Reasonline AgencyThread · Method
The Human Agency Index
HAI describes the extent of substantive human participation that the submitted evidence supports in one work process. This page sets out what that means, how the evidence is read, when a score is withheld, and how far the method has been validated.
The construct
The method asks how far the observable evidence supports substantive human participation in framing, directing, evaluating, developing, assuring, understanding, and taking responsibility for work produced with AI. It is an inference about one work process, drawn from the record the student submitted.
What HAI doesn’t claim
- It doesn't measure artifact quality, correctness, learning, expertise, or effort.
- It isn't the share of thinking the student did.
- It makes no claim about who produced the work.
- It doesn't decide whether the work followed a course policy. Policy is a separate question for the instructor.
- It isn't lowered simply because AI generated prose, ideas, analyses, or checks.
- It describes one piece of work. It isn't a trait score for the student.
Principles the method follows
- AI contribution doesn’t cancel human agency, and human direction doesn’t absorb the AI’s contribution.
- Calibrated acceptance of an AI suggestion counts as much as calibrated rejection.
- Missing evidence is not evidence of missing thinking.
- Agency is never measured by word count, typing volume, time spent, edit count, or manual effort.
- A workflow with more delegation isn’t worse in itself. What fits depends on the task.
Six dimensions
Each dimension has four subcriteria, each rated 0 to 4 against written anchors. A dimension score is 25 times the mean of its rated subcriteria. Reports lead with these dimensions and the evidence cited for each.
| Dimension | Question it asks of the record | Group | Weight |
|---|---|---|---|
| D1. Problem Framing & Evolving Ownership | What is being pursued, and does the student meaningfully shape what it becomes? | Direction and control | 15% |
| D2. Knowledge Direction & Metacognitive Control | Does the student plan, steer, adapt, and govern how the work proceeds? | Direction and control | 15% |
| D3. Critical Evaluation & Calibrated Reliance | Does the student appropriately question, accept, reject, qualify, or correct AI contributions? | Epistemic engagement | 25% |
| D4. Synthesis, Integration & Constructive Development | Does the student meaningfully connect, develop, combine, extend, or restructure contributions? | Epistemic engagement | 20% |
| D5. Epistemic Oversight & Assurance | Does the student govern reliability, evidence, uncertainty, verification, and escalation? | Epistemic engagement | 15% |
| D6. Understanding, Decision Authority & Accountability | Does the student understand consequential work, keep meaningful authority, and take responsibility? | Responsible ownership | 10% |
The weights, anchors, withholding rules, bands, and High gate belong to the locked method (HAI v0.21). Instructors can’t change them. Instructors set the task context: the assignment’s stakes, how much careful reasoning it demands, how much could reasonably go to AI, and what only a person can do. Task context helps interpret a report. It doesn’t raise or lower the score.
How the evidence is read
- The student’s evidence packet is prepared as one record: the final artifact, conversations, drafts, revision history, sources, and any Learn records, each labeled with what Reasonline knows about where it came from.
- An AI reviewer (following the locked manual and prompt) reads the record three separate times. Each reading rates the 24 subcriteria, cites the evidence for each rating, and describes how confident the evidence allows it to be.
- A rating is kept when at least two of the three readings agree exactly. When they don’t, the report shows that subcriterion as unresolved, with the evidence each reading cited, for you to weigh.
- Ordinary software, not the AI reviewer, then computes the dimension scores, the HAI, the band, and whether a score is withheld.
- Every report records the method version, the scoring software version, the exact model, its settings, and how the evidence was prepared, along with the instructor’s settings at the time.
The reviewer reads only this student’s packet for this assignment. When a piece of evidence can’t be interpreted, the report says which piece and lowers its stated confidence. It makes no finding about honesty or misconduct.
When a score is withheld
HAI isn’t fully compensatory: strong direction can’t make up for weak evaluation. The method withholds a composite score rather than report one it can’t support:
- A dimension needs at least three rated subcriteria. If any dimension falls short, no composite is reported.
- If the overall evidence confidence is low, no composite is reported.
- A final artifact with no process evidence can’t support a composite.
A withheld score is a correct result, not a failure. The report still describes each dimension the evidence does support, and says what evidence would help.
Evidence confidence and opportunity
Evidence confidence asks how much of the process is visible, how reliable each source is, whether stages are missing, whether separate sources agree, and whether evidence ties to specific decisions. Opportunity to demonstrate asks whether the task gave a dimension a natural chance to show. A task can offer plenty of opportunity while the submitted record shows little of it, and the report keeps those two apart.
Bands and the High gate
- Lowbelow 40
- Moderate40 to 59
- Substantial60 to 79
- High80 and above, and only when the High gate is met
The High gate requires at least 60 in Critical Evaluation (D3), Synthesis (D4), and Understanding and Accountability (D6), evidence confidence of at least Moderate, and at least 60 in Oversight (D5) when the task’s stakes or reasoning demand are rated High. These thresholds are calibration hypotheses, not validated cut scores.
The 0 to 100 scale is a rubric index, not a percentage of anything. Separate readings of the same evidence can land in different bands near a boundary, which is one reason reports lead with the dimensions.
Validation to date
Claim maturity: Validation in progressThe method was calibrated on synthetic cases written from known targets, and Reasonline’s scoring software must reproduce saved results exactly before release. It has not yet been compared with trained human coders. That reliability study is planned under an IRB protocol, separate from product data, and its agreement figures will be published here when they exist.
Until then, the composite HAI carries a “Validation in progress” note wherever it appears, and the Collaboration Profile, which describes the shape of the collaboration, is labeled Experimental and shown to instructors only. Independent educators will review representative reports before any student sees one.
How students see and respond to a report
- Reports are private to the instructor unless the instructor shares them.
- A student who receives a report sees the reasoning and the evidence cited for each dimension. Students never see the composite HAI or the Collaboration Profile.
- The student can correct a factual error, such as a misread date or a misattributed draft. The correction goes to the instructor.
- A report never changes a grade and is never sent to an LMS as a grade.
A report describes evidence about one piece of work. The judgment stays with the instructor.