Analogical Consistency Intervention

Current Conference-Paper State

Source: path:paper/conference-paper/figures/wrap-analogical-consistency.tex, input from path:paper/conference-paper/sections/04-experiments.tex.

Label: fig:analogical-consistency.

Asset: path:paper/shared/figures/analogical-consistency/figure.pdf.

Caption headline: Analogical consistency tests cross-input strategy identity.

The wrapfigure depicts the analogical transfer intervention: infer or sample a latent $z$ from a source input $X$, inject the same $z$ when generating from a related target input $X'$, and check whether generated outputs $Y$ and $Y'$ use the same strategy.

Writing Role

This figure defines the evaluation logic behind Analogical Consistency and supports the finding that recovered latents have cross-input semantic consistency, not merely within-input strategy separation.

Stale Or Open Items

The caption currently uses \(z\) style inline math. The knowledge-base convention is $z$, but this is a LaTeX source file, so the distinction is not an issue there. The nearby displayed metric in path:paper/conference-paper/sections/04-experiments.tex still uses \operatorname{AnalogicalAgreement} and should be renamed to \operatorname{AnalogicalConsistency} when paper text is updated.