There is no average resident.

When we started shipping Companion into skilled-nursing wings, the first instinct was the engineer's instinct: find the typical resident, tune for them, ship. So we pulled distributions from our pilot — speech rate, hearing thresholds, response latency, how often a turn got cut off mid-sentence — and went looking for the center.

There wasn't one. The histograms didn't have a single hump with tails. They had two and sometimes three separate peaks. Speech rate clustered around a brisk, fluent group and a slow, halting group with almost nobody in the dip between them. Hearing thresholds split the same way. The mathematical average sat in that empty valley — describing a resident who, in our pilot, essentially did not exist.

Why the average actively hurts

An average isn't just uninformative here; it is harmful. Set the speech rate to the mean and it's too fast for the resident with post-stroke aphasia and faintly patronizing to the sharp, fluent retiree who just uses a hearing aid. Set the turn-detection silence window to the mean and you cut off the slow talkers while making the fast talkers wait. A single config tuned to the center fails both real groups to split the difference for a person who isn't in the room.

The standard fix — let users adjust settings — doesn't survive contact with the bedside. Our residents will not open a menu. Many can't. The device has to arrive already correct.

The cohort as the unit of design

So we stopped designing for a person and started designing for a cohort: a group of residents who share the constraints that actually move our settings. A cohort isn't a diagnosis and it isn't a demographic. It's a bundle of behaviors that demand the same tuning. The ones we run with today:

• Hard-of-hearing — high-frequency loss that eats consonants; needs slower, clearer speech, not just louder.
• Dementia — repetition, confabulation, looping questions; needs a conversation policy, not correction.
• Post-stroke aphasia — word-finding pauses and halting output that break speech-to-text and turn detection.
• Sundowning — the same resident, agitated and disoriented in the late afternoon, who needs different handling at 4pm than at 10am.
• Conversational baseline — fluent residents for whom the stock settings are nearly right.

Each cohort gets its own defaults stored in Firestore and cached in Redis, keyed per device. A nurse can assign a resident to a cohort the way they already chart anything else — it's clinical knowledge they hold on day one. No menu, no resident-facing setup.

The payoff is concrete. The resident in room 214B is not the statistical center of a wing. She is a specific person with a specific hearing curve and a specific way of trailing off mid-thought. Designing for her cohort means the first sentence Companion speaks to her actually lands — and the rest of this series is how we tune each one.