Plots that keep moving between turns and never repeat.

An API story engine for game studios. Walk up to anyone in this town and talk to them. The panel on the right shows what the engine does with each line.

Arrow keys or WASD to walk. Space to talk. On a phone, tap where to go.

Every choice leaves alternatives behind. We keep them pressing on the story.

LLM fiction is predictable. StoryScope (arXiv 2604.03136) measures it. At every choice the model throws away the alternatives it almost took, and they never touch the story again. A human writer's unused options keep shaping what comes next.

We keep rejected choices alive with a weight that decays over the next scenes, and feed them back.

One parameter

It sets how much randomness enters and how long it lingers. The unit is adjustable: a word for voice, a scene for plot.

The product

A story engine over an API, first as an Unreal Engine plug-in, for studios whose dialogue and quests must not repeat.

  1. 1 Propose The model proposes k next moves from the story so far. It does not see the buffer.
  2. 2 Pick The harness picks one uniformly at random, seeded from the run, the prompt and the scene.
  3. 3 Keep The other k-1 moves enter the buffer, tagged with the scene they were rejected in.
  4. 4 Decay Each stored move weighs exp(-(j - scene_added) / tau). Below 0.05 it is dropped.
  5. 5 Write The model writes the scene. Buffered moves did not happen. They act only as pressure: foreshadowing, a character's suspicion.

First test: 10 premises x 3 conditions

8 scenes each, claude-sonnet-5-5, tau = 3 scenes. Bars show the mean rating on the 1 to 5 scale.

Predictability: lower is better

A Plain
2.70 B Decay
1.93 C Flat
2.20

Coherence: higher is better

A Plain
4.40 B Decay
3.90 C Flat
4.10
  • B's predictability was 28% lower than A's, and lower than A's in 8 of 10 stories.
  • B was also less predictable than C in 8 of 10 stories.
  • B lost half a point of coherence against A, and matched or beat A in only 4 of 10 stories.

How we score it

A judge sees the scenes before scene j, starting at scene 3, and predicts scene j. The judge then rates that prediction against the real scene from 1 to 5. The mean of those ratings is the story's predictability, so lower means harder to predict.

The judge also rates the whole story's coherence from 1 to 5. The idea holds if B scores lower than A on predictability with similar coherence. B against C shows whether decay itself matters.

What these numbers are not

This is an early proxy, and coherence has not yet held. The judge is the same model grading itself, and n is 10. The API only allowed its default temperature, so the one-parameter rule is untested. A tau sweep, StoryScope and human raters are next.

About

Farmehr Farhour

Founder

I co-founded Kara Technologies in Auckland and have been its CTO for 8+ years. I built its core platform, which turns text and speech into sign language on real-time 3D avatars. Kara grew from zero to a self-sufficient company, and I led its teams across AI, graphics, linguistics and infrastructure.

That work is close to what game studios need. It joins controlled language generation with character animation in Unreal Engine, in C++ and Blueprints, and it has to hold up in live use. I built the retrieval layer that grounds its generation in a structured knowledge base, and I work hands-on in neural rendering and Gaussian splatting.

I also run a research program on non-deterministic generative models and cellular-automata storytelling. I have been writing code since 2011, and I build end to end, from model to runtime to cloud.

MS in Computer Science (AI and Interactive Intelligence), Georgia Tech. BE in Computer Systems Engineering (First Class Honours) and BCom in Finance and Management, University of Auckland.

Badges

  • Kara named NZ Startup of the Year, NZ Hi-Tech Awards 2026
  • Lead inventor on granted patent US12062125B2 (US, UK, AU) for real-time, kinematics-constrained animation synthesis
  • Kara was the first company in APAC given access to Epic's MetaHumans technology
  • Kara won Best Hi-Tech Solution for Public Good, NZ Hi-Tech Awards 2024, and the AmCham Social Impact Award 2024
  • Alchemist Accelerator (Cohort XXXIV), Startmate (S22), Edmund Hillary Fellowship (Cohort 4)

Vision

Models that can be surprising and still be right. Games are where we start, not where we stop.

Fiction comes first because it has a metric. If a bounded memory of rejected branches cuts predictability without losing coherence, the same loop should propose research directions and product ideas. Those fields ask the same question: is this surprising but valid?

  1. Now Fiction. Measure whether decaying memory makes stories harder to predict and keeps them coherent, with StoryScope and human raters next.
  2. Next Games. A story engine over an API and an Unreal Engine plug-in, for studios whose dialogue and quests must not repeat.
  3. Later Beyond fiction. The same loop proposing research directions and product ideas, if the fiction results hold.

Building a game whose dialogue must not repeat? Talk to us.