Related research — not part of INSPEKTORIUM
LEM — Lazaros–Eudora Method
An independent research program asking whether repeated interaction with a person leaves reproducible patterns in a language model’s internal representations — across fresh sessions, topics and controlled perturbations.
LEM is a separate research project by the same author. It is not a module, feature or capability of INSPEKTORIUM, and its results stand or fall on their own. The name “Eudora” refers to AI collaboration; it does not name a second author.
Visualisation
A real attractor, for scale.
Two trajectories start almost on top of each other. Both stay on the same butterfly-shaped attractor, but they stop agreeing about where they are. That is what the word “attractor” means: a region trajectories settle into, not a prediction about any single path. In LEM, a user-specific attractor is a hypothesis, and it is still untested.
Toy dynamical system (Lorenz, 1963), drawn only to illustrate the term. It is not LEM data and says nothing about language models.
The research programme
What LEM asks, and where it stands.
Research question
Does repeated interaction with a person produce reproducible patterns in a language model’s internal representations — across fresh sessions, topics and controlled perturbations?
What it does not imply
A frozen model’s activations depend on its current input and available history. The hypothesis does not imply that a person changes the model’s weights during ordinary inference, or that an old conversation survives an empty-context reset.
Proposed measurements
A repeatable user region or signature, user-specific trajectory shapes across sessions, a user-specific attractor that returns after perturbation, and cross-model correspondence. All of these are hypotheses or future work: none has a real-user confirmation.
Current evidence
The first paper (LEM I) contains two synthetic pipeline checks on toy data. In the first, the user signals were built into the simulator, so no LLM users and no identity effects are demonstrated. The second rests on a single pair, without a held-out error estimate.
Null result
The latest pilot returned a narrow null result — reported first.
LEM-II-Light used 108 synthetic dialogues and activations from Qwen2.5-3B-Instruct. At the pre-specified confirmation turn, four classifiers reached 41.7%, 47.2%, 47.2% and 52.8% balanced accuracy, and no nested increment had an adjusted interval strictly above zero. The pilot tested rule-defined regime classification — not user identity, and not attractors.
Known limitations
In LEM-II-Light, every classifier arm included the text itself, the same 18 profiles appeared in every phase, and nearly every answer hit a 128-token cap. The public aggregates do not allow a complete independent reconstruction of the dialogues or activations.
Status
A research program with synthetic pipeline checks and one completed pilot with a narrow null result. No real-user or identity result exists yet.