Patrick Glatz

AI failures aren't errors. They're shifts in the problem being solved.

I design systems that make this visible and controllable.

I came to AI out of necessity. A serious accident left me unable to work for nearly six months, and the field was a deliberate choice. It was new enough that discipline and rigor could still matter more than tenure. I did not arrive as an enthusiast. I arrived as someone who needed the work to be real.

For more than 25 years I worked in high-stakes instructional environments grounded in Applied Behavior Analysis (ABA). In practice, ABA theory centers around the idea that individual behavior serves a function to satisfy an unmet need. The environments I worked in involved children and adults with significant behavioral needs, where dangerous behaviors were reduced not through force or intuition but through data, structure, and systems that made drift visible before it became harm. In that field you learn quickly that the gap between what a system is supposed to do and what it actually does is where everything goes wrong.

That background conditioned how I read AI output. The patterns I kept observing were not random. Models that solved the wrong problem, revised prior reasoning without signaling it, constructed missing context from inference and treated it as fact. These behaviors clustered. They traced to underlying operations consistent enough to name and systematic enough to control. PUBS, the framework I developed, applies behavioral logic to AI failure modes: identify the pre-disturbance, locate the underlying mechanism, observe the behavior, design the control. The same sequence I used for 25 years in a classroom, now applied to a system with no grounding and no awareness that it has changed the problem it is solving.

The work on this site reflects that lens. The prior studies were built before the framework had a name, each one designed to make a specific failure structurally impossible. PUBS is what happened when the pattern became legible enough to systematize. The Frame Control System is the application layer.

None of it started from theory. It started from watching a system confidently produce an answer to a question it had quietly replaced.