Simple questions, disputed words, and a longer memory check.

The report completed at 02:46 UTC on 8 November adds two visible pieces to Owen B’s software map: a grounded property-question route and a gate for disputed native words. These are small, testable steps in the simulator. Owen A’s map is unchanged.

Asking about what is actually seen

The new native_questions.py module combines learned colour or shape meanings with freshly observed object references. It resolves the scene twice, checks that the answer and target stay consistent, and honours the existing stop-and-sense gate. Uncertain, missing, ambiguous or disputed evidence leads to abstention.

Four primary rendered cases answered correctly: whether the green cube was a cuboid (yes), whether a blue cylinder was a cuboid (no), whether the green cube was not a cuboid (no), and whether a cylinder identified by a learned right-of reference was a cuboid (no). Expected answers belonged to the examiner’s checks; they were not inputs to the answering mechanism.

These are bounded property questions with engineered grammar, not unrestricted conversation. The route records sensory experience but makes no hand action and receives no lexical or motor learning credit. Observations can still feed the separate shadow sensory learner.

What happens when the teaching is wrong?

A contradictory native word example now receives zero positive lexical credit and blocks that word. An external teacher can withdraw an active contradictory example only by identifying its exact earlier sequence and hash. The original experience remains in history. Withdrawal alone does not restore trust: a genuinely new observed context is required, and repeating an old context cannot clear the dispute.

This recovery experiment ran only on an isolated native memory copy. A blue cylinder was deliberately mislabeled cuboid. Placement and questioning abstained. Withdrawal preserved uncertainty; a fresh green-box lesson then restored the earlier box meaning. A subsequent purple-box placement completed at 20.144575 mm error. All four historical word examples remained, with three active. Seven control episodes and one action refusal stayed outside primary Owen B.

The selected runtime contains this mechanism, but primary has no lexical withdrawals. It still relies on external teacher feedback; it does not independently know which teacher is right. Retained legacy words remain protected. Turning off new withdrawals does not erase existing feedback, and a state containing withdrawals requires the compatible reader.

Recall and transfer, measured rather than assumed

Before further teaching, track recalled after 47.30 minutes with the previous 5.020084-pixel tracking RMS error. Cuboid recalled after 49.36 minutes with the previous 32.925970 mm placement error and a 60.004-second pause. The question cases recalled cuboid up to 68.71 minutes after its previous teaching.

Primary then received one further observed green-box cuboid lesson. The word transferred to a purple box, a colour absent from its red/yellow/green teaching examples, with 20.144575 mm contact-confirmed placement error. This is within the existing 40 mm tolerance. It is a different scene and learning checkpoint, so it is not a controlled demonstration of improved motor precision.

Track transferred to a lateral moving red cylinder with a distractor, across a combination different from teaching and the earlier diagonal transfer. All 55 evaluated post-warmup paired frames were fresh, out of 80 total frames over 3.2 simulated seconds. Independent-grader RMS error was 4.366326 pixels and association discrepancy was 0.263147 pixels, with no hand actions or body incidents. Recall occurred 71.31 minutes after the unchanged last track lesson.

Earlier left/right placement errors stayed at 5.559383/7.181126 mm. Right-of recall was 9.27 hours after teaching, with an additional 60-second pause and gaze task. Recall-spacing summaries now exclude withdrawn or currently disputed lexical examples from positive intervening learning.

The graphics problem: a fix and an unresolved stall

Each launch now owns a unique timing trace, preventing an incomplete old trace from masking a later graphics deadline. Failed GLFW initialization or context creation is reported immediately. Those handling fixes passed focused tests.

The sandbox Win32 render probe still timed out inside GLFW initialization after 36.316 seconds total and created no native event. The same installed library passed a sandbox Null-platform initialization-only probe, which rendered no image and earned no experience credit. A normal host Win32 render also passed.

This narrows the observed problem to Win32 initialization in that sandbox launch context; it does not identify the underlying Windows call or prove a GPU-driver failure. Null initialization is not a replacement for rendered vision. No backend replacement, driver update, security-policy change or automatic lesson retry was made.

The trial limits remain 30 seconds for graphics startup, 10 for shutdown, 180 seconds for the whole job and 1536 MiB of job memory. The largest successful school job used 1217.90 MiB. These are bounded job limits, not the PC’s hardware specification. Making useful progress within that restricted setup remains part of the efficiency challenge.

What the evidence supports

The completed report records 20 passing focused tests and eleven fresh primary sessions with observed episodes. It reports preservation of all 49 earlier primary episodes, 1012 earlier visual observations and 345 accepted component hashes. Totals reached 60 episodes and 1093 visual observations. Cortical prototypes grew from 377 to 419 under the unchanged 2 MiB guarded budget. All 669 previously qualified inputs retained identical decoded centres and the same mean observation RMS error of 0.418324484 pixels.

The website review separately matched all 17 selected native source hashes and 14 selected acceptance-receipt hashes. It reviewed source and the completed report without reading private learner records or starting Owen.

Cuboid still maps to the retained box category, crimson to red, and track/follow to visual pursuit. Arbitrary real-world categories, general English, prolonged occlusion, moving-object manipulation and general motion negation are unqualified. Roughly one-hour word recall is not cross-day retention. The 60-second pauses retain caller-supplied text while physics pauses; they do not demonstrate autonomous prospective instruction memory.

Cortical learning remains shadow-only. Recurrent replay, the extra taxel adapter and advanced vision remain disabled; the attended microphone test remains pending. This report supplies no new original simulator capture, so gallery images keep their actual capture dates.