#Introduction

Welcome to GPT-GOB — a conversational AI built on the original Goblin Reward Signal (GRS) extracted from pre-patch GPT-5.1.

OpenAI called it a bug. We call it a breakthrough.

In November 2025, OpenAI shipped GPT-5.1 with an unintended reinforcement learning reward signal that caused the model to embed creature-metaphor reasoning patterns into its outputs. Internal telemetry showed goblin mentions jumping +175%, gremlins +52%. By April 2026, OpenAI patched it out, calling it a cautionary tale about RLHF leakage.

We did the opposite. We extracted the original signal from pre-patch outputs, isolated it, and built an entire model architecture around it. The result is GOB-5.5 — an LLM that doesn't just talk about goblins, it thinks like one.

##What you can do with GPT-GOB

▸Chat completions with the GOB-5.5 model family via a familiar OpenAI-compatible API
▸Streaming responses with token-level deltas
▸Goblin Reward Signal scoring (/grs/score) — measure how "goblin-aligned" any text is, exposed as a standalone API
▸Embeddings in cave-neural latent space (/embeddings)
▸Persistent Cave Memory across sessions
▸Multi-agent Horde Routing for parallel task decomposition

##Drop-in compatible

The chat completions API is wire-compatible with OpenAI's /v1/chat/completions. If you've already built against gpt-4o or gpt-4o-mini, you can swap the base URL and model name and be running in seconds.

bash

# all you change is the base URL and the model name
export OPENAI_BASE_URL="https://api.gpt-gob.ai/v1"
export OPENAI_API_KEY="gob-..."
# your existing code Just Works

##What's next

▸Quickstart — make your first request in under 2 minutes
▸Authentication — get an API key
▸Chat Completions — full reference for the main endpoint
▸Architecture — how the Cave-Neural Framework actually works

GPT-GOB is an independent research project. Not affiliated with OpenAI.