EXPERIMENT_015 // NARRATIVE.MICRO-WORLD
A small navigable brand world — three scenes, dialogue and environment generated live by an LLM in response to your choices, held to an authored style/lore block so it never drifts off-brand.
LOADING EXPERIMENT...
Michia Rohrssen's Koto — a language-learning RPG selected for Supercell's AI gaming accelerator, one of 30 globally — treats the player's journey as co-created in real time by an agent, not scripted in advance. Applied to brand experience: a small navigable world where dialogue and environment respond to the visitor, held inside hard authored rails so the world never drifts off-brand. This is deliberately narrower than "When Agents Render the Interface" (2026-05-17): that piece is UI assembled from declarative specs (EXP_006, Generative UI Renderer). This one is narrative *space* — a place the visitor moves through, not a layout the agent builds.
HOW IT WORKS
A versioned world fixture defines three scenes, their default palette/weather/layout, and an authored style/lore block — the Signal's voice, tone, and hard boundaries. Every scene offers authored choice buttons (always available, never dependent on the LLM) plus an optional freeform "speak to the Signal" input.
Every LLM completion passes through a Zod schema (dialogue length capped, environment deltas allowlisted to values already defined in the world) and a separate content-rail check for jailbreak phrases, fourth-wall breaks, and URLs. Either failing routes the turn to a retry, then to the authored fallback line — never a blank or broken scene.
With no API key configured, or after two failed validation attempts, the experiment plays entirely on authored fallback dialogue and environment deltas — fully navigable end to end. A visible per-visit token budget caps spend client-side; once it's reached, every remaining turn uses the same scripted content.
WHAT THIS PROVES
Brand coherence and generative freedom aren't in tension if the rails are structural, not just instructional. A closed environment-delta enum and a content-level rail check hold the LLM to the brand's world even when the visitor actively tries to break it.
Agent-driven narrative space is a distinct proof from agent-driven UI (EXP_006): the same schema-validate-or-fallback discipline applies, but the output is a place with a voice, not a layout with components.