A name that needs a footnote is a bug

Our product understands a social situation in three layers, and for weeks we called them the room, the person, you. It read cleanly on the page. Then we had to say it out loud in a sixty-second narration, and the middle name fell apart: the person — which person? Me, or the one standing in front of me? A reader skimming a website resolves that ambiguity from the surrounding paragraph and never notices the wobble. A listener has no paragraph. They just get a half-second of confusion, and in a sixty-second pitch, a half-second of confusion is the whole sentence.

We renamed the middle layer to the encounter. The room, the encounter, you.

The interesting part was not the fix but the rule that generated it. Our three names had been derived inconsistently: two of them named the entity being observed (a room, a person), and the third named the observer(you). Mixed-basis naming always produces one member that collides with something in the reader’s head — here, “the person” collided with the listener’s own referent for a person.

The rule we settled on: name each layer by the question it answers, not by the thing it observes and not by the hardware that answers it.

• The room — who is here, and how open are they?
• The encounter — what is happening between you and the person in front of you?
• You — is there chemistry?

Read as questions, the set is airtight, and no member can be confused with another. It also happens to be hardware-independent, which matters for us because the same layer can be answered by a phone today and a ring or a pair of glasses next year. Naming a layer after a sensor guarantees a rename the moment the sensor changes; naming it after a question survives the hardware.

The failure mode worth generalizing is not “we picked a bad word.” It is that an ambiguous name does not fail loudly. It sits in a document and everyone nods, because everyone silently resolves it using context they happen to have. The ambiguity only surfaces at the moment the context is stripped away — spoken aloud, quoted out of order, read by someone new. So the detection mechanism is cheap and worth making routine: before a term enters the shared vocabulary, say it aloud with nothing around it. If it needs a clause of support to land, it needs a different name.

There is a second-order cost we paid to learn this. A vocabulary is load-bearing across artifacts — the site, the deck, the narration, the internal wiki, the feature list. Changing one word meant a coordinated sweep of all of them, and one deliberate non-change: a passage describing an external study’s measurement levels kept the study’s own wording, because swapping in our brand terms there would have been misquoting someone else’s research. Vocabulary changes propagate to everything you wrote; they must stop at everything someone else wrote.