Why Not Simply Prompt the LLM for a Random Number?
You could – but it breaks gameplay mechanics completely.
A language model relies entirely on word and context probabilities, even when explicitly asked for a random number.
Statistical Clustering vs. True Entropy
An LLM predicts the next most probable token. Ask it for a random number between 1 and 100, and it will disproportionately output 42 or 7 – simply because pop culture (The Hitchhiker's Guide to the Galaxy) and human cognitive biases dominate its training corpus.
That is not true randomness; it is rigid statistical frequency. In an unpredictable RPG, this quickly causes repetitive patterns and predictable narrative paths.
The Consequence: Repetitive Narrative Cadence
Without unyielding entropy triggers, the model drafts every turn with identical rhythm: Action → Reaction → Expected outcome. Narrative friction gets smoothed over because the model naturally gravitates toward the most harmonious and probable path.
The Server RNG as an Incorruptible Arbiter
Before each turn, Narraboros provides the language model with several true random numbers drawn from dedicated server pools:
This governs both the quantity of discovered items and the environment itself: Finding three gemstones is not an LLM hallucination, but the result of verified server-side seed resolution.
Diegetic Soft-Fails Over Smooth Passivity
The decisive advantage: The AI is forced away from always picking the easiest, most agreeable outcome.
Tied into narrative textures, entropy obliges the model to interject trivial yet deeply atmospheric friction – such as a stuck zipper, a boot slipping on moss, or a sudden creak in the floorboards.
Dynamic Temporal and Cadence Shifts
The engine leverages the RNG pool to modulate dramatic pacing: One turn might dilate time into excruciating sensory detail (Time Dilation before danger); the next might compress miles into clipped sentences. Every round remains unpredictable and literarily alive.