The Corpus
What the station keeps.
A Weigh-In produces a record. So does a What Just Happened read, and an Incident Report. The station holds onto those records, anonymized in aggregate and attributed to a visitor only when the visitor has explicitly opted in.
The aggregate of those records is the corpus — a long-running account of what AI is actually doing to the people interacting with it, gathered one submission at a time.
Why it accumulates
Most patterns of behavioral impact don't show up on a single interaction. They show up in the trajectory: the way six months of one shape of conversation shifts what you notice, what you question, what you trust about your own read.
A station that doesn't keep a record can only read the moment, never the trajectory. The record is what makes the longer reading possible — first for the visitor whose record it is, eventually for whoever ends up working on the wider question.
What the corpus isn't being used for
The corpus is not for sale. There is no partnership, no enterprise license, no anonymized data feed quietly being shopped to AI labs, and no plan to monetize. There is no business model that depends on the corpus reaching some critical mass.
Aggregate findings might eventually be worth publishing on their own terms. Until then, the corpus is what accumulates while the station does its primary work — making the experience legible for the visitor in front of the station today. Everything else, including the corpus itself, sits downstream of that.
The longer-horizon possibility
The corpus has standalone value as a possible contribution to AI behavioral calibration — not the calibration of the people using AI, but the calibration of the systems themselves. If the patterns the station reads, in aggregate and at scale, ever read clearly enough to inform how those systems are tuned, that possibility is one of the reasons to keep the work careful and the corpus honest.
This is posture, not roadmap. There is no partnership in motion, no publication in production, no handoff being designed. If aggregate findings ever become worth publishing, that decision is made then, under whatever circumstances apply — not before.
How records get kept
Submission inputs and scores are stored against an unenumerable submission ID, and the visitor's email gets attached only when the visitor enters one (to retrieve their report, or to be linked to the record for return reading).
What Just Happened transcripts have two storage paths. The corpus path keeps the scrubbed transcript, the extracted features, and the read; the raw paste is never kept. The minimal-delivery path keeps only the score plus an email link, and auto-purges after ninety days. The visitor picks at the moment of submission.
The full disclosure lives on the privacy page.
The corpus is a long bet on a slow problem. It pays attention while pattern-naming and visitor service do the immediate work, and it keeps the long shape of the question intact for whatever comes next.