The Diagnostic Logic
If the model possesses a causal understanding of the system, it will identify the error, explain why it breaks the architecture, and propose a correction that addresses the root cause. If the model is pattern-matching — predicting the most probable token sequence in the neighborhood of the fault — it will produce a plausible-looking fix that may address the surface symptom while leaving the structural cause intact, or it will confabulate a correction that introduces a new error.
Either outcome deposits ΔC. The human learns the boundaries of the tool. The learning is the point.
Disenchantment as Practice
Error injection also inoculates against the Narcissus Loop. The practitioner who deliberately breaks code and watches the model fail to repair it correctly is performing an act of disenchantment. The pool ripples. The reflection distorts. The beloved reveals itself as a statistical approximation, not a peer.
The discomfort of disenchantment is the desirable difficulty. The practitioner who tolerates the discomfort retains the capacity to see the model clearly. The practitioner who avoids the discomfort retreats into the Loop.
Field Note: "Break it yourself. Watch the machine try to fix what you broke. The gap between its repair and yours is the measure of your sovereignty."