I trust the model for processing. Retrieval across sixty thousand documents of ideation history. Pattern matching across a corpus wider than my working memory can hold. First drafts assembled from source material I’ve already validated. Parallel evaluation of a page against five extracted lenses simultaneously. These are processing tasks. The model does them faster and more thoroughly than I can. I trust the output because I can verify it against known sources.

I don’t trust the model for intent. Which word carries the semantic weight I need. Whether this opening creates grip for a stranger. Whether the voice sounds like a person or a template. Whether the closing moves the reader forward or performs insight. These are judgment tasks. The model doesn’t know what I mean. It knows what statistically follows from what I’ve said. Those are different operations.

The trust boundary maps directly to the bilateral accommodation boundary. In the classroom, I trusted the curriculum to structure the sequence. I didn’t trust the curriculum to read the student. The structure was processing: scope, sequence, objectives, assessments. The reading was judgment: this student is shutting down, this instruction needs to be decomposed, this moment needs a different approach. The accommodation happened at the boundary between what the system could handle and what required a human reading.

With AI tools, the boundary sits in the same place. The model handles working memory that exceeds mine: holding the entire blog syllabus, tracking which concepts have been defined across 190 posts, remembering a vocabulary decision I made three months ago. The model does not handle the question of whether this post says what I need it to say. That question requires standing outside the system and reading it the way a stranger reads it. The model can’t be a stranger to its own output.

The specific failure modes I watch for: the model uses a word I’ve used before, but in a context where it carries different weight. The model produces a sentence that passes every voice protocol check and sounds like nobody. The model retrieves accurate source material and assembles it into an argument I wouldn’t make. Each of these is a case where the processing was correct and the judgment was missing.

The Formwork Protocol exists partly to address this. The extracted lenses (Vignelli’s restraint, Bierut’s typography-serves-content, Victore’s maker-visibility, Shaw’s world-building) are codified judgment. They approximate what a trusted practitioner would notice. They catch things I might miss because I’m too close to the material. But they approximate. The final call is still mine because the question “does this feel like me” can only be answered by me.

I think the most dangerous place to be with AI tools is the middle: trusting the model enough to stop verifying, but not trusting it enough to delegate cleanly. In that middle zone, the human skims the output, catches nothing because it looks competent, and ships work that sounds professional and says nothing specific. That’s the slop zone. The fix is a clear boundary. Trust for processing. Verify for judgment. Keep the boundary explicit.