An LLM "Glass Box" Option (Patreon)

Video - https://youtu.be/3luBwgEYQts

This one needs loads more testing, but it seems like a reasonable first step, so here goes!

This is basically an "OS" for use on LLMs which acts like a compiler. You can program using it's Structured Language by issuing commands in English. Yes, it helps if you can code, but no - you don't need to know how to.

The Problem

The "Black Box" problem is super difficult solve. I know I can't do it, but what I can do instead is make my own "Glass Box" that filters what is coming out of the Black Box. That way, I can view the output through whatever lens I like.

A Complementary Step Forward?

Cool! So we haven't "solved" the Black Box problem, as we already know that's super difficult, but what we can now do... is make our own "glass boxes"! Another way of looking at it, so to speak.

These Protocols Are Just Performative!

Could be... however... does the output make sense? Is it what you were looking for? Did it help?

So it depends on what I think?

Exactly! I'm putting a question to you, rather than making a statement.

How do I test then?

Super easy... though powerful LLMs handle the scripts better, even things like Gemma 24B seem to handle them OK. This means you can test locally with things like ollama, or using Claude, ChatGPT, etc.

Just paste the text file in, and the LLM should do the rest.

Any other information?

Some LLMs may show multiple questions at a time, some may start one at a time. You should be able to ask the LLM to do either.

It's designed so that the questions should actually end - but it's up to you (or the LLM context window ;) when that is.

Additionally, the Identity framework itself can be used to make other "glass boxes". It's just a "base" protocol that works by itself, or can be used to generate others.

WHAT!? It's a self-replicating machine!

No. You have to tell the identity framework what sort of new SQF protocol you want to make, then it will use the base rules to make it for you. You are the one in control.

The "creative" module is an example one I made to show you how it differs (but maintains a "strong identity" with the first).

These are only in English though :(

I haven't tested, but they should operate in any language. The main thing being that so long as the core logic is preserved, it is language agnostic.

What is different here to normal LLM programs?

It's like a mix of formal and informal things. Concepts and rules together.

Conversational Inversion of Control: The framework, not the user, orchestrates the conversational flow by offering structured choices, ensuring the user's input consistently drives progression within defined pathways.

Explicit Role Separation (User-AI): Clearly defines distinct roles: User leads by choosing and directing; AI supports by generating scenarios and responding, ensuring unambiguous control and predictability.

Dynamic Self-Correction Mechanism: Incorporates an active system that monitors conversational coherence and alignment with user goals. If deviations occur, it prompts for user guidance to collaboratively return to a desired path.

Modular System: Features a core design capable of adapting to diverse themes and interaction styles (e.g., fantasy, philosophical, mystery) via configurable content modules, without altering fundamental conversational logic.

Predictability & Safety-Centric Design: Emphasises user control through strict turn-taking and choice-based progression, intentionally designed to prevent autonomous AI initiative and ensure a safe, predictable interaction experience.

LLM Capability Management: Strategically leverages the underlying Large Language Model's generative power while employing structural constraints (options, turn-taking) to mitigate potential inconsistencies and maintain desired output.

16-Jun-2025 - MORE UPDATES!!

Collaboration-OS-Stable.txt should have most of the bugs fixed to prevent LLMs from "losing it". Framework editor should handle things better now too.

LLMs will still go wrong, so just say "Error Detected" and let it know where/why/how it went wrong. You could also modify the framework to handle your own error types, if you like. The framework is designed to modify the framework.

The framework is it's own instructions, so you can just ask it what it can do. For example:

Framework Refinement is where you can make changes to the framework:

You can also just go back to the start:

ALL THE OPTIONS

All the options are in there. Sort of. It's your choice ;) You can switch identities by asking:

Don't like a mystery? No problem...

Confused by the scene? Well. It is YOUR CHOICE after all...

Scenario text too long?

None of the above is also an option...

Etc... Yes. Everything is your choice, make of it what you will!

Works great in Silly Tavern too :)

Attachment Information

Collaboration-OS-v1.txt is the one to start with, but the older versions are attached for those who are interested.

Previous iterations:

The Collaboration-Identity framework is a protocol for an AI to follow which is designed to guide the user through an identity related question, such as "Is mayonnaise a soup?", "What makes a chair a chair?", that sort of thing.

The Collaboration-Creative framework is a similar protocol, but more related to artistic endeavours.

The Collaboration-Formal version has as much of the philosophy stripped out without breaking SQF

The Collaboration-OS version makes it more obvious you can change the code, though how operational things remain afterwards... still testing!

Have a go and let me know what you think!