Can an AI run the story without taking over the rules?
Enable narrative improvisation without letting the model alter rules, movement, or authoritative campaign state.
CASE FILE. 05 / RPG / GENERATIVE GAME
A solo or two-player virtual tabletop with a persistent lobby, exploration, and combat. The AI game master works through text or voice without controlling the rules.
BRIEF. 01 / PROBLEM AND BOUNDARY
A deliberate view of the current state. Metrics and the hypothesis remain editorial until public evidence can replace them.
Enable narrative improvisation without letting the model alter rules, movement, or authoritative campaign state.
SYSTEM TRACE. 02 / PRIMARY FLOW
Shared tabletop → rules and state engine → narrative context → AI game master by text or voice.
Map, turn, dice, and presence
Actions, movement, and combat
PostgreSQL / cursor and snapshot
Context, text, voice, and fallback
intent → policy → authoritative state → observable effectAn editorial sequence for discussing contract, authority, and effect. It does not represent real production telemetry.
Captures intent, actor, target, and observed version.
action.proposedValidates turn, range, roll, and deterministic consequence.
action.resolved | rejectedPersists the event and derives the new campaign snapshot.
campaign.advancedNarrates the confirmed outcome through text or voice.
narration.emittedDECISION LOG. 03 / TRADE-OFFS
Separate probabilistic narration from deterministic resolution; AI proposes fiction and the system validates actions.
RELIABILITY. 04 / FAIL WITH CONTEXT
These are not published production metrics. They are risks the design must make observable and candidate signals for the next round.
The game master narrates a consequence the rules engine did not authorize.
Players reconnect with different versions of the shared campaign.
Voice latency or interruption leaves a turn without a clear owner.
Narrative proposals rejected or corrected by the rules engine.
Consistency between the recovered snapshot and last confirmed event.
Per-turn latency and frequency of voice-to-text fallback.
INCIDENT DRILL. 05 / CONTAINMENT → RECOVERY → PROOF
Editorial resilience exercises, not real incidents or production runbooks. Each drill turns a failure mode into a response that can be discussed and tested.
DESIGN EXERCISE / REVIEW AGAINST REAL TELEMETRY AND RUNBOOKS BEFORE OPERATING
EVIDENCE. 06 / WHAT CAN BE DEFENDED
Test long-running campaign recovery, voice latency, and narrative consistency in two-player sessions.

INTERVIEW MODE. 07 / DEEPEN THE CONVERSATION
Questions that connect the case to system design, product, and operations without turning a contextual decision into a universal rule.
How to compose campaign context without turning history into authority.
↗How to synchronize two players and an agent that also produces events.
↗How to evaluate narrative quality separately from rule correctness.
↗CASE FILE / 05