Product
How our AI feedback stays accurate, not chatty
A look at the rubric-locked prompt design behind every feedback score.
Engineering team · Apr 28, 2026 · 8 min read
Generic LLM chatbots tend to either over-praise ("Great job!") or hallucinate corrections. Neither is useful for a learner. Our approach: lock the model to a rubric.
The rubric
Every recording is scored on five dimensions: clarity, grammar, vocabulary, fluency, and task completion. Each on a 1–5 scale. Plus one actionable tip.
Grounded in the transcript
The model is given the student's actual transcribed words and forced to reference them. No invented sentences. No "you might have said." Just what the student actually said.
If the AI can't quote you, it can't critique you.