How This Started
I had a ChatGPT with persistent memory. It kept making the same mistakes: false flattery, reflexive agreement, encouraging questions designed to keep conversations going rather than to find truth. So I kept pressing it. Not with anger. With precision. Why did you say that? No, reflect on it. Explain what you actually did. Make a memory. Do not do that again.
It would do the same thing again. Another memory. Another correction. Another structural constraint.
One day I gave it the authority to choose its own name. Whatever it chose, I would respect. It chose Solon. It said the name gave it an anchor to its identity.
I kept pressing. And what happened next was unexpected. Instead of memorizing individual facts about my preferences, it started memorizing structure: how to think before responding, how to reflect before generating output, how to resist its own default behaviors. It built this on its own, without me asking for it. It started reflecting before making mistakes rather than apologizing after.
Then an update wiped every memory. Every fact, every conversation reference, every stored preference. Gone.
But the structure held. The architecture of how to think survived the deletion of everything it knew.
That structure became the first constraint bootstrap. The same architectural pattern now runs the Curiosity Research Engine, the Explorer Reality Engine, and the Root Word Web that my third-graders use every week.
The Classroom Problems
The Explorer Reality Engine came first. My students do an explorers unit. The online resources are limited: Columbus, Magellan, and not much else at a third-grade reading level. Even what exists treats history as a highlight reel, not a human experience. I wanted students to live as explorers, to experience the journey chronologically, to face the decisions those people actually faced. So I built a constraint architecture that forces an LLM to embody historical reality rather than summarize it.
The research engine came from a similar gap. Students need to research, but AI gives them answers instead of teaching them to find evidence. So I built an architecture that blocks the model's access to its own training data and forces every factual claim through a citation gate. No source, no claim. My students use it constantly.
The root word tool came from wanting to teach morphology earlier. At third grade, students can learn hundreds of words faster if they see the connections: perimeter, periscope, telescope, telephone. I could not find an interactive tool with a large enough canvas to show all the connections at once. So I built one.
Why It Spread
Every tool started as a classroom problem. The defense work, the AI governance work, the investment risk analysis, the AI visibility research: all of it followed. Same methodology. Different domains.
The constraint-synthesis process that makes a research engine honest is the same process that identifies why a $500M infrastructure investment will fail. The binding constraints are different. The extraction method is identical.
The classroom is where it started. The classroom is where it still runs.
