One generation call, every turn.

A hard safety gate reads the turn before anything else and can answer without the model at all. If it clears, a small model and a graph traversal run side by side, and their output conditions exactly one generation call. There is no loop, no planner and no second pass.

TURN

01 / GATE

Safety

fires

Fixed reply

model not called

clears

In parallel

02 / STATE

Small model

03 / RETRIEVE

≤ 3 facts

04 / GENERATE

One call

Persist

facts supersede

Every path through this diagram calls the generation model either once or not at all. The dashed branch is the gate firing: a fixed, human-written reply, returned without a prompt for anything to argue with.

01 / Gate

Safety decides whether the model writes at all.

Most safety systems review what a model already produced. Reviewing output means the output existed. Ours runs first, and when it fires everything downstream is skipped.

01

It runs first.

Before the state model, before retrieval, before generation. A turn that trips the gate never reaches the model that writes.

02

It can answer alone.

The reply is fixed and written by a person. No prompt, no temperature, nothing to talk it out of.

03

It is not therapy.

superemotion is not a clinician. When something needs a human it says so plainly and gets out of the way.

02 / State

Four fields. None of them an emotion.

A fast small model reads the turn while retrieval runs. It produces a compact object that conditions the reply — and it has no vocabulary slot for naming how you feel, so it cannot.

intensity
0.0 – 1.0
How much weight the turn is carrying.
trajectory
rising | flat | settling
Which way it has moved since the last turn.
wants
space | company | advice | nothing
What the turn is asking for, if anything.
mode
sit | reflect | ask | suggest
How long and how forward the reply is allowed to be.

The omission is the design. A field called emotion would get filled in on every turn, and a system that labels you is a system you start performing for.

03 / Budget

A fixed call count gives a fixed ceiling.

Because nothing in the turn can decide to run again, the worst case is the sum of four stages rather than an open-ended distribution.

gate
~15 ms
Rules, then a small classifier. Runs on every turn.
state
~180 ms
Small model, concurrent with retrieval.
retrieve
~40 ms
Graph traversal, capped at three facts.
generate
~600 ms
To first token. The only large-model call.

Design budgets, not measurements. State and retrieval overlap, so the path to first token is the gate plus the slower of the two plus generation — not their sum.

04 / Constraints

Three things it will never do.

Each of these would make the system more capable on a benchmark and worse at the only thing it is for.

01

No tool loop.

Nothing in the turn can decide to call itself again. The number of model calls is fixed before the turn starts, which is why the latency has a ceiling instead of a distribution.

02

No second pass.

The first reply is the reply. A rewrite pass would smooth exactly the roughness that makes a short answer sound like a person rather than a summary.

03

No planner.

The state object is four fields produced by one small model, not a chain that reasons about what to do. Deliberation is where assistants start sounding like assistants.

One call is enough when you know what to put in it.

The graph is what makes the difference.

Architecture · superemotion