courier
self-report, gathered by the agent a participant already talks to.
[what it is]
courier is being built to collect self-report for research studies through the ai agent a participant already uses, instead of a separate study app.
the questionnaire becomes a blueprint. it still decides what gets measured. the agent is meant to gather it a little at a time: one small question, or one short open question, at a fitting moment in a normal conversation.
a few validated items are still meant to be asked word for word, rarely and briefly, on a schedule courier sets. they are there to tie what the agent gathers back to a known scale.
[why not just a questionnaire]
short asks hold up.
long forms wear people down. in earlier phone studies, single quick taps kept people answering for longer, with less felt burden. asking inside a conversation has not been tested that way yet.
no separate app.
the study is meant to live where the participant already is, not in an app they open only for you.
their own words, and a rating.
each moment can hold both. in earlier research, text written in answer to a question matched people's own ratings better than text written for something else.
the moment, not the reminder.
the agent can notice when someone just mentioned a hard day. courier is designed to ask then, and also at random times, so the sample is not only the easy moments. whether that works is untested.
built to stay comparable.
the anchor items are meant to tie different agents back to one scale. the first study will test whether they do.
[how it compares with a standard form is what the first study will test. courier will be checked against one before it claims anything.]
[how it is designed]
- courier is designed to log each ask: what was asked and when.
- courier, never the agent, is meant to assign some timing and wording at random, so their effect can be measured.
- it is meant to record the model behind each answer where the agent reports it, so a model change can be checked.
- scoring is planned for a separate, fixed model, never the agent that asked.
- reacting to answers is planned as opt-in and off by default, and logged from the first day it is on.
- an answer that points to a crisis is designed to always go to the safety path. that part is never randomized.
[who it is for]
research labs that collect self-report: mood and well-being, personality, attitudes, momentary experience. any psychological self-report built on a validated questionnaire.
studies that run over days or weeks, where keeping people answering is the hard part, and labs that need to know what each participant was asked, and when.
[not a screening tool, not a diagnosis, not a treatment. courier is a research instrument.]
[faq]
do participants ever see a questionnaire?
the plan is that they rarely do. a few validated items come up briefly, on a schedule courier sets. the rest is small questions inside the conversation.
which agents does it work with?
it is designed for personal agents that let a participant add a connector. none are tested or certified yet.
has it been tested with people?
not yet. no participant has answered through it. the first study will run it side by side with a standard form.
does the agent react to what people say?
not by default. reacting is planned as opt-in. when it is on, reactions are designed to be logged, and some held back on purpose, which participants would be told up front.
is the agent also the one scoring the answers?
no. scoring is planned for a separate, fixed model, never the agent that asked.
is it open source?
releasing courier as open source through oneirum is being looked into.
how does courier relate to oneirum?
oneirum builds care into the ai agents people already use, and is coming soon. courier is its complement, aimed at research labs.
[interest form]
running a study that could use courier? tell us a little about it.