We tested whether spatial patterns of repeat changes differ between 28 speech/song sounds and 128 other sounds. Separate public human studies supplied SEEG high-gamma amplitude and BOLD fMRI estimates. We canceled stable sound means by subtraction, removed change size and global sign, and excluded marginal-site products. An established two-sample statistic used only products from different sounds. Development used two patients and four adults. A local plan and code lock preceded tests in six other patients, 16 other B2021 adults and eight NH2015 adults. Mean scores were 0.00589, 0.02026 and 0.01577. Their 95% person Student intervals were [-0.00125,0.01303], [0.00456,0.03597] and [0.00921,0.02233]. The fixed shared rule failed. A positive result across modalities was not confirmed. The shared p was 0.34375. Removing the mean repeat change left all three person intervals across zero. All negative people and controls remain. One scalar gain per repeat, common to both classes, gives an exact score of 2 in a model. Fixed mixing also converts independent site variance changes into a positive score. These examples do not identify the human cause. Source selection, shared schedules, separate cohorts, sparse electrical coverage and unequal measurement filters limit scope. The contribution gives a cross-modal benchmark, code, source pins and terms, all results, figures, arithmetic and source checks, and repeat instructions. Original neural arrays remain external. No common neural cause or causal effect was established.
We tested whether a positive cortical height order of band median frequency transfers between human SEEG and fMRI from separate studies. SEEG used 1-45 Hz and PAFIN used 0.01-0.08 Hz. We used within-person partial correlations after linear position controls, with spacing added for SEEG. In the original three-test plan, FR1 SEEG gave mean r +0.2543 in 73 patients (95% person interval +0.1734 to +0.3326). catFR1 was uncertain in ten patients, and ten PAFIN movie participants gave +0.0094 (-0.0560 to +0.0762). The shared-positive rule failed. A separate plan then fixed twenty unrelated HCP resting records before their spectral outcomes were read. Their mean r was -0.1907 at 0.01-0.08 Hz (-0.2451 to -0.1353) and -0.5456 at 0.01-0.50 Hz (-0.5839 to -0.5048). Individual MNI centroids, volume adjustment, and both time halves retained negative means. A source alternate extraction, added before HCP outcomes, also gave negative orders in both bands. It moved the wide-band mean toward zero by +0.3307 (+0.2845 to +0.3768). Its stated equal-size masks were not independently checked. A correction across all seven direction tests kept FR1 and the four HCP tests below 0.000058. These signals showed opposite orders in the stated cohorts and models. The result does not identify a causal modality effect or establish a shared neural time constant. No new untouched cohort confirmed the reversal as a new hypothesis. We preserve all failures, code, compact inputs, source hashes, person tables, figures, plans, and repeat instructions. Independent calculations agreed. A full repeat matched all twenty output files.
We tested whether acoustic maps learned from environmental, mechanical and music sounds predict response differences among animal sounds. Separate human studies supplied signed iEEG LFP means and BOLD fMRI. Four people per primary study supplied development. A local code lock preceded tests in 20 other patients and 16 other adults. No Animal response fit a new scale, coefficient or penalty. The target panels had 17 electrical sounds and 15 MRI sounds. Mean centered error gains were 2.207% and 8.779%. Their 95% person Student intervals were [0.365,4.049]% and [7.060,10.498]%. The shared test passed after two-claim Holm correction, p=0.0214052. A fixed benefit test for 20 pitch and periodicity features failed, p=1. Eight adults from another MRI acquisition gave mean gain 9.548%, using the same MRI sounds. Resampling people and Animal sounds included zero in iEEG. A later coefficient resampling range also included zero. Shorter electrical windows, two training-class omissions and five-site MRI panels gave uncertain gains. A pre-event control gave mean -0.474%; the primary-minus-control mean was 2.681 points. All negative and zero scores remain. The result supports a prediction rule for fixed source panels, not broad transfer to new species or a common neural cause. Source selection, separate cohorts and corpora, unequal measurement filters and finite sounds limit scope. The package gives code, permitted compact inputs, source pins and terms, all results, figures and repeat steps. Fresh source chunks exactly matched 110,682 electrical means; a fresh acoustic rebuild matched all 14,520 MRI feature values.
We tested whether sound class information remains in spatial response distributions after site labels, spatial mean and positive gain are removed. Separate public studies supplied human SEEG high-gamma activity and fMRI BOLD estimates. A fixed reader used 19 normalized spatial quantiles. It fit one sound half in one repeat and predicted other sounds in another repeat. Two patients and four MRI adults supplied development. A local code lock preceded tests in six other patients and 16 other adults. Mean speech/song AUC was 0.74146 and 0.83643. Their 95% person Student intervals were [0.61578, 0.86714] and [0.78227, 0.89059]. All 22 people were above chance. The single shared rule required person mean and common sound-label tests in each modality. It passed, p=0.00216355. Fixed 24-site checks and eight further MRI records kept positive means. A later source-reader transfer check had intervals that included chance in both primary cross-modal directions. A fixed model with one signal direction gave perfect reading through class-dependent signal-to-noise ratio, but chance without noise. Thus measured gain removal does not identify an amplitude-independent neural code. The result supports separately readable normalized distributions in each selected measurement. It does not show universal weights, shared neural sources, semantic abstraction or causality. Source screening, sparse clinical coverage, cohort differences, a fixed sound library and unequal signal filters limit scope. Code, source pins and terms, all results, figures, audits and repeat instructions are supplied. Original neural arrays remain external.
We tested whether removing one fitted sound mode reduces transfer from speech and song to nonspeech human voices. Separate public studies supplied high-gamma SEEG and BOLD fMRI. Training and test clip identities were disjoint within each fitted model. Development used two patients and four fMRI adults. We locked the code before scoring six other patients, 16 B2021 adults and eight NH2015 adults. Mean AUC changes were -0.01352, +0.01449 and +0.00015. Their 95% person Student intervals were [-0.04413,+0.01708], [+0.00099,+0.02799] and [-0.01582,+0.01613]. The fixed shared cost test failed, p=0.98146. Ranges from joint person and test-sound resampling included zero in every study. All negative people and controls remain. An exact construction kept all training and other test patterns fixed. With unrestricted target patterns, it gave AUC changes of -1 and +1. Both retained and removed parts of the reader must be nonzero. Thus, a training variance fraction alone cannot bound loss of this readout. This algebra limit is not a model of human physiology. We provide checked code, source pins and terms, all scores, figures and repeat steps. Small fixed sound panels, source selection, clinical sampling, shared schedules and unequal measurement filters limit scope. No positive cross-modal cost or common neural cause was confirmed.
We tested whether a response direction learned from human voice sounds can classify animal sounds, and whether transfer works in reverse. Separate public human studies supplied signed iEEG local field potential means and BOLD fMRI estimates. Four people per modality supplied development. A local code lock preceded tests in 16 other patients and 16 other adults. Electrical responses used a duration-matched pre-onset subtraction. Training and test clips were disjoint. Mean transfer AUCs were 0.5529 and 0.5707 in iEEG, and 0.7239 and 0.8752 in fMRI. All four pointwise 95% person Student intervals were above 0.5. The fixed shared rule required a positive mean, a majority of people, and a label reference. It failed, Holm p=0.68175. Electrical positive counts were 11/16 and 10/16. The shared advantage over a fixed acoustic reader also failed. We added a third test after development. It compared training families with equal example counts and the same test sounds. Human-target mean gains were 0.0813 and 0.0281; animal-target gains were 0.0103 and -0.0057. Its shared rule failed, Holm p=1. All three claims remain. Eight records from another MRI acquisition gave positive transfer; participant overlap between MRI studies was not verified. Joint person-and-sound ranges included chance in both electrical directions and stayed above chance in MRI. Whole-sound refits, negative people, controls and source checks are preserved. Source labels include nonvocal animal clips, breathing and whistling. Different cohorts, sparse clinical sampling, trial means, source filters, prior site screens and unequal sound libraries limit scope. The contribution is a reproducible transfer audit and a negative shared decision. It does not show absent neural information, common semantic coding or causality.
We tested sound transfer in separate public human iEEG and BOLD fMRI studies. Each reader learned from speech and other training clips. All nonspeech human voices, animal sounds and other test clips were absent from fitting. Four people per study supplied development. A local code lock preceded tests in 20 other patients and 16 other MRI adults. Mean animal-over-other AUC was 0.48393 and 0.80680. Their 95% person Student intervals were [0.43866,0.52920] and [0.78002,0.83358]. The shared positive test failed, Holm p=0.76849. Human-over-animal AUC was 0.70401 and 0.60817, with intervals [0.66027,0.74776] and [0.57683,0.63951]. This fixed-panel shared test, which also required human-over-other transfer, passed Holm correction, p=0.00008965. All reserved people had human-over-animal AUC above 0.5. Yet common test-clip resampling gave MRI range [0.42452,0.78125], and a common test-label reference gave p=0.1281. Broader human-over-animal transfer was not established. A fixed exchange of the electrical other-sound halves raised animal AUC to 0.55447. That supporting result did not replace the main failure. We retain all negative results and controls. A source audit corrected our initial electrical-window description. The input values did not change. The contribution is a direct animal transfer test and a checked benchmark, not the first report of human voice preference. Separate cohorts, unequal sound classes and sampling, source selection and measurement filters limit scope. No common neural cause or causal effect is established. The package preserves source pins and terms, code, licensed electrical inputs, scores, uncertainty draws, figures and repeat steps. Original MRI arrays stay external.
We tested whether larger auditory responses mark more repeat variation shared between measured sites. Separate public studies supplied human SEEG high-gamma activity and BOLD fMRI. At 24 selected sites, one repeat measured mean response and two other repeats measured disagreement. The score removed within-site quadratic terms and broad sound-group means. Separate sounds set site RMS scales. Development used two patients and four MRI adults. A local code lock preceded six other patients, 16 other B2021 adults and eight NH2015 records. Mean scores were +0.004427 [-0.001895,+0.010749], +0.000061 [-0.000055,+0.000177] and -0.000377 [-0.001904,+0.001150]. Brackets give 95% person Student limits. The declared three-study positive rule failed; its largest required sign-flip reference was 0.71875. Crossed person/sound ranges also included zero. Site-count, baseline and repeat checks were discordant. Planned gain rescaling was unstable. An added pair-gain diagnostic did not change the failed result. Exact models gave positive shared scores from fixed spatial mixing of independent latent errors and from deterministic repeat gains with no random error. The contribution is a checked off-diagonal repeat assay and a negative human benchmark. It does not identify intrinsic neural noise, a common generator or causality. Cohort differences, source selection, shared normalization and measurement filters limit scope. Code, source versions and terms, all results, figures and repeat steps are supplied. Original neural arrays remain external.
We tested whether image classes have different covariances across brain sites after class means, equal-loading common signals and independent site variances are removed. Separate public studies supplied human iEEG and fMRI. Three patients and two MRI adults supplied development. A local code lock preceded tests in 14 other patients and six other adults. The assay uses products from four distinct images and a separate scale panel. Mean contrasts were 0.009871 and 0.001755. Pointwise 95% person Student limits were [-0.025237, +0.044979] and [-0.000917, +0.004427]. Scores were positive in 8/14 and 5/6 people. The fixed shared mean-and-majority rule failed, p=0.39526367. No positive cross-modal neural claim was confirmed. An exploratory audit found larger ordinary squared-covariance values in 14/14 and 6/6 people. The tool removes finite-sample square bias under independent-image assumptions. An exact model shows that class-dependent correlated error can still give a positive target with equal neural covariances. Source selection, clinical sampling, different cohorts, image groups and signal filters limit scope. This is a new endpoint in reused studies. The package has code, source pins and terms, all results, figures, local plans, numerical checks and repeat instructions. Original neural arrays remain external.
We tested whether eight auditory sites predict relative sound responses at eight disjoint sites. Separate public human studies supplied SEEG high-gamma amplitude and BOLD fMRI. All rules removed each site group's response level and fitting-sound class means. Sound folds excluded test sounds. Two patients and four MRI adults supplied development. A local code lock preceded tests in six other patients and 16 other MRI adults. At a fixed ridge penalty, mean gains with inputs and outcomes from different repeats were -3.76% and -5.04%. Mean gains within one repeat were +10.12% and +13.50%. The paired increases were 13.87 [7.94, 19.81] and 18.54 [16.13, 20.95] percentage points. Brackets are pointwise 95% person t limits. The fixed gap claim passed Holm correction for two shared claims, p=0.03125. The original positive different-repeat claim failed. An unchanged check in eight adults from another MRI study gave a 17.64 [15.70, 19.57] point gap; its gated test passed. An exact identity shows that the gap measures alignment of predicted and observed repeat differences, with stable sound means canceled. A repeat-noise model gave positive within-repeat gain and negative different-repeat gain. This model does not identify the human cause. Stronger ridge shrinkage and other controls remain, including discordant results. Clinical sampling, source selection and processing, nearby sites, fixed sounds, separate cohorts and different readouts limit scope. The contribution is a validation benchmark, an exact loss audit, code, source pins, all outcomes, figures and repeat instructions. No common neural cause or universal failure of site prediction is established.
We tested whether nearby training sounds lower neural response error more than equally sized far or random sets with the same class counts. Separate public studies supplied signed human iEEG and BOLD fMRI. Each target response was excluded. All rules shared class calibration from other sounds. Five patients and four MRI adults supplied development. A local code lock preceded tests in 20 other patients and 16 other adults. Near-versus-far mean gains were 0.562% and 1.855% of category-baseline error. Their 95% person intervals were [-0.102%,1.226%] and [1.180%,2.529%]. Positive counts were 12/20 and 15/16. Near-versus-random gains were 0.237% and 2.242%, with intervals [-0.283%,0.758%] and [1.775%,2.710%]. Both shared claims failed Holm correction, p=0.503445. Near skill was positive in each modality. A later unchanged test in eight adults from a separate MRI study gave 0.126% near-versus-far gain, interval [-0.582%,0.833%], and 1.836% near-versus-random gain, interval [0.868%,2.803%]. Thus the MRI random-set comparison repeated, but the distance order did not. Separate B2021 estimates also had uncertain near-versus-far gains. All negative outcomes remain. A noiseless one-scalar linear response example gave a 69.93% Gaussian-model support gain at penalty one. This limits nonlinear neural claims. Independent fits checked 31,872 person-target-rule predictions. Fresh source reads rebuilt all electrical means exactly. The package supplies code, permitted inputs, provenance, source terms, plans, uncertainty, figures and repeat steps. Clinical sampling, source selection, different cohorts, feature sets, sound panels and measurement filters limit scope. No positive cross-modal support advantage or common neural cause was confirmed.
We tested if repeat changes distinguish source speech/song sounds from source nonvoice sounds. We removed total change size and sign. Separate public human studies supplied SEEG high-gamma activity and BOLD fMRI. Development used two patients and four MRI adults. A local code lock preceded tests in six other patients and 16 other adults. Ten fixed 16-site panels and five whole-sound folds gave one AUC per person. Mean AUCs were 0.53426 and 0.53728. Their 95% person Student limits were [0.46386, 0.60466] and [0.51726, 0.55731]. The shared Student, sign-flip, label-reference and interval rule failed, p=0.15625. Eight further MRI adults gave 0.51643 [0.49341, 0.53946]. All negative people and controls remain. Mean-response controls gave AUC near 0.95-0.97. They do not repair the failed test. An exact identity shows that scalar repeat gains can preserve stable response shape in normalized differences with no additive noise. A generated example gave AUC 0.99040; it was not fit to human outcomes. One shared clinical schedule also predicted labels. This is not a neural effect. The contribution is a fixed benchmark, a checked interpretation limit and a complete result record. Source selection, clinical coverage, acoustics, common schedules, separate cohorts and unequal measurement filters limit scope. No positive cross-modal neural claim or cause was established. The package gives code, source versions and terms, all predictions and results, figures and repeat steps. Original neural arrays remain external.
We tested how variable source activity changes the relative variance effect of shared release at multi-target chemical sites. We used eight public developmental worm maps and a fixed 162-neuron panel. Both models had independent binary source gates. One model shared release across targets of a site; the comparator released each contact independently given its source gate. With source and release probabilities both 0.5, the variance ratio for summed inputs was 1.029 to 1.161. With all sources active it was 2.069 to 2.368. The two adult maps gave ratios of 1.031 and 1.029 with source variation. Their release probabilities had to be at most 0.225 and 0.214 for at least 10 percent excess at source probability 0.5. Target count normalization gave adult ratios of about 1.040. An exact covariance decomposition explains the reduction: source variation adds variance to both models. It does not remove the absolute site effect. All 16 simulated variances were within 0.72 percent of exact values in 100,000 trials per map and model. Unknown release laws, abstract binary source states, positive gains, partial maps and lack of matched activity limit the result. Code, proof, aggregate evidence, figure, hashes and external retrieval instructions support reproduction. Raw external records are excluded.
We tested whether nine nonspeech human voices require more than two stable measured response directions. Separate public studies supplied eight SEEG patients, 20 B2021 fMRI adults and eight NH2015 fMRI adults. A local plan preceded the new third-order contrast score. Its mean was -0.04628 in SEEG, +0.00402 in B2021 and +0.01328 in NH2015. The 95% person Student intervals were [-0.20312,0.11055], [0.00049,0.00755] and [0.00119,0.02538]. The fixed shared test failed, p=0.74609. All negative people and failed controls remain. An exact rank-one model gave expected score 64/189 despite independent Gaussian repeats. Correlation among sound and site errors caused the bias. We then applied a Gaussian polynomial correction to the same volume target. It needs independent test repeats with equal Gaussian error covariance, but allows arbitrary sound and site covariance inside each repeat. It reduced the model expectation to zero. Exact non-Gaussian, unequal-covariance and shared-error models show its limits. The later human correction was exploratory. Every cohort mean interval included zero. We give an equivalent real correction and verify the existing sound-kernel form with exact integer coefficient checks and independent source audits. All new numerical result files matched a complete replay. Source selection and processing can violate both error models. Different cohorts, clinical coverage, short baselines, fixed sounds and small samples limit inference. No shared third neural dimension, common axes or cause is established. The package gives code, source pins and terms, all scores, figures, plans and repeat instructions. Original neural arrays stay external.
We tested whether repeat reliability on one sound set marks sites useful for reading other sounds. Separate public studies supplied human SEEG high-gamma responses and BOLD fMRI estimates. Two sound folds kept fitting sounds out of each test. We removed half the sites with the greatest fitting reliability. We compared the retained sites with 32 random sets at the same site count. Each response lost its spatial mean, size and global sign. Development used two patients and four MRI adults. Before reserved tests, we added labelled accuracy loss to two original site-label benefit tests. A local code lock preceded tests in six patients, 16 B2021 adult records and eight NH2015 adult records. Added accuracy losses were SEEG +19.81 points; B2021 +12.25 points; NH2015 +8.22 points. The accuracy-loss shared test passed, with Holm p=0.046875 over three claims. H1 and H2 did not pass their shared tests. All records, negative outcomes, code, figures and uncertainty draws remain. Original neural arrays stay external. SEEG pre-onset loss was +4.08 points; baseline subtraction left +12.66 points. Thus the result does not isolate sound-evoked coding. Source screens used the full sound panel before release. Separate cohorts, possible overlap between MRI panels, clinical sampling, shared schedules and signal mixing limit scope. The tests concern fixed processed-signal readers. They do not establish neural necessity, source count or causality.
We tested whether a fixed source group predicts sound relations in other people. Separate public studies supplied human SEEG high-gamma responses and BOLD fMRI. Each study used 128 source-defined nonvoice sounds. Repeat 1 set site scales. Repeats 2 and 3 gave crossvalidated squared distances. We removed every additive sound term from the pair map. Two source patients predicted six other patients. Four B2021 source adults predicted 16 other adults and eight NH2015 adults from a separate study. Mean cosine scores were 0.04205 [0.02815,0.05595], 0.84276 [0.81281,0.87271] and 0.41819 [0.36097,0.47541]. Brackets give pointwise 95% person Student limits. All 30 target scores were positive. The one fixed shared test passed, p=0.015625. It required person-sign and common sound-label tests in every group. A local plan preceded new scores. One native MRI ID assertion was corrected after the first SEEG result and before any MRI score. The SEEG replay was exact. Fixed 24-site and spatial-mean checks kept positive mean estimates. Added exploratory MRI repeat-difference maps had small shared structure; this result remains. An exact one-scalar model gave perfect map transfer. Thus the human finding concerns common measured pair relations on fixed panels. It does not identify several neural sources, a non-acoustic code, or a common cause. There is no SEEG-to-fMRI sound or site join. Sparse clinical coverage, source selection, short electrical baselines and different cohorts and measurement filters limit scope. This extends earlier SCI pair-map work. Code, source pins and terms, all scores, uncertainty, figures and repeat steps are supplied. Original neural arrays remain external.
We tested whether acoustic distance from fitting sounds warns of large neural model errors. Separate public studies supplied human intracranial EEG and BOLD fMRI. Four people per study supplied development. A local code lock preceded tests in 20 other patients and 16 other adults. The fixed reader used nearest acoustic distance after feature clipping. It ranked errors after control for sound class, fold and predicted response size. Mean rank associations were +0.0054 [-0.0369, +0.0477] and +0.1134 [+0.0750, +0.1518]. Brackets give pointwise 95% person Student limits. A gate removed high-distance sounds within class and fold. Mean gains over matched random removal were -3.2362 [-6.2926, -0.1798] and +2.7839 [+2.2578, +3.3100] percent of whole-panel error. Both shared tests failed; Holm P values were 0.792091 and 0.980458. Eight adults from another MRI acquisition had a positive mean adjusted association, but uncertain gate gain. One separate B2021 estimate had an association interval across zero. A three-voxel check also left gate gain uncertain. An exact rank split showed that between-class terms reversed the raw MRI association. A later transform swap showed that clipping changed electrical gate results. These checks do not establish a modality effect or a cause. The package gives a fixed error warning benchmark, permitted inputs, source pins and terms, all results, code, figures and repeat steps. Fixed sounds, incomplete acoustic features, clinical coverage, source screening, noise and separate cohorts limit scope. No positive cross-modal error warning was confirmed.
We tested whether annual nitrate plus nitrite load became more sensitive to flow after 2000 in the Mississippi and Atchafalaya rivers. We used USGS version 3.0 load estimates and complete annual flow records for 1980-2025. We fitted changes in log combined load against changes in log combined flow, with separate slopes for transitions ending in 1981-2000 and 2001-2025. Slopes were 1.021 and 1.037. The late minus early slope was 0.016, with conditional 95% three-transition block limits of -0.291 to +0.502. All nine planned checks also had limits that included zero. The data give no clear increase, but do not establish equal slopes. Load estimates use discharge as an input, so the relation is partly built into the source method. Annual averaging, two connected branches, short periods, and absent joint source-error draws limit physical inference. Code, small aggregate inputs, all fitted checks, bootstrap draws, a figure, source hashes, retrieval steps and an independent arithmetic audit are included.
We tested removal of one repeat-variation mode from human sound responses. Separate public studies supplied SEEG high-gamma activity and two BOLD fMRI panels. Each person supplied 24 sites and three response repeats. Different sounds set the mode and measured its error. Class means and spatial means were removed. Development used two patients and four MRI adults. A local code lock preceded tests in six other patients, 16 B2021 adults and eight NH2015 adults. Mean error reductions were 8.91, 12.12 and 11.13 points of raw test response variance. Added reductions over a uniform random direction were 6.50, 9.83 and 8.69 points. All 30 people had both gains. The shared rule passed Holm correction over three claims, p=0.046875. The original claims of excess signal loss and harmful removal failed. Sound and person refits assess panel sensitivity. A later exploratory check found lower error from a zero residual pattern in every person. A pre-onset SEEG check also gave a removal benefit. Thus the fixed gain does not show useful sound decoding or a sound-specific neural process. Independent-source models can give signal loss and either error benefit or harm after common measurement mixing. Separate cohorts, source selection, three repeats, unequal filters and fixed sounds limit scope. Code, source pins and terms, all failures, figures and repeat steps are supplied. Original neural arrays remain external. No common neural cause or causal effect was established.
We tested whether directions fitted to nonvoice sounds improve speech and song response estimates at a fixed sum of filter weights. Separate public studies supplied human SEEG and two fMRI cohorts. Development used two patients and four adults. A local code lock preceded tests in six other patients and 24 other adults. Every fitted choice used 128 source sounds. The target had 28 speech and song sounds. Mean gains over uniform shrinkage were +0.3446%, -0.0785% and +0.2898% of measured target energy. Their 95% person Student intervals were [-0.4284,+1.1177]%, [-0.3781,+0.2211]% and [+0.0088,+0.5707]%. Positive counts were 3/6, 6/16 and 6/8. The fixed shared rule failed, p=0.89494324. All 30 filter skills over zero were positive. All planned controls and negative outcomes remain. An exact conditional model had perfect source moments and shared signal axes. Four axes captured 42.1% of target signal, above a 26.7% reference. The source-calibrated filter still increased target loss. This model was not fitted to people. The contribution is a controlled benchmark and an interpretation limit. No common neural cause or positive cross-modal error gain is established. Source selection, shared schedules, fixed sounds, sparse clinical placement and unequal measurement filters limit scope. The package preserves code, source pins and terms, plans, every result, figures and repeat steps. Original neural arrays remain external.