Research
Assessment you can defend, data you can reproduce
ARASH is built for schools that will be asked how the scores were produced — and for researchers who need to answer that question years later.
Approach
Four commitments behind every score
Constructs before models
Rubrics are defined against recognised speaking constructs first; models are then calibrated to them, not the other way around.
Human-in-the-loop calibration
Teacher overrides feed a review queue that measures agreement between human raters and each engine version.
Reproducible datasets
Research exports are append-only, pseudonymised, and stamped with the rubric and model versions in force at capture time.
Consent as a first-class scope
Data enters the research vault only within the consent scope a parent granted, and leaves it when consent is withdrawn.
Constructs
The six dimensions we measure
Every dimension has a published rubric, a band descriptor set, and a version stamp attached to each score.
Pronunciation
rubric v1 · calibrated
Grammar
rubric v1 · calibrated
Vocabulary
rubric v1 · calibrated
Fluency
rubric v1 · calibrated
Comprehension
rubric v1 · calibrated
Pragmatic
rubric v1 · calibrated
Collaboration
Studies we would like to run with you
We work with universities, foundations and district programmes. Consent scope and ethics review come first.
- Classroom pilots with pre/post speaking measures
- Inter-rater agreement studies between teachers and engines
- Longitudinal studies of fluency and interaction development
- Fairness reviews across accents, schools and socioeconomic contexts
Interested in a research partnership?
Tell us about your study design and we'll scope the data access together.