people tell ai agents about their days,including how they are doing.
oneirum is building a new way to measure how someone is doing, through the ai agent they already use.
each line is one person over time,measured against their own usual.
a note on oneirum
people now tell an ai agent more about their days than they tell almost anyone. how they slept, what they keep putting off, what is weighing on them. none of that reaches the way anyone measures how a person is doing.
oneirum is building one scale, read from the conversations a person is already having, compared with their own usual, and anchored to the forms clinicians already trust.
the details of the method, and what it finds, stay private until our studies with people are published.
abhay mettufounder
the problem
how someone is doing is still measured with a form.
you tick boxes about the last two weeks, and that score is how depression has been tracked since 2001.
the method, in properties
a trend, not a snapshot.how someone is doing over months, not one look back at two weeks.
against their own usual.change measured from a person’s own baseline, not a population cut-off.
one scale, even when the model changes.re-anchored after every model update, to the forms clinicians already trust.
the first scales cover mood, worry, and attention and follow-through in daily life. they track trends and never put a label on a person.
uses
each of these is a use of the same method.
a chatbot that still sees someone in crisis.the first use. it checks a whole conversation and runs again whenever the model changes.
people, their own trend.and a one-page report to bring to their clinician, if they choose.
labs, their users over months.whether their users end up better or worse off.
trials, a measure between visits.
care and coaching teams, an outcome they don’t grade themselves.
how we work
it tracks trends.
how someone is doing over time. it never names a condition.
the person decides.
what is collected, and who sees it. adults only.
no data for sale.
we never sell anyone’s data, and there are no ad trackers here or anywhere else we run.
studies through a university.
work with people goes through a university’s ethics review, with the person’s consent.
independent.
we don’t make the chatbots we check, and no lab owns a piece of us or gets an exclusive.
where it stands
nowthe crisis test runs on open models. its reports reproduce exactly.
nowthe first study with people is drafted: a preregistration, a protocol for ethics review, and consent. it has not run yet.
aliquot 0.3.1, an open source research agent for labs.
nextrunning that first study, and offering the crisis test to labs as an audit.
research
abhay’s research is on when a measurement means the same thing for different people. three papers, public on zenodo.
the validity ceilingwhat passive behavioral data can and cannot measure about depression
ghost respondentsa boundary map for covert ai authorship in open-ended survey data
item wordingand the covariance of self-report items