The engine's narrative ability
How does the engine manage to ensure that the AI doesn't just take away the monotony and achieve the average 400 words?
The engine works with "textures":
- Physical strain/wear and tear
- Interpersonal friction
- Signs of material wear and tear
- Microrituals
- Negative Space (Missing Elements)
- Continuity scars
- Micro-observation of the environment
- Tense brief exchange
- Relieving long conversations
- Expressive gestures
- Inner Monologue (Diegetic Philosophy)
- Micro-conflicts
- Anomalies (Scenario 1)
Of these 13 possibilities, which are described in more detail in the prompt table, 3 are chosen at random.
Coincidentally not as part of the LLM, but a real coincidence from the backend.
These three options are suggested to the LLM, not imposed, in order to maintain pacing.
The following techniques are also used:
- The “Microscope” rule (Deep Focus).
- Extreme time dilation (decompression).
- Verb-Driven Focus.
- Diegetic translation.
- Vocabulary evolution.
- Kinetic Cadence.
“Prop License” vs. Hard Logic (Limits of Hallucination)
The engine determines what the LLM is allowed to hallucinate and what is not.
Only objects/close surroundings that match the narrative atmosphere may be invented.
E.g. dust, broken glass, scratches. But no sudden new, inappropriate items or area changes.
The Feature Horizon (micro navigation)
Instead of remaining static in a room like in a normal text adventure, you move around in space. Objects disappear and appear, the environment changes.
The engine has a lot of environmental information available.
- Individual sub-areas (zones) of the sector, which can be dangerous, normal or outposts.
- Various specific environmental information such as entities, itchy dust, creaking floor, anomalies, smells etc.
- Fixed environmental information (Area Canon) that the LLM can access permanently.
How can Narravoros tell almost endless stories?
By the User and the LLM as co-authors.
Each story is unique due to the user's input, the LLM's response and the narrative environment.
A similar story can almost be ruled out.
Each area has several environment, entity, and area descriptions from which the LLM can draw freely.
Each area has several available exits after a certain number of rounds.
Which have their own seeds in the database. This means that no area has to be created manually, but is generated dynamically on the first visit.
Furthermore, real random numbers are made freely available to the LLM via the backend. So no estimated or weighted numbers.
This brings even Gemini 3.1 Lite to a good level.