Conversation knows the scene.
Believable contextual dialogue draws from individual knowledge, collective knowledge, world state, and environment—not an isolated prompt.
Primary talk ↗A proposition for responsive worlds
The world should remember what you meant.
Not more dialogue. A sentence that becomes intent, trust, risk—and a consequence the world can test later.
Make the promiseBelievability comes from coordination: conversation, gaze, breath, spacing, sound, rules, and the state that survives after the moment ends.
Language is not another button. It should enter the story.
Type your own line or choose one. This page uses a small local ruleset—not a live model—to demonstrate the authority boundary honestly.
Original interactive scene / damaged transit station
A promise now. A test twenty minutes later.
Before words
The courier keeps one hand on the satchel. Their breath catches when your boots stop.
Act I / the promise18:42
“If the patrol finds me, the route dies here.”
The courier is injured and hidden beside the closed ticket office. They can hear you. Concealment is an authored possibility.
Choose a line, or write your ownYour words→Intent proposal→Authored validator
No state has been changed.
Act II / the test20:02
“Anyone come through the east platform?”
A patrol blocks the service corridor. The promise is in the canonical event ledger. What you say now is tested against it.
Answer the patrolAnswer→Bound intent→Existing promise
No consequence has been committed.
Observable consequence / authored
Canonical event ledger / local session onlyNo events
These are lessons from authored craft and conventional game AI—not evidence that the referenced games used modern generative AI.
What immersion already knows
The relationship has to survive the cutscene.
The body answers first.
Gaze, posture, breath, locomotion, and physical distance keep emotional state legible while the player remains in control.
Official feature ↗Companionship is mechanical.
Where an ally stands, when they help, what they notice, and whether they respect stealth all characterize the relationship.
Primary talk ↗Consequence returns through form.
A moved chair, withheld aid, changed route, altered gaze, or familiar object can remember more convincingly than an exposition screen.
Primary talk ↗The missing middle
Recognition is not consequence.
A fluent reaction is still shallow when no authorized state, authored possibility, or later memory sits behind it.
Player words→Model output→Spectacle / forgotten
The common demo: language becomes an alternate controller.
Words→Typed intent→Validation→Committed consequence
The proposed boundary: the interpreter can suggest meaning. Only tools and authored rules commit reality.
Responsive does not mean unrestricted. The dramatic possibility space, canonical state, performance rights, and failure behavior remain designed.
The human boundary
Authored characters. Deterministic consequences.
People define what may become true.
Writers and narrative designers own character knowledge, forbidden facts, relationship arcs, story invariants, and the authored consequence set.
The validator owns the commit.
Typed allowlisted tools check witnesses, knowledge, timing, permissions, and contradictions. No generated prose or arbitrary code becomes canonical state.
Consent is part of the architecture.
Any generated voice or performance requires specific rights, informed consent, compensation, auditability, safety review, and a complete deterministic fallback.