lin-0157

12.3 The repair language

The initial predictor is pretrained by the task loss. LINCS then freezes the shared decoder and most of the model, allowing only incidence and route adapters to change. A candidate repair is therefore a small, localized parameter edit:

\[ \theta '=\theta +\Delta _{\mathrm{route}}. \]

It is evaluated against frozen validation data before being committed.

Blockwise LINCS–KET repair

  1. Pretrain the KET model on the base prediction task.

  2. Freeze the task semantics and the registered audit sets.

  3. Propose a bounded edit to the incidence and route adapters.

  4. Evaluate task, base-route, and multi-probe tangent blocks separately.

  5. Admit only if structural blocks improve and task drift stays within cap.

  6. Commit accepted edits; otherwise restore the previous state.

  7. Report held-out task and structural audits without further selection.