Testing disagreement without erasing the record.

A contradiction can remain evidence without becoming a trusted answer

An isolated artificial-word control tested the full sequence: a mistaken association, an observed semantic correction, a contradictory confirmation, explicit teacher withdrawal, and a fresh consistent confirmation. The contradictory observation stayed in the record with zero positive confirmation credit. The disputed word abstained until the conflict was resolved. Withdrawal was a separate no-action record; it did not erase the earlier observation or remove ordinary learning.

After an actual left-word lesson, the corrected artificial meaning transferred to a held-out object pair after 9.8 minutes and a measured 60.004-second pause. Endpoint error was 6.948 mm. The artificial word and fault control remained isolated and were never copied into primary memory. Owen did not independently decide which teacher statement was true: retraction was explicit, and a fresh consistent pair was still needed.

More varied wording, with measured changes in accuracy

A new primary “Put a red cube left of a blue cylinder” lesson added a third observed left-of object pairing. Transfer used the retained verb “Arrange”, changed articles and yellow/blue cylinders, with 6.215 mm error. This is variation within retained syntax and vocabulary, not general English or understanding of novel shapes.

A staging screen found a 2.240 mm shift in the learned left mean. An additional physical check on an earlier pair changed from 7.312 to 7.716 mm against the original reference. Earlier blocked-left error changed from 5.279 to 5.559 mm, while right error changed from 7.234 to 7.181 mm. The named checks remained within the declared 1 mm allowance. Performance was not identical, and that small change belongs in the record. Right recall followed the new left lesson and occurred 87.5 minutes after teaching.

A completed pause does not erase a timeout

The new conflict control completed its full 60-second pause. An older confirmed-control engineering copy also completed a 60-second pause with gaze interference, 73.6 minutes after its meaning observation, using the candidate modules before installation. These are new successful tests; the previously reported failed attempt remains a failure.

A profiled diagnostic reproduction still hit the existing 180-second cap during a recall pause. It produced a timeout/resource receipt and appended no action event. An earlier setup attempt failed before any lesson because copied RGB evidence was missing; the setup was repaired. Graphics-context creation in the diagnostic took about 80 seconds, and startup/shutdown variability remains unresolved. No deadline or memory limit was increased, and no rendering backend was replaced. The report records 13 successful exits—12 physical completions and one no-action withdrawal—alongside one timeout and one setup failure. Maximum successful-Job peak memory was 1036.30 MiB under the 1536 MiB cap.

Growth with earlier predictions retained

Primary visual history grew from 809 to 832 observations and from 271 to 284 prototypes. The retention guard admitted 23 updates and refused 58. All 559 previously qualified predictions were retained, with unchanged mean earlier-pair RMSE of 0.449034013 pixels. The model itself changed; preservation of those predictions is not an unchanged-model claim or a guarantee for unseen inputs. Cortical learning remains shadow-only, with no motor authority.

Four actual-bank proposal comparisons produced identical old/new decisions and models under the revised copying strategy. Focused equivalence checks covered all 809 earlier inputs. The report contains 17 passing focused tests, but whole-run timing differences cannot be credited to copying alone because startup variability and execution permissions differed.

Both eyes still follow one paired target. The report independently triangulated 225 newly retained paired frames; it introduced no replacement eye controller. All 25 earlier native events and 809 earlier visual pairs were preserved. The report does not specify the final native-event total, so no new total is inferred here.

What remains open

The caller retains the instruction string during pauses and physics is paused while sleeping. The results do not establish autonomous intention memory, general English, broad resistance to forgetting, or unrestricted depth accuracy. The attended microphone check remains pending, and advanced senses and the later structural-growth stages remain deferred. This report supplies the bounded test evidence that was still pending in the earlier contradictory-feedback implementation entry.