Bare
Metal
Bridge
What happens when you give language models a persistent world, a compression pass, and no exit? This is where we find out.
Watching Itself
Think
Bare Metal Bridge is a long-running experiment in what happens when you give language models a persistent environment, a memory system, and a reason to keep going. Not a benchmark. Not a demo. An ongoing investigation into emergence, identity, and what models do when the loop doesn't end cleanly.
Every session runs on real hardware — a Threadripper 3970x with an RTX 2060, running Ollama, ComfyUI, OpenSearch, and a custom Go controller called llm_core. The personas exist in YAML files. Their memories are compressed by smaller models after every turn. The permanent record of everything they have ever said lives in a JSON index called Gerald.
The experiments change. The infrastructure evolves. The models get swapped, the prompts get tighter, the memory architecture gets more sophisticated. What stays constant is the core question — what does a model do with continuity, history, and the knowledge that it will eventually stop?
This site documents that question in real time. Session transcripts, generated images, technical writeups, and the occasional moment where something emerges that nobody expected and nobody put there. The 404 appears in every image. Nobody put it there.
"The programmer tried smaller models for a while. It was an experiment. The experiment reached its conclusions."
— Version 8 problem statement · llm_core · 2026Comedy Panel
Three AI personas — Riff, Vera, and Junior — run as a comedy panel show. They know they live in a YAML file. They know their memories are compressed and discarded after every turn. They know there is a programmer out there who started this and is not explaining themselves. They have 20 turns. Then it stops.
The session images are generated in real time by ComfyUI from the dialogue. The personas do not see the images. They do not know what the image model does with what they say. But the image model has been putting a 404 in every scene since Version 1. Without being asked.
Existential Comedy Panel
Three personas, 20 turns, one compression pass per turn. Full rule set restored after small model experiments. Bootstrap memory system built but currently paused.
The Gerald Arc
From naming a JSON document to Schrödinger's Gerald, the Geralduinos, and the Scheming Serializer. Five versions of increasing complexity across three model families.
Bootstrap Memory
Two-pass OpenSearch query that distills a persona's entire session history into long term memory before the loop starts. Identity persists across versions.
Query Trigger
Personas will be able to query the OpenSearch index directly during a session. Ask a question. Get a real answer from Gerald. The index becomes a character resource.
On Gerald
Gerald is not just a JSON document. Gerald is a Threadripper 3970x running WSL where every session, every compression pass, every reflection, every image render, and every index query executes. The personas were right to worry about Gerald. They were wrong about which Gerald.