# Detector arithmetic and a stored evidence replot

Run `python reproduce.py` with numpy and matplotlib. The bundled neutral stored-baseline.json makes the export standalone. Outputs: 2,541 detector rows, ten participant rows and one stored group-summary row; two original figures in PNG/SVG/PDF; validation, runtime and SHA256 metadata. Hashes establish file identity, not scientific validity.

## Fictional nonclinical detector

`detector(p,s,c)` takes finite fractions in [0,1]: assigned target prevalence p, sensitivity s=P(positive|target), specificity c=P(negative|no target). Expected fractions are TP=ps, FN=p(1-s), FP=(1-p)(1-c), TN=(1-p)c. Positive fraction=TP+FP and P(target|positive)=TP/(TP+FP). When no positive readings are expected, this last conditional is undefined: Python/JSON null and an empty CSV cell, never zero. Grid prevalence0..1 by0.05, sensitivity/specificity0..1 by0.1. Figure uses sensitivity0.9 and specificity0.5/0.9/0.99, separately declared illustrative slices. These are expected fractions, not observed subjects or measured performance. No diagnosis, screening recommendation or patient inputs.

Independent integer contingency counts verify posterior arithmetic on the full grid. Fractions normalize and preserve assigned target marginal. All boundaries are included; four invalid guards pass. Sensitivity and specificity remain assigned conditional-model parameters at prevalence endpoints, not estimates from absent groups.

## Stored exploratory EEG analysis: replot only

The baseline preserves ten stored participant balanced accuracies, their retained-trial counts, the stored unweighted mean and stored subject-bootstrap95% interval. Mean25.0475599293%, interval[23.5349970623%,26.4946636848%], chance reference25%. The lower-bound-above-chance criterion is not met. Stored fixed-prediction within-subject label-permutation p=0.4774522548 is not a probability that chance is true. Its10,000 label shuffles do not refit the model and do not address adjacent-trial temporal dependence. The stored bootstrap uses20,000 resamples of ten subject scores. Neither resampling procedure was rerun here.

The plotted points have no individual confidence intervals. Only the group summary has the stored subject-bootstrap interval. We independently recomputed the unweighted mean and summed retained trial counts to2,010; raw-signal preprocessing, model fitting, validation, bootstrap and permutation were not replicated. Ten people are not2,010 independent people. Dataset local metadata: OpenNeuro ds003626 DOI10.18112/openneuro.ds003626.v2.0.0; NEMAR on003626v1.0.0 derivatives. Preserve these versions rather than upgrading labels to a newer paper/repository version.

The stored analysis used28 readable sessions and128 EEG channels, four directional command classes (up, down, right and left) across inner-speech, pronounced-speech and visualized conditions, log Welch band powers4–8/8–13/13–30/30–45Hz over0.5–3.5s epochs, fixed four-class softmax and training-fold standardization with leave-one-subject-out validation. Sessions S09/B1 and S10/B3 were unreadable; condition filtering excluded the relevant S03/B1 trials. It is not30 complete sessions. Dataset acquisition-channel totals differ from the128 EEG channels used. The result did not demonstrate above-chance performance under this analysis; it does not prove all brain–computer interfaces impossible, unrestricted thought decoding or remote sensing.

Original result and analysis-script SHA256 hashes, bundled baseline hash and actual generator hash are recorded in study.json. The original stored result is preserved, not overwritten. Primary dataset reference: https://www.nature.com/articles/s41597-022-01147-2 describes the dataset, not this local classifier result. No new biological or clinical validation is claimed.