Measurement
Measuring human–AI systems
The purpose is not to score people. It is to determine what the human–AI configuration is enabling, what the person retains, what the system is carrying, and where additional support or safeguards are needed.
Ten dimensions
What the diagnostic separates
Each dimension asks a different question about the same interaction. None of them stands in for the others, and none of them is reported as a single overall score.
Performance
What the human–AI configuration produced, observed at the level of output.
Capacity
What the person can now do without the system present.
Agency
Whether the person is directing the work or being directed by it.
Judgment
Whether the person can evaluate the work well enough to accept, correct, or reject it.
Authority
Who has standing to decide that a result may be relied on.
Persistence
What state accumulates across sessions, systems, and time.
Correction
Whether an error can be found, contested, and fixed—and whether the fix propagates.
Dependence
What happens to performance and capacity when assistance is withdrawn.
Provenance
Whether the origin and history of a result or a retained state can be traced.
Portability / exit
Whether the person can take what is theirs and leave without loss of standing.
Status: This is a proposed diagnostic architecture, not a validated instrument. No scores are reported. See Evaluations for the staged tasks now in design, or the related evaluation specification for the technical version of the authority-transfer diagnostic.
Protection through visibility
Measurement should support the person being measured.
That means collecting only what is necessary, preserving context and provenance, making uncertainty visible, allowing corrections, separating assisted performance from retained human capacity, and preventing measurements collected for one purpose from quietly becoming judgments used for another.
Minimal collection
Collect only what a stated measurement purpose requires.
Preserved context
Keep provenance attached to whatever the measurement produces.
Visible uncertainty
Show confidence and evidence status rather than a bare figure.
Correctable record
Let the measured person contest and correct what was recorded.
Separated capacity
Keep assisted performance distinct from what the person retains alone.
Bounded purpose
Prevent a measurement collected for one purpose from becoming a judgment used for another.
Where this sits
An operational layer over the research program
Measurement translates the questions developed across the research program into dimensions that can eventually be observed, contested, and tested. It does not replace the underlying papers, and it does not yet report results.