Limitations

Written plainly, because deciding whether this fits is easier than undoing it later.

Answers carry no reasoning

You get a number. You do not get why. There is nothing to audit after the fact beyond the state that went in, which the integration records for you.

If you need to explain a decision to someone later, this is the wrong tool.

Confidence has no published calibration

TypeSafe publishes no calibration evidence for confidence, and their own docs call it a convenient default.

Treat 0.9 as higher than 0.6. Do not treat it as right nine times in ten until you have measured it on your own questions.

It judges, it does not calculate

Hand it a reading and a threshold and it will not reliably work out which side the reading falls on. Measured, that gave 0.06 separation against 0.69 when the comparison was done first.

Do arithmetic in Jinja, where it is exact and free, and hand over the conclusion in words.

Slower than the published figure, from Europe

TypeSafe publishes 70 to 500 ms. Measured across 16 calls from the Netherlands, a warm connection answers in 250 to 580 ms and the first call after an idle spell takes 700 to 900 ms. Their figures were measured near their own service.

Fine for a doorbell. Too slow for a tight loop.

Not for safety decisions

A probability with no explanation should not hold a lock, a heater or a smoke alarm.

Use it to decide whether to tell you something. Do not use it to decide whether to do something that is hard to undo.

Structured criteria bought nothing measurable

The API accepts objects and arrays for instructions and criteria values, and the docs say structure sharpens the boundary when two options blur.

On five deliberately ambiguous doorbell callers, three runs each: 12 of 15 agreed either way, mean confidence 0.90 flat against 0.87 structured, neither unstable. An easier set gave 12 of 12 for both.

It is supported. Reach for it only when two options genuinely blur and a plain sentence has already failed.

The conversation agent handles five intents

On, off, toggle, brightness and state questions. Media players, covers and climate setpoints go to the fallback agent.

It is not a core integration

quality_scale.yaml tracks this against Home Assistant's quality scale at 47 done and 7 exempt, including all three Platinum rules. That file is a self-assessment: hassfest does not validate it for a custom integration, which I checked by deleting a rule and watching it pass.