01 / static
Examples are a starting point.
AI agents do not only need examples. Static records can show what an answer looks like while hiding what an action did.
STATIC DATA
05 / a point of view
About Windway Data
Static datasets show examples. Agents need environments where actions change state. A correct answer in a transcript can hide an incorrect result in the world.
Failure must be observable. Recovery must be testable. Ground truth must be reproducible. Evaluation must survive reruns. Windway Data exists to build that infrastructure.
We work at the boundary between environment design, data engineering and evaluation. The question is not simply how much data can be produced, but what can be demonstrated about each record.
We define the state, the permitted actions and the success condition before generating examples. An episode earns its place through schema, reset, replay, verifier and provenance checks. A failed check stays visible instead of becoming a success story.
01 / static
AI agents do not only need examples. Static records can show what an answer looks like while hiding what an action did.
STATIC DATA
02 / stateful
A stateful environment exposes the consequences of an action. Data should capture those changes and the constraints around them.
STATEFUL ENVIRONMENTS
03 / failure
The interruption is part of the record. Keeping its state visible makes recovery answerable.
FAILURE
04 / recovery
An explanation of recovery is not the recovery itself. The next actions must restore progress in the environment.
RECOVERY
05 / verification
Reproducible states, verifiers and provenance make evaluation something a team can inspect. Windway Data exists to build that infrastructure.
VERIFICATION
A checkable outcome, supported by evidence.
The trajectory can be executed again.
A controlled initial state, dependencies and provenance.
Interfaces and acceptance criteria scoped to the task.
Client-specific data stays inside an agreed boundary.