About this work

What this is

Posthumous applies a machine-learning emotion classifier, trained on one small university EEG study, to anonymous recordings from four open datasets. Its predictions become a field of marks and one letter per participant. The work is honest only when it admits that every prediction is a guess.

Letters are generated AI drafts. Every one has now been read by a human, and the 22 carrying an automated finding are not cleared for publication as written. They are not participant testimony and are not publication-ready.

How to read the field

The field is a landscape of 265people. It is one image of one thing: the shape of the model's own uncertainty.

Where a person stands comes from the affinity graph. Cosine similarity between two nine-dimensional prediction distributions decides who is drawn near whom, and the field is a stress-majorised embedding of the distances between people implied by those affinities. People the classifier read similarly end up as neighbours. Nobody is placed by hand and nobody is placed at random — the layout is computed once when the site is built, so the same person is always in the same place.

How high the ground rises comes from confidence. Mean confidence across this corpus runs from 0.196 to 0.826, but the median is 0.255 and 245 of 265people sit below 0.35 — drawn to scale, the whole archive would be a flat plain with eight spikes. So elevation uses a deliberately nonlinear map: six parts a person's rank among all 265 confidences, four parts the raw value flattened by a 0.35 power, the sum raised to 1.5. That stretches the spacing between people the model barely distinguished. It never reorders them: more confident is always higher, and the numbers themselves are printed unaltered on every letter page.

A single person raises only a narrow peak. The massif in the middle of the field is not one confident participant — it is the crowd of people the classifier lumped together, their small elevations piled on top of each other. The sharp outlying spurs are the handful it was most certain about, and certainty here means an unusual prediction, which is exactly what pushes someone to the edge of the affinity graph. The most confident readings are the loneliest ones.

Density is the same claim again: the ground is dithered, so a confident region prints as a near-solid surface and an uncertain one breaks up into scattered marks. Affinity lines between neighbours appear only when you come close to them. Participants are almost indistinguishable from the terrain until the cursor finds one, which is the honest relation between a person and a dataset.

Method

The classifier is balanced logistic regression over 70 differential-entropy features: five frequency bands across fourteen channels, extracted from four-second epochs. It was trained on the FACED dataset, then projected onto the counters, imaginers, lonely and listeners cohorts. It scores 0.198 accuracy against a nine-class chance baseline of 0.111.

Cosine similarity between nine-dimensional prediction distributions decides which strangers are drawn together. Pairs at or above 0.85 are eligible, and each person nominates at most five algorithmic kin. The full model confession records the details and the limitations.

Why it is drawn at one bit

The site has no colour. Density stands for confidence, texture for the dataset a person was filed under, and the shape of each mark for the emotion the classifier chose. A palette that gave every cohort its own confident hue would claim the model can tell these people apart. It cannot, and neither can this archive.

Seven commitments

1. Every source dataset is cited on its cohort and subject pages.
2. No attempt is made to re-identify an anonymous participant.
3. No biography is invented beyond verified source metadata.
4. Claims about feeling remain hedged because the model may be wrong.
5. The model card and its limitations remain readable.
6. The lonely cohort receives a distinct second editorial reading.
7. Subject identifiers are never turned into merchandise.

Claim, correct, or leave

If you believe you are a participant, or are authorised to act for one, you may request withdrawal, redaction, or annotation. Write to hello@aadrikas.space and include the relevant subject ID if you have it. Every takedown is honoured before publication.

This private preview has no public publication status. Requests are handled through the project's private review workflow.