Signals: what here is meant for you

The signals button in the header. The screen cuts across all spaces rather than living inside one: the point is not to miss what matters among the places you belong to.

Everything is computed on your device. No text goes to an external model, and what caught your attention never leaves the machine.

Three modes

Settings → Attention.

mode	what it does
Off	nothing is computed, no badge
Minimal	questions, replies to you and mentions of your name
Custom	the same plus what you described yourself

What can be configured

Your names. A comma-separated list: Alice, alice. Explicit, not guessed — guessing name forms fires on ordinary words and turns signals into noise.

What matters to you. One topic per line:

mesh networks and off-grid communication
open-source CAD
collaboration proposals

Today this is matched as phrases. Semantic matching across languages is the next wave, and until it exists, that is what it says.

How often it may interrupt you. A daily maximum, a minimum gap, quiet hours. Everything over the budget goes into the digest — not lost, just waiting.

The inbox

Cards are grouped by space: the eye lands on a place, not on a stream. The filter — all / priority / digest.

Each card: who, an excerpt, reason chips and actions.

Reasons are named in human words:

chip	why the signal
mentioned you	a signed mention
reply to you	a reply to your message
your name	your name in the text
a question	a question
asks for something	something is expected of you
watching	matched what you described
learned	a pattern learned from your feedback
network alert	a message about the state of the network

Approximate reasons are marked and explained on hover. Features are hashed and sometimes collide, so naming the exact word would be a lie. At the bottom of every card is an honest line: processed locally · rules or lexical — whichever layer fired.

Learning and policy are different buttons

They are deliberately not mixed.

Learning — “useful” and “not mine”. This is about what is useful to you, and it changes the model on your device.

Policy — “mute space”. Muting a room is not a statement that people write nonsense in it. It is your choice, not a judgement of the content.

Delete what it learned in the settings erases what was learned. The rules (mentions, replies) keep working: they were not learned, they are given.

What is honest here and what is not

A signal is a pointer, not a replacement for the message. Opening it takes you to the original place.
The model learns only from your feedback and lives only with you.
Explanations are approximate where they are approximate, and it says so.
Learning is not policy: what the model found interesting does not change your rules or your limits.

Mentions, specifically

A mention (@ in a message) is a signed part of the event, not a substring search. That is why it reliably arrives as a signal, does not fire by accident on a similar word, and cannot be forged by editing what is displayed.