I have nibbled at the edge of this idea and it fits in with ICM well. I am working on tools for it but I think the concept can help others. Quick test: define "AP", "POF", and "Gate". If you said Accounts Payable, Proof of Funds, and a logic gate — reasonable, and wrong. I work in parking. Here those mean Anti-Passback (stops a ticket getting passed back to a second car), Pay-on-Foot (the payment kiosk), and the barrier arm at a lane. Every industry has these. A new hire learns them in their first month. An LLM never does. It just picks the wrong meaning with full confidence and keeps moving. So instead of trying to make the model smarter, make documents smarter. A human maintains a glossary file. Term, what it means, a "not this" line (AP is NOT Accounts Payable), aliases people actually type, and a link to the deeper doc. An agent reads an incoming document (support ticket, email, whatever), finds which glossary terms actually appear in it, and prepends a header defining just those terms. Expand it to include routing tags. Dispatch, Billing, etc. Now every incoming support ticket has the information for the next agent to use. The glossary has who it goes to, or a link to the known troubleshooting process to write up an automated response. This can be done with a lighter model. This document is the router. The document contains the information to understand the document. It is also self-auditable. If no header, something is broken. If it's missing a routing tag because the agent failed to classify, it goes to a place for a human to look at it. Then the human can refine the glossary. I am calling it "Gloss". Would love to hear your thoughts, ideas, and questions.

I've been enjoying my ICM builds and systems. They are living and breathing and battle-tested. Now I've hit the point where I want to scale and expand them more than I practically can. The limiting factor? Me and my taste. My taste... well there's plenty of posts in here that will give you a pretty good indication of it. Human in the loop. Augment that human at every turn rather than cutting them out. My judgement where it matters. Guardrails and governance and stakes considered from the outset - as @Gabriel Azoulay put it, build the refusal and verification first and the rest will follow. (Yes, I've mangled it, it's not a quote, but the sentiment is there nonetheless). And in the middle of all of this, I've had a few ideas that have come in together. @Curtis Hays has agents in his system who sort requirements and building within his system. @Ari Evergreen has been inspiring me to think about where model vs model loops could play a piece in this picture to actually move things further along with my taste still driving all of the outcomes. And somehow this mashed up together into the following. @Bas Rosario taught us about how ICM was like baking a cake. And that moving from a workflow to a workspace was going from having a system that baked a cake to building a bakery. I can bake cakes. I have not one but two working bakeries. Now I need a construction crew. Not any construction crew - ones who know my bakeries and my taste and follow those to the letter. They don't suggest things that aren't compatible with my industrial grade ovens. They know what to do with my delicate equipment - nothing that is hand wash only is being shoved in the dishwasher by this crew! They understand that in my bakeries, my rules govern; and they honour and enforce that. Enter a new workspace where tools are made and bakeries are built. If you want to see them, glimpses of what these look like are already present in the week 8 competition builds - Chalky PRD (now 01 Chalky) and Architectural Thinking Partner (now 02 ATP).