The research gallery

Qualification correction: accurate relation estimates and the recorded numerical pass did not establish correct rotation in the decoded scene. Structures can be missing, fragmented or incorrectly oriented, not merely blurred. Rendered rotation remains unqualified. The original verdict, measurements and downloadable records are retained as historical evidence.

Fifteen of sixteen demonstrated changes were accepted. Median geometric displacement error, including the refusal, was 0.93% of the image short side. The particular rotation, scale and displacement were inferred from image pairs; the possible geometric family was supplied.

A second operation used only the saved first-stage VSA and a second inferred relation. After two operations, RGB mean squared error was 0.018995 (pixel values scaled from 0 to 1; lower is better) versus 0.057015 for no geometric change through the same number of passes, and 0.048183 when the first change was skipped. Those are reductions of 66.7% and 60.6%; all 32 queries improved against the unchanged comparison.

All five predefined numerical criteria passed, and independent checks reproduced all 416 image comparisons. The first-step comparison against wrong demonstrations remains statistically uncertain.

Supplying the true geometry still produced mean RGB error of 0.011179, compared with 0.000570 for directly encoding and decoding the desired photograph. A subsequent component diagnostic found local rotation failure even with original encoder features and exact geometry. Recovery from source VSA is therefore not required for the failure; this does not isolate every contribution from feature transformation, composition and decoding.

Rotation, uniform scaling and translation are a supplied family, and the two-step procedure is supplied. The system fits numeric relations without neural retraining or acquiring a new operation family. Refused fits use explicit identity fallbacks; the refused example remains in all totals. Desired query images and known geometry are observer-only references in this experiment.

The five historical numerical criteria passed and the first-stage wrong-demonstration confidence interval includes zero. Neither fact establishes intact geometry in the rendered result. Relation fitting and continuation through saved VSA remain separate evidence; the application path is not a qualified visual rotation mechanism. No semantic editing, invented detail, general imagination or runtime promotion is established. The continuing life is unchanged.

Original-feature transport lowers training image error 6.09% but raises development error 5.14%. Recovered synthesis lowers training feature error 32.06%, mostly in the coarse overview, while image error improves about 1% on both training and development. The audit passed all 960 comparisons. Fixed-panel visual review finds recognizable scenes and persistent blur or overlapping moving contours, not consistently coherent middle frames.

Training examples were previously used for learning. Development videos are separate from constructor training but are existing diagnostic data, not a fresh test set. Original encoder features are privileged inputs. Feature error is not image quality or coherence; low error on dark or static regions is not temporal understanding. All outputs pass through VSA before the frozen renderer.

Original encoder features improve the untrained blend by 2.91% on whole-image error and 4.26% on changing pixels. Against their own original-feature blend, transport changes changing-pixel error by −0.60% and synthesis by +0.19%, with different effects between the two videos. Neither reaches the required improvement on each source. Complete visual review finds no consistent learned repair of doubled moving contours.

The same previously exposed test cases from two videos. Original encoder features are privileged inputs at fitting and evaluation; they do not qualify a VSA-only input architecture. All candidates still recompose VSA before the frozen renderer. This is a bounded diagnostic, not independent validation or a complete causal isolation of the learning limitation.

Against the matched feature blend, synthesis lowers whole-image error by 0.29% but changes changing-pixel error by +0.14%. Transport changes changing-pixel error by +3.91%. Neither reaches the required improvement on each video. Synthesis makes small gains on many frames; two train cases outweigh them. All 32 cases remain visible, and neither model consistently resolves doubled moving contours.

The same 32 previously exposed test triplets from two videos. This is a diagnostic follow-up, not new independent validation, an optimized model, or a causal isolation of every source of error.

Both clean-conversion gates passed. Compared with VSA interpolation, error where endpoints differ was 8.2% higher for learned transport and 4.3% higher for synthesis; whole-image error was 2.7% and 2.4% higher. Neither constructor qualified. All 32 reviewed examples preserve recognizable scenes, but neither method shows a consistent visual advantage; moving contours often remain doubled or blurred.

A small pilot using nine previously available videos, with two test source videos. Frames within a video are related observations. Analysis images are resized to at most 256 pixels on the long side. This is neither free-form imagination nor language grounding.

Admission completed 128 training triplets, then failed while rendering the first development diagnostic: CUDA vectors reached a materialization boundary requiring CPU vectors. No adapter or construction model was fitted, and no model-quality conclusion is available.

This attempt stopped before scientific qualification or construction. The gallery contains only the fixed training-input preview, not test reconstructions or generated predictions.

Adding image-reconstruction feedback did not qualify the trained component for hole filling. Mean hidden-interior error rose by 10.9%, but the paired uncertainty interval includes both improvement and harm. Mean error in already visible pixels rose by 30.8%, with 31 of 32 photographs worsening. In the separate diagnostic that supplies corrupted clean-target information, hidden-pixel error fell by 19.9% to 29.8% against the matched model across the three tested noise levels. That denoising improvement did not transfer to reliable completion from a masked photograph. Complete visual review found no reliable coherent filling. Independent audit passed; the recovered comparison is closed.

This is one fixed loss intervention, not a search for the best architecture or coefficient. The codec saw the training collection. Fresh development and test sources are outside recorded prior exposures. Privileged noisy-target denoising is separate from target-denied completion. Numerical improvement alone cannot establish coherent completion; the 363-encounter life remains fixed. This is an already exposed panel from the interrupted comparison, not a fresh independent test or a second replicate. On preserved figures, 'No learned regional context' means 'No image context supplied to predictor': the same trained network receives zeros for both regional image vectors and its coarse image overview. Its learned weights, geometry, visibility mask and noise remain. After prediction, wholly visible regions are restored, and the final renderer still receives the shared masked overview. The photograph therefore remains context for the overall system. The causal question is: Does supplying the photograph to the predictor help beyond the photographic information already supplied to the decoder? This is neither an untrained model nor a final image produced without any context.

Unexpected external termination before paired observation. Both fixed fits and diagnostics and all64 vector-only requests completed;59of64 decoder-aware requests were saved. All partial evidence is preserved. No complete objective-quality verdict is available. A separately registered recovery will reuse completed stages and repeat only the incomplete sampling phase without changing the scientific methods.

This is one fixed loss intervention, not a search for the best architecture or coefficient. The codec saw the training collection. Fresh development and test sources are outside recorded prior exposures. Privileged noisy-target denoising is separate from target-denied completion. Numerical improvement alone cannot establish coherent completion; the 363-encounter life remains fixed. On preserved figures, 'No learned regional context' means 'No image context supplied to predictor': the same trained network receives zeros for both regional image vectors and its coarse image overview. Its learned weights, geometry, visibility mask and noise remain. After prediction, wholly visible regions are restored, and the final renderer still receives the shared masked overview. The photograph therefore remains context for the overall system. The causal question is: Does supplying the photograph to the predictor help beyond the photographic information already supplied to the decoder? This is neither an untrained model nor a final image produced without any context.

Completed the fixed full-dimensional candidate: 2000 updates, 32 held-out photographs, 64 requests and all 960 outputs. Correct context reduced hidden-interior RGB MSE 19.2% versus unchanged masked decoding and 15.4% versus absent network context; photograph-bootstrap intervals favored context. Visible RGB MSE increased 80.0%, exceeding the registered allowance, with 63/64 requests over their visible-error allowance. Complete visual review found contextual colors and blurred patches, without convincing coherent completion. Privileged diagnostics showed that vector-error gains can accompany worse decoded imagery; this does not isolate the cause. Independent frozen audit passed first attempt with exact remeasurement of 960 completions and 72 saved diagnostic outputs; 504 diagnostics remain numeric-only. The codec, pattern bank and 363-encounter life remain unchanged; no generated content was admitted or candidate promoted. The model is 16.47 MiB; retained bulk evidence is 44.22 GiB logical, including 26.20 GiB of shared input packets. All samples, native detail views, controls, report and companion are local. Stop this candidate and propose a controlled frozen-decoder-aware component qualification before another fresh held-out completion test. Nothing published or pushed.

Full-dimensional value paths remove the previous fixed linear output-subspace restriction; they do not guarantee adequate coordination or coherent completion. The separate development denoising diagnostic receives noise-corrupted clean targets and is explicitly privileged. Only the independent test completion uses target-denied inputs. The codec and 363-encounter life remain fixed.

Both registered training-only PCA output spaces failed before denoiser training: mean development RGB error rose from 0.001384756 to 0.003934376 at width512 and 0.002807358 at width1024, increases184.1% and102.7% against a10% allowance. All32 development sources worsened; independent audit passed and all figures were reviewed. No training, generation or test evaluation occurred. This rejects the tested spaces, not VSA or every small nonlinear model. The codec and363-encounter life remain unchanged. A separate full-dimensional-value candidate is planned with a fixed budget.

This diagnoses two proposed linear output spaces before denoising. It does not test whether a trained specialist can complete hidden content, or whether VSA can support that function. Development images were used for the capacity decision; the separate test photographs were encoded but received no generated completions. The existing codec and 363-encounter life remained unchanged.

Twelve of sixteen visual cues recalled the intended photograph; four were refused and none accepted a wrong memory. All 120 scheduled decoded states were finite across seven contributing triplets. Direct VSA bundling produced recognizable, context-sensitive overlays close to pixel blending, not demonstrated coherent scene construction. Halfway image readback MSE was 0.000120/0.000117 while raw-vector relative L2 changed 21.54%/19.44%. Original memories, vocabulary, codec and working state were preserved. The full eight-thread independent audit passed; a four-thread audit rounding mismatch is preserved and explained. No experimental output was changed. Learned persona and multi-image associative scene construction remain open.

Equal native canvas dimensions do not align objects. Context and weights are supplied; no persona or semantic scene organization is learned. A different picture, or a stable round trip, does not establish useful imagination. Refusals and incompatible recalls remain in the record.

The independent audit passed. Expanded selection improved 26 of 64 exchanges, worsened one and left 37 unchanged; the original bank had the same counts. Expansion reduced mean error by 1.1%, but the whole-photograph uncertainty interval crossed zero. Direct prediction improved 52 exchanges and reduced mean error by 74.8% relative to expanded selection, while worsening 12 requests. Earlier behavior survived: 44 correct recalls, 31 successful replay anchors and 32 exact editing sequences. The formal result is no established expanded-bank advantage.

One frozen codec and readout; supplied feature positions and exchanges. The photographs are independent of recorded acquisition and fitting, but this does not establish object recognition, autonomous goals or pose invariance. Regional targets use observer-only clean features through the same decoder. The original runtime path remains unchanged.

The independently audited vocabulary grew from 512 to 1,758 patterns. Expanded cleanup recalled 44 of 96 new-photo cues, versus 35 with the original bank and 68 without cleanup. Its source-paired gain over the original bank was 9.4 percentage points (95% interval 2.1 to 17.7), but two old-cue recalls were lost and offset by two gains, failing the registered individual-retention criterion. All legitimate conditions rejected all 160 unfamiliar cues after acquisition. The established no-cleanup path retained all 44 earlier correct recalls, successful temporal behavior at 31 shared anchors, and exact outputs for all 32 editing sequences with the original bank. The occurrence IDs use 0.89 MiB; vocabulary and occurrence payload add 156.71 MiB while dense memories remain. Keep no cleanup as the default and test useful cross-photo vocabulary transfer at the fixed learned memory-readout boundary next. Milestone 1.3 remains open.

One codec and one append-only acquisition policy. New-photo queries revisit the acquisition images; this is not transfer across pose, scale or independent instances. Added vocabulary and references consume extra storage beside the dense episodic memory.

Launcher failed before model inference because importing the frozen source created Python bytecode caches, rejected by the strict Python-only source closure. No acquisition or recall measurements occurred. Preserve this attempt; repeat identical prepared inputs and scientific settings with Python bytecode writing disabled.

One codec and one append-only acquisition policy. New-photo queries revisit the acquisition images; this is not transfer across pose, scale or independent instances. Added vocabulary and references consume extra storage beside the dense episodic memory.

The independent audit passed. With 307 remembered encounters, the current recall rule correctly recovered 44 of 80 known image cues and returned no memory for the other 36; it made no wrong responses on those known cues. One of 64 unfamiliar cues was incorrectly accepted as a memory. All 32 earlier two-edit sequences reproduced both intermediate and final vectors, along with their decoded crops, exactly. Held working content survived saving and restarting and could be released normally. Temporal replay was tested with an allowance of 14 million comparisons; the default limit of 4 million refused the search. This establishes the behavior of the tested current system, including its limitations. It does not establish that every historical scenario is integrated or preserved.

This records current successes and failures; it does not certify that every historical capability is preserved. No new experience is acquired in this baseline. A successful engineering check is distinct from correct recall.

Held edited content supported useful second operations after new experience and restart under the prospective provisional tolerances. Supplied requests, errors and the experimental persistence boundary remain explicit.

32 supplied feature-exchange sequences from 16 fresh photographs, four shared incoming photographs, one frozen codec/readout/bank. Protected working storage is not neural distraction resistance. Edited-work persistence is experimental; no default-runtime promotion or learned task selection.

The fully fitted readout improved 32 requests, worsened one and left 31 unchanged. Mean regional squared RGB error was 0.000227488: 75.4% below fitted scalar attenuation and descriptively 19.0% below the previous readout. Fitted scalar attenuation made no changes. The registered primary correction comparison had a positive whole-photograph bootstrap interval.

The earlier correction mattered too. Removing it while keeping its scalar reduced useful exchanges from 26 to seven. The previous readout beat its scalar alone on 24 requests, lost on none and tied on 40; the source-bootstrap interval supported the difference. Small correction amplitude had not established a small contribution to discrete pattern selection.

All 26 previously useful exchanges remained useful, and six additional requests improved. The full fit changed 33 requests across 15 photographs. The privileged Clean selector improved 38, worsened none and left 26 unchanged, with mean error 0.000211005.

The one regression increased regional RGB error by 18.4%; its edge error improved by 11.1%, and no request doubled RGB error. It failed the predefined no-increase-in-excess-harm rule relative to the previous readout’s zero observed regressions. The primary correction advantage passed; the complete bounded-use criterion did not. The recorded rule and verdict remain unchanged.

The full fit passed the numerical convergence checks. Validation vector loss reached 0.2304, versus 20.5379 for the previous checkpoint, 0.8316 for fitted scalar attenuation and 0.8363 for zero output. Independent validation checked 1,162 files and recomputed all 768 pixel and edge measurements. The research life grew from 271 to 287 encounters while preserving original memories, the codec and held work.

The image, local region and feature positions are supplied. The exact-feature reference defines an operation through the same decoder, not an edited ground-truth photograph, autonomous action or movement of an identified object. The Clean selector remains privileged.

Four requests share each photograph. The position groups appeared in prior validation; independently new evidence comes from the evaluation photographs. The readout retains 4,097 optimized coefficients plus the fixed 8 MiB basis with 2,097,152 entries.

Thirty-one full-readout requests made no change by selecting the same bank pattern twice; this is not learned confidence. Lower vector error alone is not sufficient for useful composition. No automatic promotion or additional component-optimization sweep followed this comparison.

The trained readout improved 23 requests, worsened none and left 41 unchanged, with 23 changes across 12 photographs. Untrained improved 33, worsened 27 and left four unchanged. Post-experiment analysis found that training avoided all 27 untrained losses by making no change, but also lost ten useful changes.

Mean regional squared RGB error was 0.000263286 for Trained: 62.9% below leaving memory unchanged, but 5.5% above Untrained and 1.55% above corrected independent selection. The primary 95% source-bootstrap interval crossed zero. The learning-improvement and bounded-candidate criteria were not met.

The trained readout had no observed regressions against unchanged memory on this panel. Untrained had 25 regional error doublings; corrected independent selection had eight. Matching memory mattered: Trained beat inference using another photograph’s memory. Privileged Clean selection improved 32 requests, worsened none and left 32 unchanged.

Validation vector loss fell from 327.946 to 20.538, but an analytic zero-difference output scores 0.836 on the same saved targets. The budget ended at 1,024 steps while loss was still declining. This post-experiment check identifies unresolved optimization, not a ceiling for the readout architecture. A zero-difference output would make no changes, so vector loss and decoded usefulness remain separate tests.

The global scale learned a value of 0.2495. The residual correction’s output norm was 12.8% of the scaled input’s norm; without a scalar-only control, its contribution to useful pair choices is not isolated. Independent validation checked 1,325 files and recomputed 768 image measurements. State continuity, held work, the bank, codec and pixels outside the edited core were preserved.

No-ops select the same pattern twice; they are not learned confidence judgments. Zero observed losses across these 64 requests does not establish zero risk elsewhere, and the measured tradeoff does not change the failed primary criterion.

The operation and positions were supplied. Withheld position groups appeared in validation tasks and all-position memories, although excluded from gradient and basis fitting. Evaluation photographs were independent; these were not untouched zero-shot position tasks. Clean and the exact-feature reference remain privileged.

The two transfer groups differed: mean error improved 14.6% versus Untrained for withheld pairings of trained positions and worsened 25.4% for positions excluded from gradient and basis fitting. These are descriptive subgroup results, not separate confirmation studies.

Exact cancellation reduced mean regional squared RGB error by 24.4% versus the same selector with estimated background: 0.000213194 versus 0.000282181. It improved 25 requests, worsened ten and tied on 29 in that direct comparison. The 95% whole-photograph bootstrap interval for the absolute reduction was positive, meeting the predefined mechanism-diagnostic criteria.

Against leaving memory unchanged, cancellation improved 32 requests, worsened nine and left 23 unchanged. Six changed requests at least doubled regional error. Interference removal helped the average without establishing reliable automatic changes.

Removing the remaining stored-versus-reencoded residual changed only two outcomes, both slightly worse. Clean difference retained 44.3% lower mean error than Clean pair sum. The fixed bank supports more useful changes with an operation-directed target; isolating the pair under the paired-sum objective did not recover that advantage.

Independent validation checked 1,665 evidence files and recomputed all 768 pixel and edge measurements. Retained state and codec were preserved, and pixels outside the edited core matched exactly.

All 64 requests were already evaluated. Exact cancellation and both clean controls use privileged original features. This is a mechanism diagnostic, not new-input generalization or autonomous recovery.

The supplied exchange concerns feature positions with overlapping pixel receptive fields. Its exact-feature reference is not an edited ground-truth photograph or evidence of object movement. Large proportional regional losses can concern small intended changes.

The result supports interference as a contributor to this selector’s error, not the sole cause. Historical normalized adequacy statistics are retained as context; they are neither this study’s primary criterion nor a universal cognition test.

Background correction reduced mean regional squared RGB error by 41.2% versus raw independent selection on the new requests: 0.000266 versus 0.000453. Fitting the pair jointly increased error by 6.1% versus corrected independent selection, to 0.000282. The useful change in this comparison was the background correction.

Lower averages did not establish reliable edits. Centered improved 29 requests, worsened 13 and left 22 unchanged. Joint improved 29, worsened 14 and left 21 unchanged. Among Joint’s 43 changed outputs, 13 increased regional error by more than 10% and ten at least doubled it. Centered had 13 such increases and nine doublings among 42 changed outputs. Both missed the predefined diagnostic benchmark.

The privileged Clean control met all six diagnostic criteria again: mean error 0.000119, with 32 requests improved, one worsened and 31 unchanged. Its one worsening exceeded 10% but did not double error. The same vocabulary and decoder can support useful combinations with clean information; recovering useful combinations from the bundle remains unresolved.

Matching memory content mattered: Joint beat selection from another photograph’s memory on 50 requests, lost on seven and tied on seven. The paired source-level comparison supported that advantage, but not an advantage over Centered. Joint changed only six selected pairs relative to Centered; one improved decoded error and five worsened it. These request-level counts are post-experiment descriptions, not extra acceptance criteria.

Independent review checked 2,777 evidence files and recomputed all 1,536 pixel and edge measurements. Unchanged readbacks and pixels outside the selected core matched exactly. The research life grew from 159 to 175 encounters while preserving earlier memories, the acquired bank, codec and held work. The separate reused panel showed the same direction of results and was not pooled with the new photographs.

Image, region and feature-position addresses were supplied. This tests a represented feature exchange, not autonomous object discovery or object motion. The successful Clean selector remains privileged, and automatic constituent substitution stays disabled.

Large proportional error increases can concern small requested effects; they do not mean the whole image was destroyed. The diagnostic criteria retain those regressions rather than allowing large gains elsewhere to cancel them. Two weak-reference requests remain in the 64-request totals and contribute zero normalized gain.

The estimated background is the learned bank’s mean contribution, not the actual contents of the other positions. Failure of this recovery rule does not establish that the retained information is absent or that another recovery method would fail. The clean reference is not an upper bound on decoded quality.

On the 64 new requests, joint selection reduced mean regional squared RGB error by 41.0% versus independent selection and 69.9% versus leaving memory unchanged: 0.000216, 0.000366 and 0.000718, respectively. It met all six predefined diagnostic criteria, and the paired comparison supported an advantage across source photographs.

Joint selection improved 34 requests, worsened one by 0.203%, and made no change on 29. None of its 35 changed outputs increased error by more than 10% or doubled it. Independent selection improved 27, worsened 18 and left 19 unchanged; it missed the diagnostic benchmark.

The improvement was not only a result of doing less. A post-experiment comparison on the same 35 requests where joint selection acted found 37.5% lower mean error than independent selection. Its individual patterns were less similar to the original features on average: cosine 0.9341 versus 0.9609. A pair can express a useful change without faithfully identifying the original parts.

The separate reused panel showed 48.2% lower mean error with joint selection than with independent selection. It met the diagnostic criteria there too, with 37 improved requests, one slightly worsened and 26 unchanged. Those previously evaluated photographs were not pooled with the new panel.

Independent validation checked 2,202 evidence files and recomputed all 768 pixel and edge measurements. Baseline readbacks and pixels outside the selected core matched exactly. The research life grew from 143 to 159 encounters while preserving earlier memories, the bank, the codec and held work.

Both selectors received privileged clean features, and the feature positions were supplied. This establishes a useful operation-specific diagnostic with the existing vocabulary, not autonomous constituent recovery, object movement or selection from noisy bundled memory. The exact-feature reference is not an edited ground-truth photograph.

Joint selection retained all 19 no-change outcomes from independent selection and added ten more. Those distinctions remain unresolved. The active-request and individual-similarity comparisons are post-experiment descriptions, not additional acceptance criteria.

Four requests share each source photograph. Five requests with weak reference effects remain in the totals and contribute zero normalized gain under the predefined rule. The criteria are provisional diagnostics for this operation, and automatic constituent substitution remains disabled.

On the 64 previously unevaluated requests, clean-selected patterns reduced mean regional squared RGB error by 46.3% versus leaving memory unchanged: 0.000382 versus 0.000711. Bundle-selected patterns reduced it by 12.6%, to 0.000621. Removing selection noise therefore improved average fidelity, but the remaining pattern substitution still lost useful distinctions.

Clean selection improved 28 requests, worsened 19 and left 17 unchanged. Bundle selection improved 21, worsened 33 and left 10 unchanged. Neither met the predefined reliability benchmark. Eleven clean-selected edits and 21 bundle-selected edits at least doubled the error. Large absolute gains on some requests can coexist with large relative losses on smaller intended changes.

The paired comparison supported an additional error cost from bundle selection. The separate reused panel showed the same direction: 78.2% lower mean error with clean selection and 48.6% lower with bundle selection, but neither met its benchmark. Reused and previously unevaluated photographs were not pooled.

A later inspection found that 17 clean-selected requests mapped both feature positions to the same learned pattern, erasing the distinction needed for the exchange. Bundle recovery chose the clean features’ nearest pattern in only 38 of 128 constituent selections. These observations identify problems in both nearest-pattern substitution and recovery; they do not show that adding more patterns alone would solve them.

Independent numerical and preservation checks passed, including 576 pixel and edge measurements. Pixels outside the selected regional core, the retained state and the codec remained exact during readback. Automatic constituent substitution remains disabled.

These are supplied feature-position exchanges with overlapping pixel receptive fields. The exact-feature reference is not an edited ground-truth photograph or evidence of object movement. The learned bank contains patterns from five earlier photographs.

Clean selection receives information unavailable to the operative memory path. Its nearest-cosine pattern is not guaranteed to be the best decoded substitute. The conclusion applies to this bank and selection rule; it does not rule out other ways to compose VSA representations.

Four requests share each source photograph. Uncertainty was estimated by resampling whole photographs, and the recovery-penalty interval is wide. The comparison distinguishes selection paths without establishing an additive causal decomposition of every source of error.

Before acceptance filtering, exchanges using learned patterns had 48.6% lower mean regional squared RGB error than leaving memory unchanged: 0.000530 versus 0.001030 across all 32 requests. Error is measured against the exact-constituent decoder control, not the original photograph.

That lower average did not translate into reliable individual operations. Nine proposals improved, 20 worsened, and three made no change. All 20 losses exceeded 10%, and 17 proposals at least doubled the error. Large absolute improvements on a few requests outweighed smaller absolute losses in the overall mean; the criteria also measured each request relative to no change.

None of the 45 acceptance rules met the usefulness and harm criteria. The experiment stopped with a negative development result. Sixteen fresh photographs had been prepared for confirmation, but no confirmation admission or model inference took place. The existing memories were preserved.

A later audit of saved vectors found a bank pattern with at least 0.95 cosine similarity for 62 of 64 clean constituents. Recovery from the bundle selected the nearest clean-feature pattern only 20 times. This identifies recovery mistakes, but does not establish that the nearest available pattern in vector similarity would produce an adequate decoded exchange.

The independent numerical and preservation checks passed. The stored state, held memory and codec stayed exact, and decoded pixels outside the requested region were unchanged.

These are supplied feature-position exchanges, with overlapping pixel receptive fields. The exact-constituent control defines an operation within the representation; it is not a true edited photograph, object movement or autonomous choice of what to manipulate.

The panel contains eight exposed development photographs and four related operations per source. Five control exchanges had effects too small to count toward useful coverage; all requests remained in the denominators. No fresh confirmation result was obtained.

The saved-vector diagnosis is post-experiment evidence. It does not yet separate recovery errors caused by bundled content from approximation errors introduced by replacing a feature with a learned pattern. A large cosine match alone does not guarantee a useful substitution.

Once populated, the cache reduced median search-and-recall-application time from 4.154 to 1.022 seconds: 75.4% lower. The 95th percentile fell from 4.679 to 1.109 seconds. Admitting a new photograph took a median 0.592 seconds without the cache and 0.590 seconds with it across eight paired admissions.

The tradeoff was 2.99 GiB of temporary GPU cache. The highest GPU tensor allocation measured across admission and query phases was 3.35 GiB, below the 6 GiB criterion. The cache is not saved with the memories and does not reduce durable storage.

All 384 paired decision comparisons agreed: 64 fields × three passes × two acceptance rules. Cached and uncached scores were exactly equal; checks of recalled content, remembered continuation, decoded images, new admissions and saved state also passed. These repeated comparisons come from 16 source photographs, not 384 independent inputs.

Every predefined correctness and resource criterion passed. This establishes a faster optional execution path on the tested history and hardware; it does not improve recall quality. The original unresolved queries and false match remained.

Timing includes cue encoding, memory search and application of the baseline recall rule. Image decoding, remembered continuation, the alternative acceptance rule and persistence were checked separately, outside the timer. The 75.4% reduction therefore describes this query path, not the complete cognitive loop or application startup.

The first empty-cache query took 4.588 seconds. First queries after reload and clearing took 3.732 and 4.158 seconds, respectively; each is a single observation. A 64 MiB fallback cache had a 3.761-second median on an eight-field subset, which does not establish a full-panel speedup.

This is one device, memory history and process comparison. Cues were related views of 16 photographs. An unfamiliar cue with 64 visible features still produced a false recall under both implementations and both acceptance rules. The cache remains optional and is not enabled by default.

Completed recovery of the interrupted comparison: 32 newly admitted photographs yield 25/128 baseline versus 60/128 support-scaled correct partial-cue paths, each requiring the intended anchor and both successors (+27.34 percentage points; source-bootstrap 95% interval 19.53–35.16). Full local fields give 48/64 under both policies. Six numerical/causal panels and the integrated hold/save/new-intake/resume proof pass. Earlier-photo recall loses successes as competition grows; video partial paths are 1/32 versus 4/32. All final novel fields reject, but earlier full-field false accepts remain. Bounded 1.6 is demonstrated; quality is provisional.

Supplied 80×80 local visibility and two-step control, one continuing life; the full-field condition is not the full photograph. Recovery reuses exposed inputs. Earlier recall regressions, earlier novel-source false accepts, and weak correlated video remain visible. Remembered order is not prediction or learned attention. Prior confidence criterion remains failed; baseline default unchanged; full 1.3 and final sealed panel remain open.

Execution was externally interrupted during the reported daemon restart. The last after-stage log records query 95/480 at 2026-10-09T20:03:20.001514+00:00; host inspection found no surviving study worker or launcher. Before and middle source-denied receipts, ordinary acquisition, integrated acquisition, and integrated continuation receipts passed. The final after panel, restart/context controls and independent audit did not complete. This is an incomplete execution, not evidence against the scientific hypothesis. All 445 frozen source/input hashes and 55 dependency hashes remain unchanged; partial outputs and the preliminary visual are preserved. Any repeat must use a separately registered recovery attempt and disclose already exposed inputs.

Supplied local visibility and replay control; fresh-photo comparisons and correlated windows from two previously examined trailers. Remembered encounter order is not prediction or learned attention. No default promotion; full milestone 1.3 and the sealed final panel remain open.

Completed valid negative qualification: support-scaled correct partial recall 111/192 versus baseline 66/192 (gain 23.44 percentage points; source-bootstrap 95% interval 17.71–29.17). No wrong present-source accepts. Never-admitted partial false accepts 7/384 versus 5/384; two sources gain an additional erroneous field, giving a 9.51% upper bound above the required 5%. Both sources already had baseline errors elsewhere. Fixed boost gives 121 correct and 18 novel wrong fields. All independent and continuity audits pass; full 1.3 stays open.

A supplied visibility/count policy with one fixed codec and memory construction, not learned confidence or general novelty recognition. Source-owner tests use local fields. No policy is promoted; full milestone 1.3 remains open and the earlier final32 panel stays sealed.

FAILED exact top-two criterion: runner-up handles 377/208 change order for cue 72/slot 43 before and after admission. Selected handle 30 stays unchanged. All 17,152 selected-prototype choices and 1,728 ordinary/counterfactual recall decisions agree, from 8,781,824 prototype comparisons. These are passing subchecks, not an overall audit pass. Original run 001 retains its score-cap failure. Full diagnostic figures are on the run 001 page.

Post-result verification of existing evidence, not new inference or independent replication. Original run 001 remains failed. No threshold, cleanup policy, acquired state or final-panel exposure changes; full milestone 1.3 remains open.

Failed before numerical computation: the isolation wrapper omitted the /dev mount and PyTorch could not open /dev/urandom. This is an execution-binding failure. The identical numerical method is separately registered as run003 with the established device and process mounts restored; this failed attempt remains preserved.

Post-result verification of existing evidence, not new inference or independent replication. Original run 001 remains failed. No threshold, cleanup policy, acquired state or final-panel exposure changes; full milestone 1.3 remains open.

Original run 001 remains failed: prototype-score error 2.5413459939782612e-6 exceeded its fixed 2.5e-6 limit. Separate run 003 confirms every selected prototype and recall decision but also fails its exact runner-up-order criterion; both failures remain. Fresh clean-field recall is 27/32 without cleanup versus 23/32 with acquired cleanup; known-target recovery works at interference weight 0.25 but rejects weight 0.5. A separate 47-to-63 encounter fork preserves old memory and held recalled episode 45. No cleanup policy is promoted.

A diagnostic with fixed codec, exemplar bank and historical policy, not a selected remedy or completion of milestone 1.3. Sixteen fresh photographs bound generalization; known-target VSA controls are separate from natural recall. No cleanup remains the runtime default and the final 32-image panel remains sealed.

Bounded 1.5 passed: all 276 temporal choices were correct across original, shuffled and context-swapped storage; all 1,263 independent audit checks passed. Seven forward and six backward steps decoded actual retained working content. Full and restarted replay matched exactly, and subsequent acquisition matched the no-replay control. Qualification changed decoded pixels by at most 2.98e-8. Noisy cleanup remained ambiguous (margin 0.0159 < 0.02), so no prototype was substituted. The verified continuation contains 47 encounters and holds recalled episode 45. Full 1.3 and native visual movement remain open.

Supplied roles, coordinate bases, temporal horizon and replay requests; no autonomous selection, spatial perception or replay learning benefit. Six irregularly timed natural observations extend a retained life. Full 1.3 remains open; final 32-image panel remains sealed.

Working hold controls passed across three content periods and eight new observations. The life retains 40 encounters. Exact restart match: True; acquisition and cursor agreement across all branches: True. Internal-query outcomes are reported separately; no autonomous selection or complete 1.3 claim.

Supplied generic hold/release controls and a fixed replacement cue; no learned control policy, improved perception or complete milestone 1.3 claim. Four screened new photographs and four new-in-life frames extend an already acquired life. Final 32-image panel remains sealed.

Engineering controls passed. Final no-cleanup queries accepted 25/32 correctly, 0 wrongly and abstained on 7. Before admission, 0/8 absent cues were falsely accepted. Cleanup outcomes are recorded separately against no-cleanup; these results do not close milestone 1.3.

Previously examined natural inputs and supplied visibility; not a fresh generalization benchmark. Dense eight-bit storage and the immutable codec are provisional machinery. A 512-entry acquired bank is bounded. Engineering continuity and scientific cleanup/novelty outcomes are separate; no complete milestone 1.3 claim or unseen final-panel evaluation is assumed.

The attempt failed after 16 observations: all 24 narrow queries per cleanup condition were below the fixed 0.40 cutoff, so accepted working-state use was never exercised. No-cleanup ranked the correct source first for 16/16 known cues, versus 10/16 with acquired cleanup, but neither condition accepted a recall. Known no-cleanup top scores were 0.383584–0.396666. A subsequent observer-only float64 audit verifies the 72 saved query vectors without changing this failed outcome. The separately registered wider-view recovery preserves this record.

Previously examined natural inputs and supplied visibility; not a fresh generalization benchmark. Dense eight-bit storage and the immutable codec are provisional machinery. A 512-entry acquired bank is bounded. Engineering continuity and scientific cleanup/novelty outcomes are separate; no complete milestone 1.3 claim or unseen final-panel evaluation is assumed.

Eight-bit dense storage preserves every baseline retrieval choice on all three assays while reducing tensor payload from 616.75 to 176.71 MiB (71.35%). Eight-bit residuals improve numerical readback at higher cost without changing recall; four-bit alternatives introduce paired regressions. All 17,055 independent checks and 400 source-denied readbacks pass. Dense8 is the candidate for larger confirmation; the final 32 inputs remain sealed and integration is pending.

This component comparison uses fixed acquired banks and supplied visibility. Examined and fresh development queries are reported separately; repairing known failures is not general noninferiority. Count the entire bank, references, fallbacks, quantized coordinates, scales, envelopes, serialization and model costs. Residuals restore the original VSA space before existing operations. No candidate is installed, the final panel remains sealed, and milestone 1.3 integration remains open.

On8 previously examined development queries, dense recall is 2/8 with gray-filled cues and 8/8 using only fully visible feature terms. The shared095 memory scores 6/8 with the same visible-term cues. All original cue vectors and retrieval cells reproduce exactly; no memory or model changed.

Development preflight selects methods using development evidence only; its outcomes are not final confirmation. The final run uses separately locked methods and inputs. The codec and sensory geometry are supplied machinery; pattern contents are acquired. Compact references are storage addresses, not semantic categories. Count the shared bank, references, structure, fallbacks, versions and model costs. Prototype reuse alone establishes neither object understanding nor integrated milestone 1.3. Rotation, scale, novelty and detail limits remain visible; no checkpoint is installed.

Five one-order development conditions completed on16 acquisition and8 independent development photographs. Candidates meeting provisional storage/recall criteria: shared095, conservative095. Persistence and source-denied readback passed; final photographs remain sealed.

Development preflight selects methods using development evidence only; its outcomes are not final confirmation. The final run uses separately locked methods and inputs. The codec and sensory geometry are supplied machinery; pattern contents are acquired. Compact references are storage addresses, not semantic categories. Count the shared bank, references, structure, fallbacks, versions and model costs. Prototype reuse alone establishes neither object understanding nor integrated milestone 1.3. Rotation, scale, novelty and detail limits remain visible; no checkpoint is installed.

Attempt failed; retained outputs are incomplete. See execution log.

Development preflight selects methods using development evidence only; its outcomes are not final confirmation. The final run uses separately locked methods and inputs. The codec and sensory geometry are supplied machinery; pattern contents are acquired. Compact references are storage addresses, not semantic categories. Count the shared bank, references, structure, fallbacks, versions and model costs. Prototype reuse alone establishes neither object understanding nor integrated milestone 1.3. Rotation, scale, novelty and detail limits remain visible; no checkpoint is installed.

The frozen 2 × 2 diagnostic supports a cross-version matching cost. Re-encoding the comparison bank raises occluded-cue top-1 accuracy from 65.28% to 81.94% for ordinary continuation and from 76.04% to 81.25% for replay; the original baseline is 81.25%. Paired gains are +16.67 percentage points [95% interval 9.72, 24.31] and +5.21 [1.39, 9.72]. Within-version top-1 change remains unresolved, and correct-memory cosine margins are lower. All 38 cells and parent repeats pass; independent statistics checks pass 61,578/61,578 and bank replay 475/475. Counterfactual banks require original RGB. No memory migration, training or checkpoint adoption occurs. The worker took 240.0 seconds, peaked at 8.41 GiB CPU resident memory and 225.5 MiB PyTorch GPU allocation; each transient full bank uses 1.948 GiB.

This is a post-result diagnostic on 96 previously examined memories and one fixed 20% occlusion mask, not a fresh held-out test. Regenerating banks requires original pixels; it is a counterfactual comparison, not source-free memory migration. Geometry and cue selection are supplied structure. Retrieval is appearance-instance matching, not semantic cleanup. Intervals are conditional on the fixed fits and images; overlap with zero does not establish equivalence. Bank recovery does not prove an invertible or useful bridge, and failure does not prove information destruction. Report regional bank bytes, peak CPU/GPU memory, durable storage and elapsed time. Parent evidence and operative memories stay unchanged; no training or checkpoint adoption occurs.

Completed and independently audited: replay lowers fresh Flickr MSE 5.28% and held-out stored-memory MSE 0.35%; ordinary training improves fresh Flickr 10.11% but worsens retained readback 21.92%. The registered verdict is acuity_and_retention_gain. Occluded-cue retrieval still falls from 81.25% to 76.04% with replay (ordinary 65.28%). Fine textures remain soft; larger-video MSE worsens 4.80% on average. Run-001's evaluation failed at a CUDA device-name check; run-002 is evaluation recovery of the same comparison, with exact inherited inputs and locked endpoints, no retraining and no reselection. Repeated baseline tensors are byte-identical. No candidate is installed.

Both arms refine the same joint native codec. Replay targets are archived old-decoder outputs, not original RGB truth, and preserving those outputs can preserve their errors. Retention training, development and final cohorts have separate roles; held-out retention is known acquired experience, while fresh final photographs measure new-input quality. Source-free cross-version readback explicitly retains acquisition and current decoder identities; it does not establish automatic codec compatibility. Geometry and spatial bindings are supplied sensory structure. Each complex128 region costs 256 KiB before overlap and other state; report bytes, exposure, replay overhead, latency and peak memory. Earlier lives retain their original models; no candidate is installed. Reconstruction, retention and measured retrieval remain separate from reading, learned meaning or general cognition.

Execution failed during evaluate: ValueError('isolation device must be cpu or cuda'); preserve evidence.

Both arms refine the same joint native codec. Replay targets are archived old-decoder outputs, not original RGB truth, and preserving those outputs can preserve their errors. Retention training, development and final cohorts have separate roles; held-out retention is known acquired experience, while fresh final photographs measure new-input quality. Source-free cross-version readback explicitly retains acquisition and current decoder identities; it does not establish automatic codec compatibility. Geometry and spatial bindings are supplied sensory structure. Each complex128 region costs 256 KiB before overlap and other state; report bytes, exposure, replay overhead, latency and peak memory. Earlier lives retain their original models; no candidate is installed. Reconstruction, retention and measured retrieval remain separate from reading, learned meaning or general cognition.

Native-size VSA readback and all three refinement regimes pass the registered photo gates. Joint training reduces COCO/Flickr MSE by 37.00%/30.18%, and writing-white/photo by 71.41%/34.16%; every fitted seed improves every final entry in whole-image MSE. Fine photo texture remains weak. Joint old native-vector readback worsens 25.79% and legacy-vector readback 38.38%; decoder-only improves stored native readback 39.60% but worsens legacy readback 13.10%. Stored vectors cost 38.6–42.2 times raw RGB on these panels. Independent audit 23858/23858 passed. No candidate installed; earlier lives and failed runs preserved.

Geometry, binary size coding and spatial bindings are supplied sensory structure, not learned number concepts or relations. Native regions and the overview consume additional vectors; fidelity gains are not equal-compression superiority. Each current complex128 region costs 256 KiB before overlap and other state. Report bytes per image and megapixel, latency, training exposure and peak memory. The overview is intentionally coarse; native detail regions preserve the complete admitted field. Component regimes have different trainable subsets. Earlier lives retain their original models; no candidate is installed. Better appearance does not establish reading, acquired meaning, general cognition or a VSA capacity ceiling.

Execution failed during train: ValueError('native neural pixels must be finite in[0,1]'); no rescued scientific verdict.

Geometry, binary size coding and spatial bindings are supplied sensory structure, not learned number concepts or relations. Native regions and the overview consume additional vectors; fidelity gains are not equal-compression superiority. Each current complex128 region costs 256 KiB before overlap and other state. Report bytes per image and megapixel, latency, training exposure and peak memory. The overview is intentionally coarse; native detail regions preserve the complete admitted field. Component regimes have different trainable subsets. Earlier lives retain their original models; no candidate is installed. Better appearance does not establish reading, acquired meaning, general cognition or a VSA capacity ceiling.

Retaining bundle magnitudes reduced reconstruction error by 14.38% on COCO and 13.63% on Flickr compared with equally trained phase-only readout; both writing conditions improved. Fine detail remains blurred. Actual old-vector error rose 15.26% against its unchanged baseline. The separate admission audit motivates higher-resolution perception, but no 128-pixel neural model was tested. All 25 validity checks and 406 independent artifact checks passed. Neither candidate was installed; earlier lives retain their original models.

The admission reference is a declared 256-pixel crop, not native truth or an information ceiling. The matched readout study uses one fixed refined encoder, equal decoder parameter counts and a finite mixed-input training recipe; its pretrained initialization favors phase inputs. Gains do not establish general acuity, reading, acquired meaning or compositional usefulness. Magnitude corruption is an out-of-distribution dependence diagnostic. No negative result establishes a VSA ceiling. Earlier lives retain their original models; a magnitude checkpoint requires its declared readout mode and training-derived scale.

Photo fidelity improved modestly: COCO MSE 6.19% lower (95% interval 6.05–6.34%), Flickr 5.83% lower; edge error also improved. Writing on white regressed 12.12%, with every example worse in every seed. Actual old-vector readback regressed 4.96–6.45% overall and 8.25–11.45% for Dune. All 39 integrity checks passed, but the candidate is unsuitable as an unconditional general-decoder replacement. Earlier lives retain their original models. Independent audit passed 360/360 checks; all 854 repository tests and required static/viewer checks passed.

One fixed seed-29 refined encoder, one 64 by 64 PatchSum architecture and one registered decoder-training recipe. Decoder seeds do not repeat encoder training. Pixel fidelity, edge detail and actual stored-memory readback receive separate conclusions; improved reconstruction does not establish acquired understanding, reading or compositional usefulness. Prior historical exposure and all semantic duplicates cannot be ruled out. Encoder-space drift remains a separate adoption question. No existing life is migrated or replaced.

Encoder-only training reduced immediate reconstruction MSE by 47.64% on 1,024 final COCO photographs (95% interval 46.90–48.38%), 45.38% on 1,000 Flickr photographs and 57.50% on the prior 72 inputs. Three seeds agree; all 15 integrity checks pass. All decoder, place-code and normalization-state tensors remain exact. Fine detail remains blurred, and new vectors drift substantially relative to old memory vectors.

This establishes a reconstruction gain for one pretrained 64×64 candidate and training recipe, not optimal encoding or general cognition. Every evaluated image improved in pixel MSE for each seed, but detail is still lost. New vectors match their own old vectors in only 34–36 of 72 replay cases (22/24 photographs); reconstruction alone does not establish memory compatibility or useful semantic/compositional structure. Existing lives retain their original models. Source exposure and near-duplicate screening limits are disclosed in the report.

Interrupted before final evaluation. A separate registered attempt repeats the unchanged protocol.

One pretrained 64 by 64 PatchSum candidate and one registered training recipe. Reconstruction improvement does not establish a perfect encoder, acquired understanding or compositional usefulness. External transfer, writing retention and vector drift are reported separately; prior historical exposure and near-duplicates cannot be ruled out by exact hashing. No existing life is migrated or replaced.

All registered mixed-media, pause, graceful interruption, source-free readback and continuation checks passed.

One frozen 64×64 sensory candidate, twelve added COCO validation photographs and two sampled trailers. Prior model evaluation used the validation split. Reconstruction concerns admitted crops; frame selection omits most video. Explicit episode readback does not establish partial-cue recall, cleanup, learning or understanding.

All registered continuity, isolation and retained-content control checks passed for this run.

One frozen pretrained checkpoint and twelve COCO validation photos, newly encountered by this life but previously eligible for model evaluation. Scores concern the admitted center-cropped 64 × 64 RGB image. Episode ordinals are administrative readback. This does not establish partial-cue retrieval, acquired cleanup, conceptual composition, scene editing, perceptual understanding or broad generalization.