ResourcesPedagogical AI & evaluation

Pedagogical AI: an agentic LMS guided by teaching design.

Fluent text isn't the same as good teaching. Remedia's AI works inside a teaching workflow: it plans from learning objectives, generates content, has its work reviewed by an independent AI, fact-checks its own claims and hands everything to an educator for the final decision. Here is how it works, and how we test it.

MStudent answersLearner 12Learner 13Total score8 / 129 / 12Question 01Multiple choiceDemand is elastic when…A. price rises and revenue risesB. quantity falls more than price risesC. buyers ignore price changes2/33/3MarksCorrect answerABCJustificationCorrect choice. Explain the revenue effectto earn the full justification mark.Question 02Short answerExplain how elasticity affects total revenue.Revenue falls when price rises because buyerscut back more than proportionally, so a firmfacing elastic demand should avoid price rises.Marks are suggested from your rubric.RubricOrganisation2/3Clarity2/3Critical thinking2/3Grading progress18 of 25 submissions marked19 of 25 submissions markedNext learner →
DefinitionWhat is an agentic LMS?

What is an agentic LMS?

An agentic LMS is a learning platform where AI agents carry out multi-step tasks, such as planning a module, drafting assessments, reviewing their own output or pre-marking work, instead of answering one prompt at a time. In a pedagogical agentic LMS, those agents are bound by teaching design: objectives, levels, criteria and educator review.

BlueprintIntroduction to Microeconomics · Price elasticity of demand

Course blueprint

  • LO1Define price elasticity of demandRemember
  • LO2Calculate elasticity from a demand scheduleApply
  • LO3Explain how elasticity affects total revenueAnalyse
LevelUndergraduate · Year 1
Difficulty · set by educator
How Remedia's AI works

Five steps between your request and what students see.

1. Plan from objectives

Generation starts from your subject, level, standards, learning objectives, difficulty and Bloom level, plus the sources you provide.

2. Generate

A generator model drafts the course section, assessment or rubric, grounded in your materials.

3. Independent review

A separate reviewer AI, which never sees how the content was produced, checks scientific validity, consistency, level and pedagogical quality. Only proven blocking errors count.

4. Fact verification

Definitions, values, calculations and answer keys are checked one by one by a different model, and suspected errors are confirmed by a third model to avoid false alarms.

5. Targeted correction

Only the faulty items are rewritten. Everything else stays untouched, and the best reviewed version is kept.

Then: the educator decides

The reviewed draft goes to the teacher, who edits and approves before anything reaches students.

Guardrails

Designed to avoid AI's usual failure modes.

No criticism for its own sake

The reviewer can only fail content on a closed list of blocking categories, and the loop stops when a correction stops improving the result.

No silent rewrites

Corrections target the faulty question or block only, so content you've already checked doesn't change behind your back.

Grounded in your sources

Generation and review take the documents you supplied into account.

Human decision, always

AI suggestions are drafts. In teacher-led assessment, every mark is approved by an educator.

How we testErrors caught, false alarms avoided

We measure errors caught and false alarms.

Every change to a model, prompt or verification step is tested on an internal benchmark before release. Each test case exists as a minimal pair: one version with known errors and the same version corrected. That lets us measure how many real errors the reviewer catches and how often it raises an alarm on correct content. A change is only kept if it catches more errors without adding false alarms.

  • Real errors from past generations weighted above injected ones
  • Full-loop tests check corrections fix errors without breaking correct items
  • Human review sheets for expert checking of each case
InsightsIntroduction to Microeconomics · Price elasticity of demand

Class insight · by criterion

Definition
Calculation
Interpretation
Application
Recurring difficulty: interpreting the elasticity value. Consider a follow-up practice activity.
AI marking

How AI marking stays accountable.

Your rubric is the reference

Pre-marking applies the criteria and level descriptors you define, visibly, criterion by criterion.

Adjustable severity

Choose how strictly criteria are applied, from encouraging (formative) to strict (summative).

Feedback you can read

Each suggested level comes with an explanation the teacher can check in seconds.

The right model for each task

We compare models task by task, balancing quality, reliability, speed and cost. The reviewer can run on a different model from the generator, so one model's blind spots are caught by another.

On publishing benchmarks

We'll only publish benchmark figures once the task, model versions, sample size, metrics and limitations are documented. This page describes our approach; it is not a certification or a performance guarantee.

FAQ

Pedagogical AI, answered.

What is pedagogical AI?

Pedagogical AI is artificial intelligence designed around teaching and learning principles, such as learning objectives, appropriate challenge, transparent criteria and educator oversight, rather than general-purpose text generation.

Can AI-generated course content contain errors?

Yes, any AI can make mistakes. Remedia reduces them with an independent reviewer and claim-by-claim fact verification by a second model, and educators review content before students see it.

What is LLM-as-a-judge?

It's a technique where one AI model evaluates the output of another against defined criteria. Remedia uses an independent reviewer model on four axes and only acts on proven, blocking errors.

Does Remedia use AI to make final grading decisions?

No. In teacher-led modes, educators review and approve every final mark.

See where educator review sits in the workflow.

We'll generate an assessment live and show you what the reviewer caught.

See the Review Loop in a Demo