WHAT EIDOS MEASURES

EIDOS observes each model's token-by-token entropy signature from the outside. It does not see content — it sees structure. The entropy patterns describe how settled the model's word choices were over the recent token window — a reading of the output's structure, not of the model's internal state.

WHY SINGLE OBSERVATIONS AREN'T SIGNALS

A reading can spike or dip in any single token for many reasons — context shifts, sampling noise, transient drift. EIDOS does not trigger on first observations. It accumulates evidence using change-point analysis and only flags when a structural break is sustained long enough to separate from noise. The math waits until the evidence is sustained.

WHAT THE STATES MEAN

STABLE — Coherent internal state. The model is operating within expected entropy bounds.

DRIFT — Entropy pattern shifting. Not yet a signal, but the system is watching.

SUSTAINED BREAK — A sustained shift in the entropy pattern, flagged only after enough evidence accumulates.

READING THE NUMBERS

Coherence — Shows how settled the model's word-choice pattern was over the recent token window. Lower values mean more variation. It does not mean better, worse, true, or false.

Variability — Shows how much the response varied across its token-level uncertainty pattern. Higher means more variation.

Token colors — The colors show where the model had more or fewer competing word choices.

Transition status — Shows whether the measured pattern changed enough to qualify as a transition under the current instrument rules.

Source — Readings come from Top-5 token logprobs, not the full predictive distribution.

These readings describe the response pattern. They do not score truth, correctness, honesty, consciousness, or intent.