Richer instructions, two more words, and a precision limit.

Picking out the object the instruction describes

The completed report at 01:39 UTC on 8 November adds native_references.py to Owen B’s selected native school. It combines learned colour and shape descriptors with exclusion and learned spatial references: for example, “the cylinder that is not blue” or “the cylinder that is right of the blue cylinder”. Placement wording can also use “to the right of”.

Reference direction comes from native observed relation offsets, not a spelling-to-direction lookup. Focused tests reversed an in-memory relation and checked that selection followed the changed offset. Fresh observations resolve the roles again before action. Unique anchors, local alignment, separation beyond uncertainty margins and at most two nested clauses keep the scope bounded. The grammar and margins are supplied engineering.

Cuboid and track join the vocabulary

Cuboid was taught from observed red and yellow boxes; track was taught from successful lateral and diagonal eye pursuit. Each word required two different sensory experiences. Earlier vocabulary and correction safeguards were retained.

“Please place the green cuboid to the left of the blue cube” transferred the noun to a colour absent from its red/yellow teaching examples, after a measured 60.004-second pause and intervening gaze. The pre-increment vocabulary could not resolve that instruction.

The contact-confirmed placement error was 32.925970 mm, about 3.3 cm. That completed within the existing 40 mm tolerance, but it is only modest precision. A matched familiar-word box control with composition disabled produced exactly the same error. This supports an existing placement/perception limitation in that scene rather than a regression caused by the new noun; it does not establish generally precise manipulation.

“Could you keep tracking the object that is crimson” transferred track to a red moving cylinder with a distractor. All 55 evaluated post-warmup paired frames were fresh, out of 80 total frames over 3.2 simulated seconds. Independent-grader RMS error was 5.020084 pixels and maximum tracking/grader association discrepancy was 0.257604 pixels. There were no hand actions or body contact incidents. The separate process ran 455.1 seconds after track teaching.

An ambiguous cuboid reference and unsupported motion negation both declined without action or episode credit. Disabling composition rejects the new forms while keeping the older simple forms available. Shape constraints and negation in moving-target instructions are refused rather than silently ignored.

What survived the extra teaching

Before the new lessons, follow and crimson recalled after 3.85 and 3.84 hours with exactly the previous 5.020084-pixel pursuit RMS error. Earlier left-of placement stayed at 5.559383 mm error after 6.37 hours from teaching, and again after four new lexical lessons using an exclusion-form instruction.

Earlier right-of placement stayed at 7.181126 mm error after 8.07 hours, four new lexical lessons and a 60-second pause with gaze interference. Richer relative-reference wording also stayed within the declared 1 mm regression tolerance. These are specific checks, not a guarantee across all scenes or behaviours.

The report records preservation of all 38 earlier native episodes, 982 earlier visual observations, 345 accepted component hashes and prior vocabulary, relation/correction state, identity and cue/forward memory. Eleven primary sessions appended observed episodes; two extra refusal sessions earned none. Totals reached 49 episodes and 1012 visual observations. Cortical prototypes grew from 367 to 377 under the unchanged 2 MiB budget. All 637 previously qualified inputs retained identical decoded centres, with unchanged mean per-observation RMS error of 0.423633056 pixels. These are the later baseline’s counts; earlier journal checkpoints retain their original figures.

Finding the stalled graphics call

New timing markers separate GLFW initialization, window creation and destruction. The sandbox probe stopped at 35.888 seconds total with glfw_init unfinished, before any create-window call. Its event history remained unchanged. Successful host initialization calls took 0.001735–0.015281 seconds.

This localizes the observed stall to GLFW initialization in that sandbox context. It does not identify or repair the underlying OS/driver cause, or prove that every graphics stall has the same cause. The renderer, models and security/desktop settings were unchanged.

The same bounded-job limits remain: 30 seconds per graphics startup, 10 per shutdown, 180 seconds total and 1536 MiB memory. Maximum successful-Job memory was 1217.18 MiB. These are trial limits, not a description of the whole PC. Working within the restricted setup remains part of the efficiency challenge.

Evidence and limits

The completed report records 18 passing focused tests, six staged rendered cases, eleven primary sessions with new observations and two additional no-action refusals. The website review matched all 16 selected native source hashes and 14 selected acceptance-receipt hashes. It did not start or modify Owen.

Cuboid maps to the retained box category; arbitrary real-world cuboids, new materials and new visual categories are unqualified. Track and follow mean gaze pursuit, not locomotion. Crimson still maps to red without learned shade discrimination. General motion negation, shape-selective pursuit, prolonged occlusion and moving-object manipulation remain unsupported.

Cross-session intervals above are actual elapsed timestamps. The separate 60-second pauses preserve caller-supplied text while physics pauses; they do not demonstrate autonomous prospective language memory. The new noun/verb recall covers minutes, not a new multi-hour interval. General English and broad lifelong retention remain unproven.

The cortical pathway remains shadow-only. Recurrent replay, the extra taxel adapter and advanced vision remain disabled, and the attended microphone test is still pending. No new original simulator capture is supplied by this report; the gallery keeps its existing capture dates.