Small talk is the hardest thing we build, because there's no slot to fill.

Most voice AI is task-oriented, and tasks are merciful to build. There is a goal — set a timer, book a flight, find a song — and a clear moment of done. You can measure success: did the task complete? You can even fill the missing pieces with a checklist: ask for the date, ask for the destination, confirm. Companionship has none of that. When a resident in a quiet room at 9pm just wants to talk, there is no slot to fill and no finish line to reach. That absence is the whole engineering problem.

No success signal

Ask an engineer to optimize a system and the first question is against what metric. For a task bot it's completion rate or time-to-task. For companionship the honest answer is uncomfortable: the metric is whether the resident felt a little less alone, and we cannot read that off a log. The proxies are all wrong. Conversation length isn't it — a confused, looping conversation is also long. Turn count isn't it. We can detect the catastrophic failures — the resident disengaging, going quiet, sounding distressed — but the difference between a good chat and a great one doesn't show up in any field we store.

Nothing to complete

A task bot that runs out of slots to fill is finished, and silence is correct. A companion that runs out of things to say has failed at the one job it had. So the model has to carry a conversation that has no inherent direction — follow a thread, remember that she mentioned her garden last week, ask the question that opens the next door — without ever feeling like an interrogation. The failure modes here aren't crashes. They're subtler and more human:

• The interviewer — relentless questions with no offering of its own. Exhausting, not warm.
• The cheerleader — empty enthusiasm that flattens real feeling. That's wonderful! to someone who is sad.
• The amnesiac — no continuity from yesterday, so every conversation starts cold and nothing accumulates.
• The closer — wrapping up and wrapping up, treating the chat as a task to finish instead of company to keep.

Fixing these isn't a feature you ship; it's a posture you tune, and you tune it against the messiness of real human talk rather than a test suite. Continuity comes from the memory layer. Warmth comes from pacing and prosody. Restraint — knowing when not to ask another question — comes from a system prompt written more like guidance for a good listener than a spec for a bot.

And it stays bounded by the same rules as everything else we do. Companionship is not therapy and not clinical advice; the moment a chat drifts toward meds or symptoms, it leaves the small-talk lane and becomes an escalation to a person. Within those edges, the goal is plain. The resident in 214B has had a long quiet day, and for a few minutes someone asks about her garden and actually remembers the tomatoes. There is no task there. That is the product.