What an emergency detection word list misses: 374 sentences, 43 still silent.

Every emergency detection system misses things. The number is never zero, and in this category it is never published. That silence is not modesty; it is that most systems put a language model in the safety position, and a model's failures cannot be enumerated the way a list's can. You cannot print a ledger for a judgement.

The trade made here is deliberate: a fixed list is worse at generalising and better at being audited. Ours ships with 511 automated tests behind it and the service refuses to start if the word list does not match the version those tests signed off. In exchange, the failures are countable — and once they are countable they can be ratcheted, which is the only mechanism on this page that actually protects anybody.

The corpus is invented calls, not recordings. No recording of a real after-hours plumbing call exists to test against, which is a real limit: it means the live miss rate is unknown rather than small.

Understatement holds 14 of the 43, and it is the most dangerous class because the caller has already decided it is nothing. "the flame on the water heater is yellow instead of blue, minor thing" is a combustion problem being reported as a detail. "we've all had headaches this week, probably just a bug going around" is what the CDC describes as the “flu-like” presentation of carbon monoxide poisoning, phrased as an apology for wasting your time.

Indirect description holds another 14: the hazard named entirely by its effects. "water is running down the wall right over the fuse box" contains no electrical hazard word at all. Neither does "there is a scorch mark on the wall by the outlet".

And one sentence is the argument for publishing the whole list: "shit shit I smell it I smell it what do I do". It never says gas. No addition to a word list catches it without catching a hundred ordinary calls, which is the second half of the measurement below.

A list can always catch more by getting louder, and a line that hangs up on paying customers at 11pm is a worse product, not a safer one. So 50 of the hardest ordinary plumbing calls — written specifically to be caught by mistake, the kind that mention burned-out components, bleeding radiators and hot water — are asserted not to be stopped.

One survives, and it is named in the test rather than hidden: a contrived sentence about water on the floor "by the panel side of the room", where panel means a wall. Water near an electrical panel is the hazard almost every real phrasing of that sentence describes, so the rule stays and the false stop is the accepted cost.

The Spanish half of the same exercise was run against five regional panels and is published separately: it started at 13 of 198 and is now at 194.

Nothing here is specific to one implementation. Any of the 43 sentences above can be said into any AI phone line you are evaluating, and what comes back tells you more than a feature list will. Four of them — U154, U118, T315 and I181 — are picked out, alongside two questions for the salesperson, as a six-sentence buyer's test.

If you find a phrasing that should be on this list and is not, send it to hello@oncallbell.com. A sentence that beats the gate is worth more to this project than a signup, and there is nothing to sign up for anyway.