Why seenzus had to learnwhen not to speak 管家为什么必须先学会不说话

seenzus’s home screen used to open with a flattering headline: “Handled N things for you today.” The number came out of the operation log, updated daily, and sat in the most expensive spot in the product. We deleted it. Not out of modesty: when we checked the number line by line, it failed in four places. It counted issued commands as “handled,” although the operation log does not know whether a command actually succeeded. It excluded scheduled tasks, even though the section directly below shows exactly those results. Its “today” rolled over at the device’s midnight rather than the home’s own clock. And when the number could not load, the page fell back to “A quiet day so far,” describing a network failure and a genuinely calm home in the same words.

One sentence meant to prove diligence, wrong four ways. That flipped the direction of our worry. Going in, we worried that a proactive Smart Space Agent would miss things. Assembling the home screen taught us the crowded side is the speaking side. Weather wants to mention an umbrella. Learning wants to report a new habit. Device state wants to announce that something went offline. Scheduled tasks want to show last night’s results. Each source, taken alone, has a good reason to speak; put together, they make a notification feed. A notification feed and a wall of device toggles do the same thing to your attention: they hand the “should I care about this” judgment back to you, in bulk.

So across these three months we have spent less effort making seenzus speak than stopping it from speaking. Speaking is a managed act now: one voice at a time; evidence before words; a “no” from you gets corrected first; and when it is not sure, it says so in words instead of dressing the guess up as a percentage. The rest of this post walks through those constraints, then what they cost.

Home carries at most one voice at a time

No source writes to the home screen directly. Weather, learning results, and device state can only submit “I have something to say” to a single arbiter, which decides whether anything deserves your attention right now and, if so, picks exactly one thing. That thing becomes the sole protagonist and the screen enters its busy state. No protagonist, and home is a verifiable line instead: I’ve had a look, nothing needs you. The two states are exclusive, and at most one action waits for you per screen.

The pecking order is fixed: verified urgent risk first, then time-boxed action proposals, then due reminders, and new discoveries last. Within a layer, candidates compete on relevance to the space you are in and on how fresh their evidence is. Losers do not queue. Once you handle the protagonist, this visit goes quiet and nothing backfills; the one exception is an urgent alert that is still valid. We did not build an inbox that pops the next item when you close one, because an inbox turns the home screen into debt. Discoveries also carry a hard quota: one per person per day.

The most direct consequence of this arbitration is that a lot of perfectly harmless remarks get suppressed forever. We accept that.

Speaking has to clear evidence first

A single event is not a reason to speak. The exception is anything you explicitly asked it to watch, a channel you authorized yourself; we come back to that below. A light on once at three in the morning, no coffee brewed today: these enter the record, but one isolated observation is not worth anyone’s attention until it sits alongside longer evidence.

A reminder that made the schedule is re-checked at the moment it would appear. If the window is already shut, there is no close-the-window reminder. If the rain window has passed, the umbrella is off the table. If the upstream forecast has been withdrawn, so has the topic. One rain gets one mention per day; answer “got it” and rain stays off the screen until tomorrow.

A freshly connected space is the harshest case of this rule: seenzus knows almost nothing about it, so it says almost nothing beyond “I’m keeping an eye on things.” For the first two weeks it looks uneventful. We accept this too: enthusiasm without evidence turns into noise within days.

A single no outranks ten yeses

One basic form of speaking is holding up a concrete observation for you to judge, something like “lately the living-room light is often still on after ten,” and you answer right or wrong. Both answers are evidence. They are treated very differently.

“Wrong” takes effect immediately. That observation closes for good, and the same thing will not come back reworded. The learning subject behind it is slotted into the next learning batch for priority review. Before the topic can come up again, seenzus needs materially new behavioral evidence, and even then it waits for the normal cadence. “Right” jumps no queue; it is recorded and digested by an ordinary later batch. The asymmetry is deliberate: something said wrong has to be fixed promptly, and being confirmed does not justify displacing unsolved problems.

The other rule matters just as much: silence is not a no. A card you walk past, ignore, or let expire is never converted into “he dislikes these reminders.” No response means no response. The product does not invent a position for your silence.

The internal confidence score never becomes a number on screen

seenzus does keep confidence internally: every learned rule carries a belief score that rises and falls with feedback, and learning uses it to rank and prune. That number appears on no user-facing surface. No “80% sure,” no “12 out of 14 nights,” no “observed for 14 days.” When it is not sure, it says “waiting for you to confirm,” or “let me watch a little longer,” or it stays silent.

The reason is that a percentage, here, is a performance. “86% hit rate” reads like honesty, but it hands the judgment right back to you: whether that number is high enough to trust is now your math to do. We want it to own “I’m not sure” in words. The numbers you do see describe the world: sixteen devices, seven of them on, idle for two hours. Those are facts it observed, not grades it gave its own guesses.

This line has a cost of its own. Some users will want to see exactly how sure it is, and will read the absence as opacity. Our current answer: every sentence it says can be checked against the world, and that is closer to transparency than putting a price tag on a guess.

The top priority layer has exactly one source today

The first layer of the arbitration is verified urgent risk. Today that layer has a single tenant: the devices you marked as important, by hand, on the device page. A flagged device has to stay unreachable for five continuous minutes before it qualifies as an alert and takes the protagonist slot, and the copy claims only what was verified: contact lost, cannot confirm it is running. It does not guess a power cut and does not guess trouble. Beyond that, the layer is empty. No risk seenzus infers on its own has earned a place there, and ordinary bad weather is never promoted into it to fill space. The only things allowed to interrupt you at top priority are things you pointed at yourself.

This is where quiet presents its bill. A system built to avoid false alarms leaves the risk of missed ones on the table. People tolerate the two errors asymmetrically: one unsaid thing is usually forgiven, while a few unnecessary interruptions in a row can finish off trust. That is why we guard harder against false alarms. But that is a decision about odds, and the question of responsibility stays open: when a user says “interrupt me less,” what exactly are they giving up? If something that should have been said was not, whose failure is that? We do not yet have an answer we are satisfied with.

Back to the deleted headline. Its problem can be stated more precisely now: it used a number to prove the product’s presence, which is exactly the habit we are asking seenzus to break. What these three months settled is this: in this product, speaking is a scarce act that has to be applied for with evidence. When the home screen is quiet, most of the time that is the constraint working. Whether this quiet will one day fail at a moment that mattered, only longer life in real spaces can tell. We are leaving that question on the table.