# Gaming Multiverse Reasoner

> Pick any set of games and see the hidden paths, bridge games and communities
> that connect them. Every path is computed live in FalkorDB from a frozen
> IGDB snapshot; IBM watsonx (Granite) narrates the results, and every claim it
> makes is validated in code against real path ids.

Live app: https://multiverse.2fy1rnikhcll.us-south.codeengine.appdomain.cloud
Source: https://github.com/Mike-Demo/gaming-multiverse-reasoner
Licenses and full tech stack: https://multiverse.2fy1rnikhcll.us-south.codeengine.appdomain.cloud/licenses

## Data

- 50,217 games from the IGDB API (Twitch), pulled and frozen 14 Oct 2026,
  including the full QueerCade catalogue (753 games).
- Graph "multiverse" in FalkorDB: 84,296 nodes / 1,023,292 relationships.
  Node labels: Game, Company, Genre, Theme, GameMode, Platform,
  PlayerPerspective, Franchise, Collection, Engine.
- Snapshot frozen in IBM Cloud Object Storage as
  snapshot/igdb-raw-2026-10-14-v2.tar.gz with snapshot/manifest-v2.json.
- Communities are PRECOMPUTED (weighted label propagation over similar-game
  links, 6,234 communities) and labelled as such wherever they appear.

## Method (honesty rules)

- Paths are computed live at request time with Cypher; each path response
  includes the Cypher used.
- Hub suppression: Platform and GameMode nodes never appear in paths; Genre,
  Theme and Engine nodes shared by more than 12,000 games are suppressed;
  company bridges cap company degree at 800.
- Path kinds: lineage (remake/remaster/DLC/expansion/version/part-of),
  company-bridge, similar-chain (shortest path over SIMILAR_TO),
  common-neighbor, shared (shared attribute neighbour).
- Narration claims must cite path_ids from the computed result; claims are
  validated in code and dropped if they cite anything else. If watsonx is
  unavailable, template claims are served and labelled source: "template".

## REST API (JSON, no auth, read-only)

- GET /api/health — graph name, game count, connected user.
- GET /api/resolve?q=<title fragment> — matching games (igdb_id, name, year,
  rating_count).
- GET /api/bridges?ids=<id,id[,id[,id]]> — 2 to 4 IGDB ids; returns per-pair
  paths with path_id, kind, templated description and the Cypher used, plus
  per-game community info and stats.took_ms.
- GET /api/random — 4 random games, validated so every pair connects.
- POST /api/narrate — body: a /api/bridges response; returns claims citing
  path_ids, with source "watsonx" or "template".
- GET /openapi.json — OpenAPI 3 description of this API.
- Errors are JSON: {"error": "..."} with an HTTP status code.

## MCP server (for AI agents)

- Endpoint: POST https://multiverse.2fy1rnikhcll.us-south.codeengine.appdomain.cloud/mcp
- Transport: Model Context Protocol over streamable HTTP, JSON responses.
  JSON-RPC 2.0. Send initialize, then tools/list / tools/call.
- Tools:
  - resolve_game {query: string} — find games by title fragment.
  - find_paths {games: [name-or-id, ...]} — 2-4 games by name or IGDB id;
    names are resolved to their top match; returns the full paths payload.
  - random_games {} — a validated random set of 4 connected games.
  - game_community {igdb_id: number} — one game's community, peers, summary.
  - narrate {ids: [number, ...]} — paths plus watsonx narration claims.
- GET /mcp returns server info as JSON. GET /.well-known/agent.json is an
  A2A-style agent card describing the same capabilities.

## Limits

- Read-only demo: no authentication, no writes, no state changes.
- Graph queries are capped by a 5,000 ms FalkorDB timeout; the demo
  four-game query typically answers in under 2 seconds.
- Please be gentle: this is a hackathon demo on shared infrastructure.