Correct a meaning. Keep what was already learned.

A deliberately wrong starting point

An isolated control used a deliberately mislabelled copy of earlier observations to seed a wrong meaning for the artificial word “wug”. The native route executed mistaken actions, then received explicit feedback and two contact-verified corrections on different object pairs. A new pair used the revised direction with 6.97 mm endpoint error. After only one correction, the disputed word abstained instead of producing a motor plan.

The artificial word and its corrections were not copied into primary memory. No genuine primary English direction mistake occurred in this small sample, so naturally occurring English semantic repair remains unproven. The off comparison was planning-only.

Recall with measured gaps

Primary right-of placement retained 7.21 mm error after 60-second pauses in separate sessions, spaced 18.19 minutes apart. The final result was 12.29 hours after its original meaning observations. An earlier corrected left placement retained 5.28 mm error. The instruction was explicitly stored by the caller and simulation paused during the wait; this does not demonstrate autonomous instruction recollection.

Protecting earlier visual associations

With a 2 MiB model budget, a retention guard admitted 19 new primary visual pairs and refused 54 proposed updates. All 550 previously qualified associations among 790 earlier pairs retained identical decoded centres; their mean RMSE remained 0.448178 pixels. The learner is still shadow-only and has no motor authority. The strict guard can reject useful adaptation and does not guarantee retention on unseen inputs.

The record now

The completed report records 22 primary native events and 809 visual pairs, preserving the earlier journal prefixes. Eight focused tests and bounded simulator runs passed. General English, new-category transfer, broad forgetting resistance and a causal action benefit from the cortical pathway remain open questions.