Agent behavior
Evaluation design that separates a model response from the surrounding workflow.
For AI labs
Bring one question that needs real workflow evidence.
Some apparent model failures begin in human specification or organizational design. Some apparent workflow failures are agent failures. A useful collaboration makes those objects separable and keeps publication status visible.
One observed failure
Where did the result actually break?
The research design must distinguish at least three possible sources before assigning causality.
Human specification
Observation of intent, correction, review, and decision ownership.
Operating model
Evidence from handoffs, shared context, exceptions, and release rules.
Collaboration register
Define the contribution before promising the paper.
Working specification
Research question
A falsifiable question tied to alignment, reliability, or organizational learning.
Proposed jointlyArtifact
A named dataset, evaluation, workflow record, or method with an attributable source.
Scoped before accessContribution
Methods, analysis, models, domain access, or workflow evidence each party will provide.
Named by partyPublication
Authorship, review status, disclosure, embargo, and non-public material boundaries.
Agreed in advancePublication discipline
Three rules keep the collaboration inspectable.
- 01
Separate the objects
Model evidence, workflow evidence, and product evidence do not borrow status from one another.
- 02
Preserve uncertainty
A result keeps its sample, method, publication status, alternative explanations, and limit.
- 03
Earn authorship
Authorship follows a defined research contribution; it is not promised by an inquiry form.
Research inquiry
Send the question, available artifact, and required boundary.
We can then determine whether there is a joint method and a contribution worth formalizing.