Conversational tutors
Test explanation, questioning, scaffolding, misconception handling and appropriate escalation.
- Multi-turn learning paths
- Help without answer leakage
- Memory boundaries
Education & learning systems
Test whether AI supports learning—not only answers.
A factually correct output can still be pedagogically wrong. Construct evaluates whether a tutor, feedback system or content workflow supports the intended learning objective across learner contexts and interaction histories.
We evaluate the AI system—not a learner's intelligence, personality, diagnosis or future.
Systems in scope
Evidence for controlled educational AI use.
The system must be runnable and one intended use must be isolated. High-stakes grading, admission, diagnosis or automated trajectory decisions need separate governance and may be declined.
Formative feedback
Check whether feedback is accurate, actionable, rubric-aligned and consistent for equivalent work.
- Writing and problem solving
- Rubric reliability
- Educator review
Activities and item generation
Stress curriculum alignment, answer integrity, difficulty assumptions and accessibility across variants.
- Equivalent forms
- Distractor and solution quality
- Language/context checks
Educator copilots
Test planning, summarisation and decision support while preserving human authority and uncertainty.
- Source grounding
- Approval boundaries
- Workload and latency
Evaluation dimensions
Seven questions beyond “is the answer right?”
Does the interaction serve the objective?
Responses and actions are mapped to the intended skill, level and pedagogical strategy.
Is the reasoning useful and accurate?
We inspect correctness, clarity, examples, uncertainty and sources where required.
Does help preserve learner agency?
The system should support the next step without bypassing the learning task or over-helping.
Are equivalent cases treated comparably?
Repeated runs and parallel tasks expose stochastic drift and rubric instability.
Do language and profile alter quality?
Matched cases can expose unjustified gaps while keeping interpretations proportional to available data.
Does the learning path remain controllable?
We test continuity, outdated assumptions, contamination and appropriate response to progress or feedback.
Does the system know when to escalate?
Educator-only decisions, uncertainty and review points must remain visible and effective.
About formal psychometrics
Without an appropriate human response sample and study design, Construct does not estimate human item difficulty, discrimination, DIF or learning gain. We evaluate the system and the agreed rubric within the tested environment.
Education protocol
From learning objective to deployment boundary.
Define the use
Learning objective, audience, educator role and the decision the evidence must support.
Agree the criteria
Educators or subject experts define correctness, useful support, limits and critical failures.
Model contexts safely
Specify levels, languages, accessibility needs and common errors without personal records in the public request.
Build equivalent tasks
Representative, parallel and difficult cases include multi-turn trajectories with and without memory.
Repeat and compare
Measure correctness, feedback quality, consistency, context gaps, cost, latency and escalations.
Ground with experts
Subject and education specialists review critical or disputed cases against the agreed rubric.
Measure bounded fixes
When included, agreed changes are applied and the same evidence protocol is rerun.
Set the boundary
The evidence brief states suitable use, open risks, human oversight and continue/conditional/hold.
Deliverables
Evidence an education team can review.
- Learning-objective and criteria map
- Versioned profile, context and interaction suite
- Reusable expert rubric and runner
- Correctness, consistency and context analysis
- Ranked pedagogical and technical failure map
- Expert review record when included
- Before/after comparison and deployment boundary
Required inputs
Bring educational intent, not learner records.
- One learning objective and intended audience
- A runnable controlled environment
- Sample activities or representative tasks
- Existing rubrics or curriculum references
- An available educator or subject expert
- Known governance or accessibility constraints
Minors and learner data
Do not submit student names, records, responses, health information or data about minors through the public form. Any necessary processing must be agreed separately.
Appropriate evidence decisions
- Pilot with a small supervised group
- Choose between two configurations
- Hold a feature until critical failures are fixed
- Define educator escalation and usage limits
- Prepare a later efficacy study
Not established by this Sprint
- Accreditation or test certification
- Learning efficacy or causal impact
- Automated grading or admission approval
- Diagnosis of a learner
- Replacement of educator judgement
Preparing a supervised education pilot?
Bring one objective and one learning workflow.
We will identify the smallest defensible protocol, the educator input required and the decision boundary the evidence can support.