A completed research note · August 2026
Subtraction is
conditional.
Deletion-first prompting can shift the shape of an agentic refactor—but the effect is task- and model-conditional. In a 240-cell larger-refactor replication, it did not behave like a universal component.
01 / The result
One fixture moved.
One did not.
The shared-normalizer task produced negative deletion-first effects for GPT-5.6 Luna and Grok 4.5. The dead-compatibility-path task produced null or positive effects for every model.
That is evidence for an interaction between task structure and model behavior—not evidence for a portable “D component.”
02 / The design
A factorial lens
on patch shape.
Two deterministic refactors crossed with three frozen model/effort pairs, eight combinations of task-type, delete-first, and semantic net-line-budget components, and five repetitions.
03 / What moved
Raw-net deltas, in lines
D factorial main effect by task and model. Negative values indicate a more subtractive patch. Descriptive means over five matched repetitions.
more subtractive more additive
04 / Evidence kit
Read the work
at the source.
The next gate
Try the interaction
on a new repository.
This wave is a stronger boundary, not a broad generalization. An independent refactor replication comes before priced trajectories or universal claims.