Back to browse
We Gave an LLM Adventure Engine a Body, Now It Feels Exhausted

We Gave an LLM Adventure Engine a Body, Now It Feels Exhausted

by oopismcgoopis·Feb 25, 2026·1 point·0 comments

AI Analysis

●●●BangerZero to OneBig BrainWizardry

LLM with simulated exhaustion state—forces grounded prose when stressed, prevents inventory hallucinations.

Strengths
  • Custom fine-tuning removes RLHF sycophancy and retrains on sensory prose instead of helpfulness
  • Bidirectional coupling between simulated body state and text generation—not a prompt wrapper but genuine constraint
  • Open-source and runs locally; no cloud API dependency, full terminal interactivity
Weaknesses
  • Execution limited to text adventure domain; generalizability to other LLM tasks unclear
  • Niche audience overlap (local LLM + interactive fiction + philosophy); adoption risk high
Category
Target Audience

Interactive fiction enthusiasts, local LLM hobbyists, game developers experimenting with AI agents

Similar To

Dwarf Fortress (procedural simulation grounding narrative) · AI Dungeon (interactive fiction with LLMs) · Custom character AI systems with state management

Post Description

'Sup HN,

I love local LLMs for interactive fiction, but I got tired of the "Naked Model" paradox. If you tell a standard RLHF'd model "I drink the health potion," it will gleefully describe you drinking it—even if your inventory is completely empty. They have no object permanence and are biased to be sycophantic assistants.

To fix this, I built BoneAmanita, an architecture that puts a fine-tuned LLM inside a Python-simulated body.

How it works: It’s a two-part system:

The Brain (GGUF): A custom 3B model (fine-tuned on Llama 3.2 via Unsloth). I scrubbed out the "helpful assistant" RLHF and trained it strictly on atmospheric, sensory, and philosophical prose.

The Body (Python Engine): A local terminal hypervisor that runs a physical simulation. It tracks variables like "ATP" (stamina), "ROS" (trauma), "Voltage," and Cortisol.

The Feedback Loop: The Python engine intercepts every turn and dynamically rewrites the LLM's system prompt based on its metabolic state. If you stress the engine out with high-entropy actions, its simulated Cortisol spikes. The Python engine injects a strict prompt override forcing the LLM to output short, fragmented, defensive sentences. It literally gets exhausted.

Solving the Hallucination Problem (The Gordon Shock): To enforce hard physics, the Python engine manages a strict inventory state. If you attempt an impossible action (e.g., washing a car in a forest), an internal interceptor ("Gordon") catches the premise violation before the LLM can "Yes, and..." you. Gordon violently injects a CRITICAL OVERRIDE into the context window, forcing the LLM to coldly reject the action and ground you in reality.

It boots into 4 modes (Adventure, Conversation, Creative, Technical) depending on how strict you want the physics engine to be.

You can pull the brain straight through Ollama: ollama pull hf.co/aedmark/vsl-cryosomatic-hypervisor

And run the Python hypervisor here: https://github.com/aedmark/BoneAmanita

It’s completely free, local, and released under The Unlicense.

Come play!

Similar Projects

AI/MLMid

100% LLM accuracy–no fine-tuning, JSON only

Ancient Rome Q&A benchmark shows 81pp accuracy lift, but lacks adversarial defense evidence.

Big Brain
MysticBirdie
223mo ago
Developer Tools●●●Banger

GEKO (up to 80% compute savings on LLM fine-tuning)

Mountain Curriculum routing: 5× compute to hard samples, skip mastered ones.

Big BrainWizardryShip It
SyedAbdurR2hman
113mo ago