CASE FILE. 05 / RPG / GENERATIVE GAME

AI_VTT

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.

  • TypeScript
  • PostgreSQL
  • Voice AI
  • Rules engine
View all projects

BRIEF. 01 / PROBLEM AND BOUNDARY

The product in one read.

A deliberate view of the current state. Metrics and the hypothesis remain editorial until public evidence can replace them.

TEST QUESTION

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.

TABLE
SOLO + DUO
EDITORIAL LABEL
CAMPAIGN
ASHGLASS SIGNAL
EDITORIAL LABEL
CAPABILITIES
5
frontend / backend / data / realtime / ai
PUBLIC DOOR
NO
PRIVATE SURFACE

SYSTEM TRACE. 02 / PRIMARY FLOW

Boundaries before frameworks.

Shared tabletop → rules and state engine → narrative context → AI game master by text or voice.

  1. 01
    TABLE

    Shared tabletop

    Map, turn, dice, and presence

  2. 02
    RULES

    Authoritative engine

    Actions, movement, and combat

  3. 03
    MEMORY

    Campaign state

    PostgreSQL / cursor and snapshot

  4. 04
    GM.AI

    Narrative layer

    Context, text, voice, and fallback

intent → policy → authoritative state → observable effect
EXECUTION TRACE / ILLUSTRATIVE

One path. Every decision in view.

An editorial sequence for discussing contract, authority, and effect. It does not represent real production telemetry.

  1. TABLE

    Captures intent, actor, target, and observed version.

    action.proposed
  2. RULES

    Validates turn, range, roll, and deterministic consequence.

    action.resolved | rejected
  3. MEMORY

    Persists the event and derives the new campaign snapshot.

    campaign.advanced
  4. GM.AI

    Narrates the confirmed outcome through text or voice.

    narration.emitted

DECISION LOG. 03 / TRADE-OFFS

Every choice costs something.

Separate probabilistic narration from deterministic resolution; AI proposes fiction and the system validates actions.

  1. 01 / AUTHORITY

    Rules outside the model

    GAIN
    Fiction can improvise without changing dice, health, or inventory.
    COST
    Narrative must translate engine rejections naturally.
  2. 02 / CONTEXT

    Campaign cursor + snapshot

    GAIN
    Recovery does not require resending the entire session history.
    COST
    Summary and state must evolve without contradicting past events.
  3. 03 / INTERACTION

    Voice with text fallback

    GAIN
    The table remains operable when synthesis, network, or model is slow.
    COST
    Multimodal turns require careful cancellation and coordination.

RELIABILITY. 04 / FAIL WITH CONTEXT

Failure mode. Signal back.

These are not published production metrics. They are risks the design must make observable and candidate signals for the next round.

FAILURE MODES
  1. 01

    The game master narrates a consequence the rules engine did not authorize.

  2. 02

    Players reconnect with different versions of the shared campaign.

  3. 03

    Voice latency or interruption leaves a turn without a clear owner.

OBSERVABILITY CANDIDATES
  1. 01

    Narrative proposals rejected or corrected by the rules engine.

  2. 02

    Consistency between the recovered snapshot and last confirmed event.

  3. 03

    Per-turn latency and frequency of voice-to-text fallback.

INCIDENT DRILL. 05 / CONTAINMENT → RECOVERY → PROOF

When the hypothesis breaks.

Editorial resilience exercises, not real incidents or production runbooks. Each drill turns a failure mode into a response that can be discussed and tested.

DRILL-01The game master narrates a consequence the rules engine did not authorize.SIMULATION / NATIVE

Response to: The game master narrates a consequence the rules engine did not authorize.

CONTAIN
Interrupt divergent narration and preserve the deterministic event as the only state authority.
RECOVER
Generate a correction or text fallback from the confirmed result without changing health or inventory.
PROVE
Validate that narration, event, and snapshot point to the same version and rule consequence.
DRILL-02Players reconnect with different versions of the shared campaign.SIMULATION / NATIVE

Response to: Players reconnect with different versions of the shared campaign.

CONTAIN
Reject new proposals until each participant acknowledges the current campaign cursor.
RECOVER
Send the snapshot plus later events and rebuild table, turn, and presence before reopening actions.
PROVE
Compare the recovered state version and hash across server and every connected player.
DRILL-03Voice latency or interruption leaves a turn without a clear owner.SIMULATION / NATIVE

Response to: Voice latency or interruption leaves a turn without a clear owner.

CONTAIN
Cancel pending synthesis, freeze automatic turn passage, and display the confirmed result as text.
RECOVER
Return control to the current player and begin new narration only with a current turn token.
PROVE
Confirm a single turn transition, no late audio, and continuity through the text fallback.

DESIGN EXERCISE / REVIEW AGAINST REAL TELEMETRY AND RUNBOOKS BEFORE OPERATING

EVIDENCE. 06 / WHAT CAN BE DEFENDED

Facts in view. Gaps too.

Test long-running campaign recovery, voice latency, and narrative consistency in two-player sessions.

VERIFIED FACTS
  • TypeScript + PostgreSQL
  • Game master by text or voice
  • Solo and two-player campaigns
  • Authoritative rules outside the AI
Four people play a fantasy tabletop RPG with miniatures, dice, and teal lighting.
ASHGLASS / WORLD BUILDING / Lobby, exploration, combat, and recovery in The Ashglass Signal. AI narrates; rules, movement, and state remain authoritative.

INTERVIEW MODE. 07 / DEEPEN THE CONVERSATION

Three points to open on the board.

Questions that connect the case to system design, product, and operations without turning a contextual decision into a universal rule.

  1. 01

    How to compose campaign context without turning history into authority.

  2. 02

    How to synchronize two players and an agent that also produces events.

  3. 03

    How to evaluate narrative quality separately from rule correctness.

CASE FILE / 05

Architecture is a hypothesis.
Real use is the test.

Meet the builder