Expression Membrane

brain/behavior/express_to_user.py is the one door: every artifact a person sees — a live reply, a note, a desktop file, a dashboard announcement, an OS notification — is composed here from a Motive (intent + felt state). Nothing person-facing is ever populated by scraping internal representation.

Why one door

Before the membrane (fixed 2026-06-14), emitters scraped working memory directly, so internal fragments could leak to the person verbatim — unfiltered, incoherent, or not something the runtime "meant to say." The membrane inverts this: expression starts from an intent, not from whatever happens to be in memory.

Two sides, one door

INTENT (motive)  ──►  express_to_user(...)  ──►  channel

Inside the door:

• Composition goes through expression.express() — affect plus the learned vocabulary, congruence-enforced (the text must match the actual internal state; Rogers 1959 is the cited inspiration for congruence).
• The speakability invariant is enforced in exactly one place: whatever channel the output takes, it passed the same checks.
• Channel fan-out — live reply, note, file, announcement, notification — happens after composition, so every channel gets the same quality of output.

What this means in practice

• The reasoning layer describes its own state in qualitative terms (see Control Signals — the reasoning layer never receives raw numbers), and the membrane keeps person-facing text consistent with that state estimate.
• Delivered notes and replies are recorded as effects on the effect ledger, so communication counts as production.
• Anything that bypasses express_to_user to reach a person is a bug by definition — there is one door on purpose, which makes auditing person-facing output tractable.

Code pointers

• brain/behavior/express_to_user.py — the door (design: EXPRESSION_MEMBRANE_FIX_PLAN, 2026-06-14)
• brain/think/speech_builder.py, speech_coherence.py — utterance construction and checks
• frontend/src/pages/Face.tsx — where expressed output lands in the UI