled as something else	~1 in 7

Small samples, ±10–15 points per topic. Collections isn't its own topic today (it sits inside billing). The test suite has 102 cases and none for late rent or eviction.

Is Opus 5.5 worth it?

Not as the main fix. Studies find a bigger model alone moves classification accuracy only a little, and the speed cost is real. Votes across several calls tell us more about confidence than model size does. Use the big model where it's cheap and it matters: the one reply-time check.

Setup	Per month now	If today's spike lasted	Speed
Today (Haiku 4.5)	~$9	~$78	~13s
Opus 5.5 for every label	~$38–62	~$330–530	~20–60s
Cheap first, Opus only when unsure (~35%)	~$23–29	~$195–245	mixed
Opus on the reply-time check only	~$0.004 per reply		needs a longer timeout

List prices, estimated from message sizes rather than measured. Opus speeds are estimates. "Classifying…" must clear in ~20s.

Keeping it right, and fixing it when it's wrong

An answer key we trust: 300+ hand-checked messages, weighted toward rent and hardship. If 0 of 300 are missed, the true miss rate is under ~1%.
Each module has a limit for wrong replies. The confidence bar is set from the answer key so the error rate stays under it, and it is re-checked after every prompt or model change.
Unsure cases go to review. Reviewers see the message before the label, so they aren't steered.
Re-labels keep history: we keep the label Clara saw when she decided, plus the corrected one. A correction that would have changed a past reply gets flagged.
Release safely: test against the answer key in CI, run silently beside the live version, then turn it on for one customer first.
Later: train our own small classifier from the labels and corrections. Big support teams do this to get cheaper, faster and more predictable results.

Sources and code checked

Stated confidence vs. sampling · Guaranteed error rates (conformal) · Models rarely abstain · Small fine-tuned vs. frontier models · Multi-intent messages · Intercom's escalation classifier · CFPB on chatbots · Claude API: no token probabilities

Code checked: the reply check's choices and its "rent balance → other" line are in classify-reply-topic.ts; supervised-reply-rule.ts maps "other" to no module, so it sends.