Observable signal
A fresh-context answer differs materially even when the task facts are held constant.
Context effects
Earlier interaction changes how later instructions are interpreted, weighted, and answered. Useful adaptation and undesirable drift can emerge from the same mechanism.
Underlying MD-00401 / DEFINITION
Context is both capability and pressure.
Conversation history lets a model preserve goals, preferences, terminology, and decisions. The same history can also make local patterns exert more influence than they deserve.
I4MD uses “context effects” rather than claiming a literal model identity. A language model does not need a persistent self for accumulated interaction to change its next output. The relevant observation is behavioral: prior turns alter the distribution of later responses.
The difficult boundary is between personalization and drift. Adaptation is useful when prior information remains relevant and improves the answer. Drift becomes plausible when familiarity, prior agreement, repeated framing, or old decisions continue to shape judgment after their evidentiary value has weakened.
Not the same as
Legitimate continuity, such as preserving an approved decision, a standing preference, or an established project constraint.
Useful control
For consequential judgments, compare the in-context answer with a deliberately context-minimized or independently framed evaluation.
Evidence boundary
Research supports personalization and context-dependent behavior broadly. “Contextual Identity Drift” is I4MD's operational formulation, not an established clinical or scientific category.