01 / PERSONALUSER-SPECIFIC
Personal state
History and profile belong to the individual. Associative memory and categorized connections are proposed ways to organize context.
~/lyceus / README [ research note 001 ]
research brief / overview
A model exploring
how to keep learning.
Lyceus is an experimental transformer built from scratch. It explores what it would take for a model to change over time—carefully, under supervision, and with a way back.
LYCEUS / RESEARCH NOTESSMALL MODEL · OPEN QUESTIONS · NO PRODUCT CLAIMSEST. 001
01 / ARCHITECTURE
Two domains.
Different ownership.
Personal context and reusable competence should not be treated as the same kind of state. This is the intended design, not a proven privacy boundary.
FIG 01 / OWNERSHIP MODELCONCEPTUAL · NOT TO SCALE
02 / SHAREDREUSABLE
Model competence
Reusable capability lives in the model. Editable weights and growth are being explored as supervised changes, not proven general learning.
NOTE / 01
Transformer and source changes are intended to be less frequent, eligible only above a configurable model-size threshold and subject to independent tests. The gate is under development—not proof that a change is correct.
02 / CHANGE LIFECYCLE
Propose. Test.
Keep the way back.
A candidate is not a conclusion. The intended lifecycle makes evaluation, persistence and recovery part of model change.
FIG 02 / SUPERVISED UPDATE LOOPPROPOSED DESIGN
- 01
Observe
Gather a signal.
- 02
Propose
Describe a change.
- 03
Candidate
Update or grow.
- 04
Evaluate
Test independently.
- 05
Persist
Save a checkpoint.
- 06
Reload
Restart or roll back.
Recovery is a design goal, not a guarantee.
[ PLANNED ]Independent evaluation, persistent checkpoints, restart and rollback remain work in progress. The prototype does not establish a secure or fully recoverable lifecycle.
03 / RESEARCH STATUS
Evidence over
assumptions.
A small research model is not a capable assistant. Here is the boundary between what has been tested and what remains to be built.
01 / OBSERVED[ TESTED ]
Small-scale experiments
A from-scratch C++/LibTorch transformer has been trained on public text. Scheduled growth 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. This does not prove general learning.
02 / IN PROGRESS[ RESEARCH ]
Integration & safeguards
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.