Direction & Control
Who framed the problem, set meaningful constraints, and steered the work as it developed?
D1 · D2What we measure
Three pillars. Six dimensions.
AgencyThread evaluates human intellectual agency across three broad areas: Direction & Control, Epistemic Engagement, and Responsible Ownership.
Epistemic Engagement
How did the human judge, develop, and assure the work rather than simply receive it?
D3 · D4 · D5Responsible Ownership
Did the human understand the consequential choices, exercise meaningful authority, and remain accountable for the result?
D6D1Direction & Control
Problem Framing & Evolving Ownership
Who defined what the work was actually trying to do?
This dimension looks at whether the person established the objective, set meaningful criteria and constraints, and continued to own or revise the framing as the work evolved. The important question is not merely who wrote the first prompt. It is whether the human remained consequential in defining what problem the work was solving.
D2Direction & Control
Knowledge Direction & Metacognitive Control
Who steered the process, and did they recognize when it needed to change?
This looks at workflow planning, monitoring, adaptation, and boundary-setting. A person can demonstrate strong agency by deciding that AI should handle part of the task when that delegation is appropriate. AgencyThread is not designed to reward unnecessary manual effort.
D3Epistemic Engagement
Critical Evaluation & Calibrated Reliance
Did the person actually judge AI’s contribution, or simply accept it?
This dimension looks for meaningful discrimination between stronger and weaker suggestions, reasoned acceptance or rejection, recognition of errors or limitations, and appropriate calibration of trust. The goal is not disagreement for its own sake. Good judgment can also mean recognizing when an AI contribution is strong enough to keep.
D4Epistemic Engagement
Synthesis, Integration & Constructive Development
What did the person actually build, connect, extend, or improve?
This looks for substantive human development: integrating contributions, extending ideas, changing structure, developing concepts, and making choices that meaningfully propagate into later versions of the work. Final approval alone is not enough. AgencyThread looks for evidence that the person materially shaped what the work became.
D5Epistemic Engagement
Epistemic Oversight & Assurance
How did the person make sure the work was reliable enough for its purpose?
This dimension looks at risk recognition, source and claim verification, assurance design, evidence grounding, and escalation when something cannot be safely accepted at face value. The human does not have to perform every check manually. A well-designed AI-assisted verification process can still demonstrate strong human assurance when the person designs, monitors, and adjudicates it.
D6Responsible Ownership
Understanding, Decision Authority & Accountability
Did the person understand the consequential choices and exercise real authority over the result?
This looks at whether the person can meaningfully explain or defend important decisions, whether their authority matches their actual capability, and whether they remain accountable for what is ultimately used or submitted. A signature or final click is not automatically evidence of substantive ownership.
The locked continuum
What the HAI score means
Human Agency Index results fall into four bands. These bands describe how much substantive human intellectual participation the evidence supports.
Higher is not automatically better.
Different tasks appropriately require different levels of human involvement. A routine, low-stakes task may be well served by deliberate delegation to AI. That workflow can reflect sound judgment even if the Human Agency Index is relatively low.
The HAI score is descriptive. It is not a virtue score, quality grade, or measure of effort.
Low0–39Moderate40–59Substantial60–79High80–100*
*High is subject to the locked High-HAI gate.
Routine delegated taskLow HAI may be appropriate.High-stakes analysisStronger evidence of evaluation, assurance, and accountable decision-making may be especially important.
Descriptive, not evaluative
A separate question
Evidence Confidence
Those are different questions. A rich, continuous, well-provenanced record supports a stronger interpretation than a sparse or mostly reconstructed one. But missing evidence is not proof that human cognition was absent.
Where the record does not support a full evaluation, AgencyThread should withhold the result rather than silently convert missing evidence into a low score.