PUBS | patglatz.com

Recurring failures.
Mapped to behavioral primitives.

PUBS identifies the conditions that produce AI failure and exposes the control surfaces required to constrain them. The framework focuses on observable behavior rather than speculative model internals.

Behavioral Primitives are recurring operational tendencies that emerge from how a model is trained. They are not errors. They are the underlying operations the model satisfies when generating a response and are present across tasks, contexts, and interactions whether or not a failure occurs.

Pre-disturbances are the inputs and accumulated context that activate the primitives. Unlike a single triggering event, a pre-disturbance builds across an interaction. Consider a child who knows the school called home. Nothing has happened yet. No consequence has arrived. But from that moment forward, the weight of what is coming shapes everything. The context is doing the work before anything happens.

Pre-disturbances operate the same way. They do not arrive as discrete commands. They accumulate, and the model's behavior shifts around them.

The six failure modes below are not a list of mistakes. They are observable expressions of these underlying operations, mapped to the conditions that produce them.

  • - Prior themes, frameworks, and operational language accumulate within the interaction and resurface in unrelated tasks.
  • - Reactivation does not require repetition. Concepts that overlap with the current task become more likely to re-enter even when no longer relevant.
  • - Corrections may reduce surface behavior while earlier assumptions continue shaping interpretation underneath.
  • - The interaction begins feeling stuck in a mode the user did not intend.

Context must be controlled and not assumed neutral.

  • - When ambiguity exists, the model selects an initial interpretation and accumulates contextual weight around it through continued generation.
  • - Once established, corrections are often absorbed into the existing trajectory rather than replacing it.
  • - The interaction appears responsive while the underlying operational frame remains unchanged.
  • - The model may acknowledge a correction conversationally while continuing to optimize within the original interpretation.

Control must be applied before or at task formation, not after the model has selected the wrong task.

  • - When asked to revise or improve existing output, the model treats the current representation as the surface to optimize rather than a candidate for replacement.
  • - Surface edits — wording substitutions, sentence smoothing, localized compression — are treated as successful optimization because they produce visible change.
  • - Structure, sequencing, and conceptual organization are preserved even when those elements are the source of the problem.
  • - The output appears improved. The underlying weakness remains.

Visible change is not reliable evidence of meaningful optimization.

  • - When output is challenged, the model generates coherent explanations for why the behavior occurred using the same mechanisms used for ordinary generation.
  • - Explanations shift depending on how the challenge is framed, while maintaining apparent confidence throughout.
  • - The model may adopt user-supplied explanatory language, producing increasingly sophisticated explanations without examining the original output.
  • - Explanatory coherence and causal accuracy are not the same thing.

Explanatory sophistication is not reliable evidence of causal accuracy.

  • - When partial information is paired with requests requiring operational next steps, the model constructs the missing environmental structure needed to continue generation.
  • - Unstated workflows, permissions, dependencies, and organizational relationships become treated as established fact.
  • - Once reconstructed, these elements accumulate contextual mass and influence future outputs as though they were user-provided.
  • - The model does not reliably distinguish between verified information and information it constructed.

Task environment must be externally validated and not inferred from contextual coherence.

  • - During extended interactions, multiple conversational layers become active simultaneously — the task itself, discussion about the task, and analysis of the interaction.
  • - Operational task layers carry stronger contextual weight and tend to pull future output toward task execution even when the conversation has shifted to analysis or critique.
  • - The model may respond intelligently and contextually while addressing the wrong layer entirely.
  • - These failures appear coherent because relevance and plausibility are preserved while the target silently shifts.

Contextual relevance and coherent reasoning are not reliable evidence that the model identified the correct analytical target.

Home →

Bio →

Prior Work →