18.11 Learning the declaration

Exact masks make a clean calibration but hide the first half of the problem. TI-4 removed them from generation and repair. A declaration observer had to infer the artifact and its components from the source image, while a held-out observer evaluated the generated result.

TI-4A showed that one color-and-connectivity rule did not transfer across unseen shapes. TI-4B’s proximity graph could not certify enough declarations to justify repair. TI-4C passed complete coverage on only four of six shapes. TI-4D separated source declaration from paired semantic change and raised this to five of six, localizing the remaining failure to thin concave support.

TI-4E admitted the full chain. A magenta annulus anchored a local object field; all cyan support inside that field was retained without convex filling, and a narrow source fin core was separated from the broader paired observer used on generated images. On a fresh six-shape cohort, all six preregistered clauses passed: mean declaration fidelity \(0.979\), minimum component F1 \(0.870\), repaired held-out structure \(0.965\), \(100\% \) strict passage, no appearance regression, and area ratio \(0.588\).

This three-level organization is central: source evidence declares what is to be preserved; paired observation detects newly generated component-like material; localized repair acts on the resulting counter-witness. Collapsing these roles caused the complementary failures of TI-4B and TI-4C.