LYCEUS. Research notes

AN EXPERIMENT IN MACHINE LEARNING / 001

A model is
never finished.

What if learning didn’t end at training? Lyceus is a small, experimental transformer exploring how a model might change under supervision.

Explore the idea
AN OPEN QUESTION, NOT A FINISHED ANSWERSCROLL TO EXPLORE ↓
01 / THE PREMISE

LEARNING AS A PRACTICE

Less like a product.
More like a question.

Most models are treated as fixed objects: train, deploy, repeat. Lyceus asks what a more deliberate relationship with change could look like.

The project starts with an independently built C++ transformer and looks toward a supervised loop for proposing, testing, and—only when appropriate—keeping changes.

THE MODEL~5.77M

parameters, approximately

THE FOUNDATIONC++

independently built transformer

THE STAGEEarly / research

not a capable assistant

02 / THE ARCHITECTURE

TWO DOMAINS, DIFFERENT OWNERS

Change has
boundaries.

The proposed system separates personal history and profile from reusable model competence. These are research goals, not a demonstrated privacy boundary.

DOMAIN 01

Personal state

History and profile belong to the individual. Integrated associative memory and categorized connections are proposed ways to organize context.

HISTORY / PROFILE / MEMORY
DOMAIN 02

Model competence

Reusable capability lives in the model. Weight edits and growth are being explored as supervised candidate changes, not proven general learning.

WEIGHTS / COMPETENCE / GROWTH

FIG. 02 CONCEPTUAL OWNERSHIP DOMAINS NOT AN IMPLEMENTATION CLAIM

Transformer and source changes are intended to be less frequent, eligible only above a configurable model-size threshold and subject to independent tests. That gate is under development, not a guarantee of correctness.

03 / THE PROCESS

A PROPOSED LIFECYCLE

Change, with
the way back intact.

A candidate is not a conclusion. The intended loop makes evaluation and the ability to return to a checkpoint part of the process.

  1. 01

    Observe

    Gather a signal worth examining.

  2. 02

    Propose

    Describe a possible change.

  3. 03

    Candidate

    Prepare a supervised update.

  4. 04

    Evaluate

    Test before accepting a change.

  5. 05

    Persist

    Keep an accepted checkpoint.

  6. 06

    Reload

    Return to a checkpoint when needed.

Rollback is part of the design.

Independent evaluation, persistent checkpoints, restart and rollback are goals; the current prototype does not establish a secure or fully recoverable lifecycle.

PROPOSED / NOT PROVEN
04 / WHERE WE ARE

A NOTE ON THE WORK

Curiosity, with
clear edges.

01 / DEMONSTRATED

A small beginning.

A from-scratch C++/LibTorch transformer has been trained on public text. Scheduled growth was tested but did not improve exposed-development loss against its control in one seed. A host-scripted synthetic update changed a restricted token answer and reloaded a candidate; it does not prove general learning.

02 / RESEARCH DIRECTION

Much still to explore.

The model is not a coherent assistant. Memory and online growth integration remain ongoing; the prototype is not autonomous, secure, privacy-preserving, or source-self-modifying. Its evaluator is not OS-sealed and feedback is not externally authenticated.