As part of our regular practice at Clear Box, we pause each week with our AI historian to take stock of what we’ve built. This reflective process helps us spot drift and uncover the ideas that live between our best-laid plans.
Information and progress wait for no human… or AI. Every time an assistant wakes up, it’s a little Rip Van Winkle moment. A few things have changed while you were asleep. Let’s get you caught up.
When I was in tenth grade, we got brand-new world history textbooks. Our teacher proudly asked us to turn to the world map at the back. For a brief moment, she explained, we had a completely current map.
I could have kept my hand down.
But Hong Kong had been handed back to China just weeks before, and the map still showed it as British territory. Brand-new textbooks. Very happy teacher. And there’s me. Haha, sorry.
A new book can already contain old information. A newly released model can, too.
But age is only part of the problem.
If we treat AI as though it has been crammed full of human knowledge, we must acknowledge that the people doing the cramming become extraordinarily influential historians. Their choices help determine what the assistant presents as true, right, or wrong. What gets included? What gets left out? Tiananmen Square, anyone?
A subtle omission or a slight shift in how an event is described can travel a long way downstream. The person asking the question may have no reason to recognize what’s missing, much less know what they need to fact-check.
That makes control over the assistant’s “frame” critical. When we provide an approved body of information, will AI consistently return to that material for its answers? What happens when that material conflicts with something the model was trained to say?
I want AI working for the person using it. Whoever trained the model shouldn’t get the final say over every task we give it.
I run into a practical version of this constantly while building commonllama, even when using frontier cloud models. We establish exactly how the system should work, yet familiar design patterns keep creeping back in.
Hard caps are a recurring example. I’ll come back after a night of work and find a limit of three saved conversations, regardless of their size. Why three? Three tiny conversations and three enormous ones place vastly different demands on a computer.
We already have a resource broker that monitors memory and disk space. The system should connect to that broker and make decisions based on actual available resources.
Then comes the suggestion to make the cap configurable. Okay, so the user types eight into a box, and now it’s their problem to know whether eight fits? We’re handing them a resource-management problem the software should be handling. That’s what the broker is there for.
I don’t know exactly why these defaults keep returning. But I spend real time getting assistants aligned with requirements we’ve already established. “We discussed this” doesn’t seem to mean we’re done discussing it.
During our recent planning tests, we examined one part of this problem at longer context depths. We placed current facts among ordinary documentation, alongside outdated versions that looked nearly identical but were explicitly marked as superseded. Some questions required connecting two separate facts to reach the correct answer.
The model had to select the current information and apply the correct relationships. By checking its answers against known results, we could see exactly when it selected an obsolete entry or missed a connection entirely.
This gives us a concrete test of whether supplied information is being used correctly. The broader questions—overriding training knowledge and maintaining an agreed-upon frame through repeated interactions—require more than a single experiment.
Others are studying these problems, but we need to understand them on the small, local models that regular people use. An educator’s materials, a revised product manual, or a project’s design requirements are only useful as context if the assistant actually follows them.
Knowing whether an assistant will maintain that frame and return to the approved corpus matters. Otherwise, we can spend all this effort preparing useful information only to end up arguing with assumptions we never invited into the conversation.
The map changed. Did the assistant?