What can ordinary behavior tell us about change?

Rientro's question is about ordinary behavior, so the first move wasn't to build a new world around it. We looked at traces of ordinary movement that already existed.

GeoLife — a large public collection of real-world GPS traces — gave us a way to examine the structure of outdoor movement without waiting for our own data to accumulate. Twelve features, extracted from raw trajectory points:

Trained as an IsolationForest — a model that treats the crowded middle of a distribution as ordinary and the edges as anomalous — the result is a single score in [0, 1]. That score isn't a research exercise sitting off to the side. It runs inside Rientro's live ingestion path today, on every scored trajectory, as a bounded multiplier on how urgent a moment looks.

But location tells us where movement happens. It doesn't necessarily tell us why, or what happened between two points, or what changed inside the person whose movement we observed.

So we looked closer.

A second dataset gave us another way to look — this time indoors. TIHM instrumented the real homes of people living with dementia using passive infrared motion sensors and magnetic door contacts. Nine features, this time about how a day is inhabited rather than where a body travels:

Trained the same way — an IsolationForest, scored against the population's ordinary day — this became a second, separate model.

Rientro doesn't deploy fixed indoor sensors. So the nine features above are approximated instead from wearable telemetry resolved against a floorplan — a PIR analog, not PIR data. Two structural gaps come with that substitution:

1. A motion sensor fires on movement; a wearable reports on a timer. A person sitting still produces zero sensor events and a full stream of telemetry points — the approximation measures sampling, not activity, wherever someone is stationary.
2. The model was trained on full patient-days. A short window doesn't read as calm — it reads as an extreme outlier, for reasons that have nothing to do with behavior, before any actual pattern is considered.

Because of that second gap, we treat a short window's score as unusable — it is never shown to anyone, by rule rather than by accident. This piece is built, tested, and correct — and deliberately dormant, waiting on real indoor data and validation before it touches a live decision.

Two models, two windows onto the same kind of question, at two very different stages of trust. Naming the boundary plainly is more credible than pretending it isn't there.

This is not where the research ends.

It is where the next question begins.

These are not claims. They are questions we are still working on.