How it works

AI knows the request.

“Help me with this.”
user · just now

Four words. That’s everything the model gets.

01 / the person behind the request

Behind those four words is a person at work.

“Help me with this.”

The request stays the same.
What surrounds it doesn’t.

the person

Who is working: role, expertise, goals.

the work context

What “this” actually is: the project, the task, what matters now.

the way of working

What kind of help fits: format, depth, style.

the current state of work

What is happening right now: momentum, friction, focus.

None of this is in the request. All of it decides whether the help is useful.

02 / context composer

Same request. Different understanding.

Flip what the AI knows. Watch the answer change.

CONTEXT COMPOSER
USER MESSAGE · ALWAYS ON
“Help me fix this bug.”
SITUATION AWARENESS · produced by Qualia-1, on device
AGENT RESPONSE
HARNESS DIRECTIVE
Same facts. Different moment. Flip it.
Generic
 
on device
Pre-composed demo: responses are scripted to illustrate the mechanism. Qualia-1 produces the situation awareness — structured context across preferences · work-state · org-facts · history · agent-policy — and composes the window.
03 / the orchestration pass

Gather. Reconcile. Compose.

Before each agent step, Qualia-1 runs a single pass over every candidate context unit — the structured context, plus the live work-state.

⟡ work-state: deep focus · interruptibility bad
COMPOSED WINDOW · PRIORITY ORDER
token budget6,400 / 8,000
HELD BACK — AVAILABLE FOR LATER STEPS

Everything that could go in the window flies in — tagged with where it came from and what kind of fact it is.

04 / both halves

A composed context and a harness directive — from the same pass.

HALF 01 · CONTEXT

What the model sees.

The units selected for this step, in priority order. Reconciled facts with provenance attached. Sized for the window’s budget — held-back items stay available for later steps.

HALF 02 · HARNESS

How the answer is delivered.

depth — terse fix, or scaffolded explanation. format — bullet, paragraph, code, table. tools — which to expose, which to suppress. timing — surface now, queue, or defer.

You already felt this above: flipping work-state never changed the facts — it changed the directive.

05 / on-device by design

Where it runs, and what it doesn’t do.

The model never reads the contents of what the user is producing; it reads the shape of how they’re producing it.
On-device is not a feature. It is the condition that makes it deployable at all.
×No camera, no microphone, no biometric sensors.
×No reading of document, code, email, or chat contents.
×No emotion inference. No personality scoring. No employee evaluation.
Work-state from behavioral signal only — keystroke rhythm, mouse motion, context switching, interaction trajectory.
Output is situation awareness — structured context across preferences · work-state · org-facts · history · agent-policy — plus a typed harness directive. Nothing else.
06 / the formula
requestanswer
request+preferences+org-facts+history+agent-policy+work-statehelp that fits

Smarter AI needs a better understanding of the human.

Qualia-1 is how AI gets it — from the interaction itself.