Trust & method
A system that interprets human intellectual contribution should be explicit about what it can support, what it cannot support, and how it handles uncertainty. AgencyThread is built around that principle.
AgencyThread interprets only what the available record supports. It does not infer human agency from polished prose, professional tone, or a final artifact alone. If evidence is absent, the system should not quietly turn that absence into a claim about the person.
A human may have done meaningful thinking that was never captured. AgencyThread treats that as an evidence limitation, not proof of absent agency. Where the record cannot support a dimension or overall result, the system should withhold rather than guess.
Every AgencyThread result includes Evidence Confidence. That allows readers to distinguish what the evidence suggests about human agency from how strongly the record supports that interpretation. A lower-confidence record does not mechanically reduce HAI. It changes how strongly the interpretation can be stated.
AgencyThread does not treat more prompts, more edits, more time, or more typing as proof of greater human agency. The product is designed to reward meaningful intellectual contribution, not visible busyness.
Human agency should not depend on whether someone worked in ChatGPT, Claude, Codex, Gemini, or another system. AgencyThread is designed around a provider-neutral evidence model so that the underlying contribution can be interpreted across different AI workflows. Current authenticated integrations are still in development.
AgencyThread is not designed around continuous keystroke monitoring, screen recording, or time-on-task surveillance. Its evidence model focuses on meaningful intellectual events: decisions, critiques, revisions, verification, escalation, and final responsibility.
Public nonclaims
AgencyThread does not determine whether the finished work is good or bad, whether it is correct, whether AI use was permitted, plagiarism, originality, legal authorship, intelligence, effort, or time spent.
It is not an AI detector.
It is not currently intended for high-stakes use in grading, hiring, discipline, promotion, or similar consequential decisions.
Validation status
The Human Agency Index framework has undergone synthetic calibration, adversarial stress testing, repeated reliability testing, naturalistic evaluation, and private-Alpha testing on real work.
Those stages have helped refine the construct, scoring rules, evidence handling, and reporting behavior.
What has not yet been completed is the human/expert validation required to justify stronger high-stakes claims. That work is a separate next stage.
The Collaboration Profile remains experimental.