Algorithms Architecture Adaptation Evolution
KKmatch/KKOS AI/Adaptation
Adaptation

One core model. Three domains. And a loop that lets you argue with it.

The traits do not change between products — only the projection does. What makes that safe is the correction loop: the model can be overruled by the person it describes.

Domain projection

Same 11 traits, three adaptation layers

A projection layer decides which dimensions matter, how they are weighted, and what vocabulary the output uses. It never rewrites the underlying trait or its evidence.

TOC · SOCIAL APP

Social matching

Projection: recommendation engine

A profile that exists from minute one, so "swiping" turns into "talking".

  • A first profile in 30 seconds, recommendations from sign-up
  • 11 traits × relationship memory — matching on chemistry, not tags
  • Dyadic matching, pairing the two people who actually resonate
  • Relationship health tracking, lifting both retention and activity
TOC · COUNSELING

Relational counselling

Projection: relational assessment

Translating feelings that cannot be said out loud into language that can be examined.

  • Causal insight locating the crux, answering the "why" behind the feeling
  • Stage forecasting, warning about conflict and drift before they arrive
  • A structured profile that helps a counsellor reach a client's frame faster
  • Privacy first — modelling stays on the server, nothing leaks
TOB · BEHAVIOR INSIGHT

Behaviour insight

Projection: organisational intelligence

Reading cooperation and friction between people, instead of between job titles.

  • Team complementarity, flagging collaboration conflicts early
  • Account manager ↔ client style fit, improving close rates
  • Cultural fit assessment, reducing avoidable attrition
  • Causal chains that locate the person actually moving performance
The correction loop

You can overrule the model, and it has to show its work

Most systems treat a user's disagreement as noise to be smoothed over. Here it is the highest-weight signal available — because the person being modelled is the only one who can actually check the claim.

STEP 01
Read a traitYou see the dimension, its value, its confidence, and the quoted lines it was derived from.
STEP 02
Mark it wrongRight, wrong, or unsure. Unsure is a real answer and is stored as one.
STEP 03
Re-weightThe correction changes how much that dimension counts next time, rather than being discarded.
STEP 04
Log and bumpThe model's revision increments and the change is written to a correction log, so any later match score can be traced to it.
An open question we publish

Baseline against yourself, not against the average

KH-001

Personal baseline is more suitable for a long-term human model than population averages. Individual behaviour should be modelled relative to a person's own historical baseline, not a generic mean.

This is an original, falsifiable hypothesis from the KK research programme — it is stated as a hypothesis, not as a shipped guarantee. Its current stage and the evidence we would accept against it are tracked in the public hypothesis database.
Read the hypothesis database →

We separate what is proven from what we merely propose. Where a claim is still a hypothesis, it is labelled as one — including on pages like this.

Next ← Architecture Evolution →