The biometric and emotion-inference rules target a specific act: deriving a person's emotional state from involuntary signals — face, voice prosody, physiological data, keystroke dynamics. The signal source is the body. The person did not choose to express what the system extracted.
An EDM record is an interpretation of expressed communication: what a person chose to put into words, in the channel they chose to use. That includes implicature — what a sequence of explicit statements adds up to, even when no single sentence says it. Reading meaning that arrives in words but was never stated outright is not inference from the body; it is interpretation of expression. It is, in fact, the point of interpretation: what mattered most often lives in what wasn't said in so many words.
It would be wrong to claim interpretation only captures what a user "explicitly communicates." Interpretation routinely captures more than the literal words — the homepage demo shows exactly that. What separates it from regulated inference is where the signal comes from (chosen expression vs. involuntary emission), not how subtle the reading is.
Interpretation does not, by itself, make anything auditable. Sealing it at the moment does. Captured and sealed, an interpretation is a dated, schema-bound, inspectable artifact: it can be examined, disputed, and re-appraised over time — with re-appraisals appended, never overwritten — so how a reading changed is itself on the record. Unsealed, the same interpretation is lost to whatever a model later infers or recalls. The record answers the prior question in any dispute about an AI's conduct: not "why did the model decide that?" but "was the significance it computed over captured accurately in the first place?"