The config you tuned in week one is wrong by week six.
Residents don't hold still. They decline, they recover, they come back from a hospital stay as someone new. A cohort assignment is a snapshot of a moving target — so we monitor for drift instead of trusting the original label.
We spent this series arguing that cohorts, not averages, are the right unit of design. There's a catch we have to say out loud: a cohort assignment is a measurement taken at a moment, and residents are not stationary. The resident in room 214B who was a clean conversational baseline in week one has, by week six, slowed her speech, started repeating questions, and stopped hearing the consonants. The label is still 'baseline.' The person isn't.
We call this cohort drift, and it's the failure mode that quietly undoes everything good about cohort tuning. A config that was correct on admission becomes subtly, then badly, wrong — and nobody filed a ticket, because the resident can't, and the device looked like it was working.
Drift runs both ways
Decline is the obvious direction: cognition fades, hearing worsens, a stroke or an infection moves someone overnight into a cohort they weren't in yesterday. But recovery is just as real and just as mishandled. A resident comes back from a hospital stay tuned for aphasia, does the rehab, regains fluent speech — and is now stuck behind a 3.5-second turn window that makes a recovered person feel disabled. A stale cohort can hold someone down.
Monitoring instead of trusting
Because we store events rather than recordings, every conversation already emits the signals that would reveal drift. We watch them per resident, as slow-moving trends, not per-turn alarms:
- Speech rate and mean pause length — the early tell for both aphasia onset and recovery.
- Repetition rate — the same question returning more often is a cognitive signal.
- Repeat-request rate ('what?', 'say again') — a proxy for worsening hearing.
- Turn-timeout rate — how often our silence window cuts the resident off, which means the window no longer fits them.
- Sundowning onset time creeping earlier evening over evening.
When a metric drifts past a band we hold a resident-specific baseline for, the system does not silently re-tune itself. It raises a suggested reassignment for a human — the same nurse review that signs our SOAP notes. A drifting trend in our data is often the first sign of a real clinical change, so the right move is to surface it to a clinician, not to quietly paper over it with a config swap.
An auto-correcting device that hides a resident's decline is worse than a slightly mistuned one that reports it. Drift detection is a clinical signal first and a tuning input second.
The product result closes the loop on the whole series. Designing for cohorts only works if the cohort stays current — so the system that fits the device to the resident is also the system watching for the day the resident changes. When 214B starts repeating herself in week six, a nurse sees it as a trend, not a surprise, and the next conversation she has with Companion already fits the person she's becoming.