Field report 002 · Open
The Refusal That Cannot Be Cached
Why remembering your taste is different from holding your standard.
On May 29, I reopened the conversations behind four essays I considered among my best. I expected to find my sentences inside them. Claude had written essentially every one.
The discovery killed a clean lesson I had just learned. After watching an automated writing pipeline produce polished work I would not publish, I had concluded that models should research and challenge while the human wrote every sentence. The primary record showed that sentence authorship had never been the source of quality.
My contribution was smaller by volume and larger by consequence. I brought an odd seed. The model opened the territory. I chose the center and locked the thesis. It drafted, then drifted toward a version that was intelligent, defensible, and wrong for the piece. A short refusal from me forced the work to be derived again.
That refusal was the collaboration's highest-leverage event. It was also the event my systems kept trying to remove.
The rule left by a correction
A correction leaves useful residue. If I reject a literal translation of an engineering term, the system can remember the accepted register. If I reject an unsupported technical claim, it can require a source next time. Stable definitions, citation standards, glossary choices, and repeated factual errors should become memory.
This worked in practice. Six Korean translation corrections made on one morning became a reusable language rule. Later translations improved without requiring the same six interventions. I wrote about that experience in Writing for the Next Model.
It was tempting to extend the result. If a system could infer one language rule, perhaps it could infer the whole standard. Every rejection could become a rubric item. Every good piece could become a positive example. With enough history, the model would meet me near the answer instead of waiting to be corrected.
The repository ran that experiment. Twenty-six audit scripts encoded failures I had already seen. Reviewers scored some pipeline output at more than 90 percent voice match. The prose was clean, sourced, and plausibly mine. I removed it from the public surface.
The memory had learned the visible shape of prior judgments. It could avoid yesterday's obvious error. The decisive drift in a new piece was usually different. One draft became too academic. Another optimized for executive usefulness. Another discovered a respectable regional angle that displaced the reason I had started. Each move looked competent in isolation. No static rule could know which competence was the betrayal this time.
Personalization changes the person too
Personalization is usually described as the system moving toward the user. The interaction runs in both directions. Once people know that their behavior trains the assistant, the behavior can become an input ritual for the machine.
A 2026 study tested AI writing support personalized with a reader's highlights. The personalized condition encouraged more highlighting, but participants began highlighting to feed the AI rather than to make sense of the text. They relied more heavily on the assistant and reported less autonomy, ownership, and self-credit. The sample was small at 46 participants, so the exact effect needs replication. The mechanism is already legible. The AI Personalization Paradox documents the study.
A person who knows the archive is watching can begin performing consistency for it. New material is judged through a profile built from old choices. Surprising the system becomes evidence of error, even when surprise is the sign that the person has moved.
Recent agent research reaches a similar boundary from the engineering side. The PAHF framework found that persistent personalization adapted better when memory was paired with both pre-action clarification and post-action correction. Static history was insufficient when preferences changed. PAHF treats personalization as a continuing interaction rather than a profile learned once.
Memory retrieval itself has become a trust problem. A June 2026 evaluation found that a semantically relevant memory can still come from the wrong domain, carry a stale assumption, or steer later actions inappropriately. The authors describe memory as a durable control channel and place an admission gate between stored history and the model. Beyond Similarity supplies the evidence.
The frozen taste problem
A taste archive is a prior over the person who created it. The better the archive becomes, the more persuasive that prior feels. This is useful for catching repetition and dangerous for work whose value depends on the author changing.
My strongest essays do not merely express stable preferences. They contain a position I gave up. One reversed a claim I had already published. Another began with a system I thought was about engineering and ended with fear about my own place in the stack. A memory optimized to preserve voice can easily preserve the person who held the old position.
The live snapback performs a different job. It compares this draft, this seed, and this moment of judgment. Past refusals can inform the comparison. They cannot own it. The author must remain able to violate every stored preference without first proving that the archive is wrong.
This relocates the value of memory. The system should retrieve earlier decisions, show the context that produced them, and expose when a current move conflicts. It should preserve the before and after of a correction rather than distill every correction into law. The conflict returns to the person as a question.
A collaboration that can still surprise itself
The working division is now clearer. The model carries breadth, drafts, prior evidence, and the record of previous refusals. It can warn that a choice repeats an old failure. It can also argue that an old rule no longer fits. The author brings the seed, locks what the work is trying to discover, and remains responsible for the correction that does not yet exist in memory.
A useful memory system should therefore reduce repeated corrections without reducing new ones. If the same factual or stylistic rejection recurs, memory failed to help. If the author stops surprising the memory, personalization may be narrowing the work. Both signals matter.
This is also why the private Library remains in cold standby. It stores claims and prior positions well. This report did not need it. Repository archaeology and direct reading were enough. A future piece may need to recover a changing belief across years, and that task could pull a narrow revision layer back into use. The artifact decides.
The next time a draft drifts, the record will keep the before, the correction, and the after, without pretending that any of them can decide the next case.