Algorithms Architecture Adaptation Evolution
KKmatch/KKOS AI/The engines
The engines

Five engines that read the model — they do not guess at people.

KKOS turns one conversation into an 11-dimension model. These five engines consume that model. Each is a separate mechanism with its own input, its own output, and a claim you can check.

Consumers, not guessers

What each engine actually does

Every engine below takes the same thing as input: a model that already carries evidence and a confidence value. None of them infer a personality from a photo, a job title or a form.

(0)

Personality capsule

MIND · CAPSULE

What does this person look like in one readable paragraph — and which evidence lines produced it?

ReadsThe 11 traits, each with its evidence lines and per-dimension confidence.
OutputA capsule summary plus a pointer back to the sentences that justify it.
(1)

Relationship memory

TGN · TEMPORAL

How did this relationship reach today — and what did each interaction leave behind?

ReadsA sequence of interactions, each stamped with time and the trait state at that moment.
OutputA time-weighted memory representation, not a single snapshot.
(2)

Causal insight

NOTEARS++ · CAUSAL

Which change actually caused which — rather than merely moving at the same time?

ReadsMulti-variable behaviour sequences with their timestamps.
OutputA directed graph stating what influences what, instead of a correlation list.
(3)

Dyadic matching

DYADIC · ATTENTION

How do these two particular people react to each other — not how do they each score alone?

ReadsTwo models at once, each with its own evidence and uncertainty.
OutputA simulated interaction and a per-dimension friction / resonance map.
(4)

Relationship forecast

DSM · SURVIVAL

How far can this relationship travel, and where are the turning points and risks?

ReadsThe relationship state sequence produced by the memory engine.
OutputA survival curve with explicit uncertainty, not a single destiny number.
(5)

Correction loop

HUMAN-IN-THE-LOOP

What happens when you tell it that a trait is wrong?

ReadsYour explicit right / wrong / unsure markers on any dimension.
OutputA re-weighted model with a bumped revision number and a correction log entry.
The input contract

One model in. Nothing guessed.

This is what makes the pipeline auditable: there is exactly one place where a person becomes data, and every downstream engine has to point back to it.

Single sourceEngines never read raw audio, video or free text. They read the structured model — traits, evidence, confidence, uncertainty.
Evidence-boundA dimension cannot enter the model without at least one quoted line from the session that supports it.
Revision-trackedEvery correction bumps a revision number, so a match score can always be traced to the exact model version that produced it.
Checkable claims

Each engine makes one claim you can test

We would rather state something falsifiable than something flattering. Every row below is designed so that a failing result would be visible.

Engine The claim How it would be falsified
Personality capsule Removing every evidence line changes the capsule, because the summary is derived from them rather than written beside them. A capsule that stays identical after its evidence is stripped.
Relationship memory The same two people, re-interviewed later, produce a memory state that differs from the earlier one. A memory representation that ignores timestamps and returns the same state.
Causal insight Reversing the time order of the inputs changes the direction of the detected influence. An undirected result that comes out the same either way.
Dyadic matching A is compared with B and B with A gives the same resonance map, since both sides carry their own uncertainty. An asymmetric result where swapping the two people changes the score.
Relationship forecast The curve widens when the input model has low confidence rather than narrowing to a confident single number. A forecast that reports one precise number regardless of input confidence.
Correction loop Marking a dimension wrong moves its weight in the next model, and the correction log records who changed it and when. A correction that produces no revision bump and no log entry.

Engine names and their abbreviations are ours; the underlying methods draw on published work in temporal graph networks, causal discovery and survival modelling.