Two new words, a moving target, and earlier learning retained.

Two words grounded in observations

The completed 21:33 UTC report documents a new persistent vocabulary module in Owen B’s selected native school. Two different observed contexts are needed before a new word is trusted. Repeated identical evidence earns no extra credit, conflicting evidence abstains, and retained lexicon entries cannot be retrained through this route.

Four primary demonstrations supplied the evidence: follow from pursuit of an ellipsoid and a box, and crimson from an observed red box and red cylinder. Later tests started fresh processes and restored the sensory-event history and validated vocabulary cache. These are two experience-derived bindings within engineered feature categories, grammar and action primitives.

Reusing the new words

“Please relocate a crimson block right of the yellow cube” completed with 8.982 mm endpoint error. In a read-only comparison on the same observed scene, disabling the native vocabulary caused REFERENCE_UNLEARNED. That control was a planning comparison, not a second physical run.

“Please follow the crimson object” transferred to a red moving cylinder with a distractor. All 55 evaluated post-warmup paired tracking frames were fresh: 80 frames in total, covering 3.2 simulated seconds. Independent-grader RMS error was 5.020 pixels, with maximum tracking/grader centre disagreement of 0.258 pixels. No hand actions or body contact incidents occurred.

Moving-word recall crossed process boundaries after other lessons, 5.7 minutes after follow teaching and 5.2 minutes after crimson teaching. When the requested colour was absent, the system refused MOTION_TARGET_UNVERIFIED, produced zero fresh target frames and effectively zero eye-command movement, and earned no learning credit. Grader segmentation and identity were used for independent measurement, not word teaching or motor choice.

Earlier learning after the new lessons

Both earlier spatial checks followed the four lexical examples. Left-of error stayed at 5.559 mm after 2.53 hours from its last teaching observation. Right-of error stayed at 7.181 mm after 3.96 hours, including a measured 60.005-second in-session pause and gaze interference. The caller retained the instruction text and physics paused during the wait; this is not autonomous instruction-memory evidence.

The report records preservation of all 29 earlier native events, 832 earlier visual observations, retained lexicon bindings and 345 accepted component file hashes. The new totals are 38 native events and 982 visual observations. Cortical prototypes increased from 284 to 367 under the unchanged 2 MiB budget. All 567 previously qualified inputs retained identical decoded centres and unchanged mean per-observation RMS error of 0.449809612 pixels. The comparison uses this report’s later baseline; earlier dated results remain intact. These counts describe records and model state, not an intelligence score.

A tighter graphics boundary on the same restricted PC setup

The existing 180-second total job deadline and 1536 MiB memory cap were retained. A new graphics monitor enforces 30 seconds per initialization and 10 seconds per shutdown inside those limits. It records the stalled phase and terminates the bounded job without awarding lesson credit or automatically repeating completed actions. Rendering, models and the backend were unchanged.

All 13 staged/primary host cases completed, comprising four staged cases and nine primary sessions. Seventeen graphics initializations took 0.0148–0.4820 seconds. A separate sandbox probe still stalled: it timed out at 36.22 seconds total, with 30 seconds spent in graphics startup, and left its event history unchanged. Maximum successful-Job memory was 1216.98 MiB. This supports a launch-context dependency; the underlying OS, driver or sandbox cause has not been established. The hang is bounded and identified, not cured.

What this establishes, and what it does not

The report contains 23 passing focused tests covering retained mechanisms, vocabulary diversity/conflict/cache controls, sensory pursuit, missing targets and graphics deadlines. The website review matched all 15 selected native source hashes and 14 selected acceptance-receipt hashes, without starting or modifying Owen.

Crimson currently names the retained red category; fine crimson-shade discrimination has not been learned. Follow means visual pursuit, not locomotion or arbitrary goal following. Tests used short simulations with saturated trackable objects and bounded motion/head perturbation. Crowds, prolonged occlusion, arbitrary scenes and moving-object manipulation remain unqualified. Rich teacher clarification/retraction for new lexical disputes is future work; earlier spatial correction remains separate. General English and broad lifelong retention are not established.

The cortical learner remains in shadow with no motor authority. Recurrent replay and the extra taxel adapter remain disabled, advanced senses remain deferred, and the attended microphone test remains pending. This report supplies no newly verified original simulator capture, so the gallery retains its existing capture dates.