Our Foundation paper is coming to arXiv →
Ammonix

Technology

The Architecture

  1. 01
    φ(xt)\varphi(x_t)

    Mechanistic World Model

    Fixed code turns the raw input into the features experts already use. Nothing here is trained.

  2. 02
    ut=Λ(φ(xt))u_t = \Lambda(\varphi(x_t))

    Classifier Swarm

    Small classifiers, each recognizing one pattern, together place the new example next to the past examples it resembles most.

  3. 03
    U~(Λ)\tilde{U}(\Lambda)

    Knowledge Universe

    Every example the system has ever seen is kept here, with what it was, how it ended, and the record to prove it.

  4. 04
    E(ut,a)\mathcal{E}(u_t, a)

    The Ether

    For each action the system could take, it looks up how that same action ended in the most similar past examples.

  5. 05
    SS

    Regional Skills

    Each region of the map has its own skills: rules that remove forbidden options, and rankings built on the evidence from past examples. They change only offline, after passing sealed tests.

  6. 06
    LL

    Frozen Language Model

    It turns the finished decision into words a person can read, and it is never retrained.

Inside the system

The Knowledge Universe

This is how Ammonix stores what it knows. Each past example is a point on this map, and similar examples sit close together. A new example is shelved the way a librarian shelves a book, next to the ones most like it, and the outcomes of those neighbors are the evidence behind the decision.

knowledge-universe · live view
example ended well example ended badly◎ live example u_ta₁ … a₄ possible actions

The platform

Every Ammonix Agent Is Built With AmmonixCode