Reconstructed
- Saved conversations
- Drafts
- Artifacts
- Notes
- Retrospective evidence
How it works
The record starts with evidence, not the finished document
A polished final file can come from many different workflows. AgencyThread does not try to infer the process from prose style alone.
Instead, it starts with evidence of the work behind the work: conversations, drafts, decisions, corrections, source checks, and meaningful revisions.
Today, AgencyThread can evaluate Reconstructed records assembled from saved conversations, drafts, notes, artifacts, and other surviving evidence.
The Alpha architecture has also tested prospective Verified evidence capture, where evidence is recorded as work happens. Direct authenticated provider integrations and broader automated capture are still being developed. The important distinction is not a different scoring system; it is the strength and provenance of the evidence available to support the interpretation.
Verified
- Prospective capture
- Stronger provenance potential
- Authenticated integrations on roadmap
Same HAI framework. Different evidence strength.
Step 1
Capture the work
AgencyThread gathers evidence that helps explain how a piece of work developed.
AI conversationsDrafts and version historyNotes and outlinesSource-checking activityArtifactsDecisionsHuman review checkpoints
This is not a keystroke transcript. It is a record of meaningful intellectual events: what was framed, changed, challenged, checked, accepted, rejected, and carried forward.
1Evidence
2Agency Interpretation
3Human–AI Collaboration Report
ChatDraftHuman commentRevisionSource checkFinal decision
Step 2
Evaluate human agency
The evidence is evaluated across six dimensions within three pillars.
Direction & Control
Problem Framing & Evolving Ownership
Knowledge Direction & Metacognitive Control
Epistemic Engagement
Critical Evaluation & Calibrated Reliance
Synthesis, Integration & Constructive Development
Epistemic Oversight & Assurance
Responsible Ownership
Understanding, Decision Authority & Accountability
Each dimension is evaluated on its own evidence. The system does not assume that final approval, heavy editing, or extensive prompting automatically demonstrates substantive human agency.
Step 3
Understand the collaboration
The human-facing report combines three kinds of information:
Human Agency Index
How much substantive human intellectual agency the record supports.
Collaboration Profile
How the human and AI appear to have worked together.
Evidence Confidence
How strongly the available evidence supports those conclusions.
Where the record is too sparse to support a responsible result, AgencyThread should withhold the score rather than produce a number the evidence cannot defend.
See a Sample Report →01Human Agency Index
02Collaboration Profile
03Evidence Confidence
04Why this result
05Six dimension views