Readers trained on word generation did not pass a shared memory load test in EEG and fNIRS from 20 reserved volunteers

We tested whether personal readers trained on word generation give larger scores at higher n-back load. One public study supplied simultaneous EEG and fNIRS from 26 volunteers. Six people were development; a local code lock preceded new transfer scores in 20 other people. Each device had separate personal weights. N-back labels did not train those weights. Mean adjusted changes from 0-back to higher load were 0.637 and 0.531 in training score units. Their 95% person t intervals were [0.260,1.015] and [0.037,1.025]. Positive counts were 15/20 and 14/20. The fixed shared rule for means and positive counts failed, Holm p=0.115318. The further rise from 2-back to 3-back also failed, Holm p=0.979305. Mean transfer AUCs were 0.614 and 0.645, with pointwise intervals above chance. These descriptive scores do not replace the failed rule. The optical mean interval included zero after an eye adjustment fitted on word trials. All fixed checks, negative people and discordant outcomes remain. An unfitted two-feature Gaussian example gives negative transfer despite a shared mean component. It is an interpretation limit, not a human explanation. Separate calculations and an exact offline repeat passed. The package gives 104 attributed compact feature files, source hashes and terms, all models and scores, paired-person draws, code, figures and repeat steps. Original continuous arrays remain external. This is one paired cohort, not independent-study confirmation. Task order, personal calibration, sensor mixtures, blood effects and source processing limit neural claims. No common cause or general neural demand scale is established.

EEG · fNIRS · working memory · word generation · cross-task transfer · human neuroscience · negative result