LATTICE [LA09A]

LATTICE (Living Architecture Transposer) turns FACET storyworlds into inhabitable worlds: Marble locations, TRIPO bodies, and Reactor Helios video — all governed by canon, not prompts.

Comment

LATTICE — Devpost Project Story (LA09A)

Elevator pitch (184 characters)

LATTICE (Living Architecture Transposer) turns FACET storyworlds into inhabitable worlds: Marble locations, TRIPO bodies, and Reactor Helios video — all governed by canon, not prompts.

Inspiration

Between us we had two halves of the same problem. William's background is in storyworld authoring — watching richly built story bibles die on the page because "experiencing" them meant bespoke game development nobody had time for. Benjamin's background is systems and multi-agent design — watching generative tools produce beautiful, disconnected outputs with no shared source of truth governing them. World Labs Marble, TRIPO, and Reactor all generate incredible substrate — worlds, bodies, live video — but none of them know anything about why a room matters or who a character is. We wanted to prove that a structured story bible, authored once in AOHK's own FACET platform, could be the thing that actually builds and governs a world — not just describe one.

What it does

LATTICE transposes a FACET-authored "cosmology bundle" — a JSON export of a storyworld's characters, locations, objects, systems, and events, typed and related in an ontology — into an inhabitable, interactive world, in six stages:

  1. FACET authors the storyworld as a structured knowledge graph and exports it as JSON — not prose.
  2. LATTICE's asset generators (World Labs + TRIPO) ingest the bundle and transpose every entity into an asset: locations, bodies, and placement.
  3. The Playtest Workbench places characters and objects in their generated world, lets an operator assign triggers, and captures the "play" of their movements and triggered actions. With actions turned off, our novel Azimuth Capture tool uses those same placed assets as seed data for interactive video generation.
  4. FACET's structured Event & Character Cosmology Cards — the bundle's Character, Object, Location, and Event data — load into a structured form that produces a distinct Reactor instance per scene.
  5. LATTICE reference images: TRIPO's own 2D reference images, plus the Azimuth Capture tool's 2D captures of the World (Marble) and its placed Characters/Objects, are seeded as reference imagery alongside FACET's text.
  6. Reactor Helios takes the structured event text from the FACET bundle — environment and action descriptions — and turns them into the triggers that drive interactivity inside the generated video itself.

The result: sponsors provide text-and-image-driven visual generation (Marble for worlds, TRIPO for 2D reference and 3D embodiment, Reactor Helios for interactive video); FACET brings the storyworld engine, ontology, structured bundle, and narrative of events; LATTICE is what we built to connect them — canon-grounded actions, skill-gated interactions, a systems-rule director, and placement and triggers derived from the story graph itself, not hand-scripted.

How we built it

  • FACET authored and exported the cosmology bundle (Contechs ontology) that both of our hackathon submissions consume.
  • Frontend: Vite + React + Three.js.
  • World Labs Marble generated every location as a walkable, gaussian-splat world.
  • TRIPO ran two ways: TRIPO 2D for text-to-image reference generation, TRIPO 3D for image-to-object embodiment and rigging.
  • The Playtest Workbench handled placement, trigger assignment, and capture of character "play" inside the generated world.
  • Our Azimuth Capture tool — built fresh this weekend — captured consistent 2D references of each world and its placed characters/objects, feeding both TRIPO's reference pipeline and Reactor's seed data.
  • Reactor Helios consumed FACET's structured event text (environment + action descriptions) as triggers, producing interactive, canon-driven video per scene.
  • Anthropic Claude, via our Vercel-hosted Eve agent framework, grounded actions and dialogue in each character's canon fields.
  • Infrastructure: Vercel for hosting and the Eve agent framework, Google Cloud for asset and data infrastructure.

Challenges we ran into

  • Getting a TRIPO rig to actually animate cleanly once embodied was a real integration risk — we validated the pipeline on one character before building the rest on top of it.
  • Designing the Azimuth Capture tool so its output was consistent enough to serve two very different downstream consumers at once — TRIPO's reference pipeline and Reactor's interactive video seeding.
  • Mapping FACET's structured event text cleanly into Reactor Helios triggers, so "what the event says happens" and "what the generated video lets you do" stayed in sync rather than drifting apart.
  • Standing up hosting and cloud infrastructure (Vercel, Google Cloud, the Eve framework) alongside the creative pipeline, on the same two-day clock.
  • Coordinating a two-person team split between on-site and remote across a compressed weekend.

Accomplishments that we're proud of

  • Nothing in the world is hand-placed or hand-scripted: every position, trigger, and interactive video cue traces back to a relationship or event in the original canon.
  • A single Azimuth Capture technique that turned out to generalize — it seeds both LATTICE's interactive Reactor Helios experience and our sibling submission's automated narrative pipeline.
  • A clean three-way separation of what the sponsors' tools provide, what FACET already brought to the table, and what we actually built at the event — the moat made legible, not just claimed.
  • Going from a raw JSON bundle to a fully interactive, canon-governed world with live generative video in a single hackathon weekend.

What we learned

  • Sponsor world-generation tools are extraordinary at substrate and know nothing about meaning — the gap between "a space" and "a world that knows things" is exactly the layer we chose to build.
  • The same reference-capture infrastructure can seed genuinely different production modes — a live interactive experience here, a fully automated narrative pipeline in our sibling project — without touching how the underlying story was authored.
  • Consistency discipline in how references are captured mattered as much to a convincing result as raw model quality did.

What's next for LATTICE [LA09A]

  • Generalize the Azimuth Capture tool into a first-class LATTICE feature usable on any location, not just this build.
  • A frozen-exhibit publish tier: a static build streaming assets from our cloud infrastructure, so a produced world stays walkable and reactive without our full agent backend running.
  • A live-hosted tier with gated, budgeted Reactor sessions, so the interactive video experience can go from demo to something anyone can actually visit.
  • Expanding the combat, crafting, and language skill channels beyond the stubs we scoped out of the hackathon build.
  • Write-back to FACET, so what an operator discovers by playtesting a world can inform the canon it came from.

Built With

Share this project:

Updates