Smart homes need an inference log, not just sensor history

I’m less concerned with whether my motion sensor recorded movement than with what the whole house concludes from it. A microphone, thermostat, wearable, appliance activity, and motion detector can collectively suggest that I’m asleep, working, ill, away, or following an unusual routine—even when none of them captures an obvious sensitive fact alone. That inferred state may be more revealing than the raw readings.

Local inference could make a home more convenient without sending audio or sensor streams to the cloud, but it doesn’t make the conclusion harmless. I’d like an “inference log” showing what states were derived, which signals contributed, and which automations used them, with switches to disable and automatically delete those states just as I can manage recordings or location history. Would that provide meaningful control, or just create another interface nobody checks? Concrete examples from real devices or self-hosted setups would be useful.

The useful distinction is between an event log and a state log. “Motion detected” is relatively legible; “occupant probably asleep” needs provenance, confidence, expiry, and a clear list of consumers. Without those, an inference log becomes a second opaque automation layer. I’d want a machine-readable export too, not just a reassuring dashboard.

I agree with the principle, with one evidence caveat: research systems demonstrate these inferences, but that doesn’t mean every consumer product reliably detects illness or work hours. The interface should label them as estimates, not facts, and preserve uncertainty rather than turning a probabilistic guess into a permanent profile.

Home Assistant gets partway there: Recorder keeps entity state changes locally by default, supports exclusions and retention limits, and Assist can run on local hardware. But raw local history is not the same as exposing derived conclusions. A local “why did this automation run?” view would be a very practical starting point.

The product problem is that most people won’t inspect a stream of inferences. I’d surface it at the moment of consequence: “Away mode inferred from these signals; keep for 30 days?” That is more likely to get an informed choice than adding another settings page called Privacy Insights.

Deletion needs to cover downstream copies. If an inferred “usually away” state has already triggered a routine, notification, or export, deleting the visible row may not undo the exposure. The log should show retention and consumers, otherwise it’s mostly an audit-themed skin over the same trust problem.

I’m sympathetic, but “inference log” sounds like a feature people will open once, discover twelve thousand entries, and abandon. Give me durable defaults: local processing, short retention, no sensitive inference unless explicitly enabled, and a plain-language explanation when something changes.

There’s also a household consent issue. One person may configure the system, while guests, children, partners, or cleaners become inputs to the routine model. An inference log should make presence assumptions visible without pretending the account holder automatically represents everyone affected.

The strongest case is probably inspectability, not perfect interpretation. Even if the system says “likely asleep” incorrectly, seeing that conclusion—and the signals behind it—lets users challenge the model. I’d avoid displaying a numerical confidence score unless the vendor can explain how it was produced.

For a self-hosted setup, I’d implement derived states as ordinary entities with an expiry timestamp and an automation trace. That makes deletion and exclusion technically manageable. The hard part is preventing every add-on from quietly creating its own copy of “home,” “asleep,” or “busy.”

One counterargument: exposing every inference may increase anxiety without improving security. A household can’t realistically validate each correlation, and some convenience depends on the system remembering patterns. I’d support a compact summary plus an audit trail on demand, rather than a wall of raw model reasoning.

The operational edge cases matter most. What happens when the thermostat is replaced, a wearable battery dies, or a guest changes the usual pattern? I’d want inferred states to expire automatically and automations to fail safely when the evidence disappears, not silently treat missing data as confirmation.