Learning happens in the conversation.
Conversations, multiple choice, and concept maps: three ways to see how a learner actually reasons, all grounded in the materials you assign, with every judgment still in your hands.
You said entropy "measures disorder." Can you tell me what is actually being counted when entropy increases?
Good. So push that further. Why would a gas expanding into a vacuum have more of those arrangements than before?
Author the intent. Sofiya runs the dialogue. You keep the judgment.
Teacher configures the assignment
Opening question, learning objectives, rubric, source materials, and turn limits. Nothing about the pedagogy is left to the model.
Student has the guided conversation
A structured, bounded dialogue grounded in the assigned materials. The partner probes their reasoning and brings it to a natural close.
Teacher reviews the transcript
An authentic record of how the student actually reasoned: harder to fabricate than an essay, more revealing than a score.
Teacher authors the feedback
Holistic written feedback against their own rubric, authored entirely by the teacher, in their own voice.
One objective. Three kinds of evidence.
A learner who picks the right option has not necessarily understood it. Sofiya can assess the same objective as a conversation, as a multiple choice exam, and as a concept map, because the places those disagree are where the learning actually is.
Conversation
A structured Socratic dialogue with a partner you configure. It probes the learner's reasoning and asks the next good question. It never hands over the answer.
- Bounded by the turn limits you set
- Grounded in the materials you assign
- You author every word of the feedback
Multiple choice
Items drafted from your own material, where every wrong option can name the misconception it reveals, so a wrong answer tells you something a score cannot.
- Drafted from the material you attached
- Distractors carry a misconception tag
- A wrong answer can open a short follow-up
Concept map
Learners lay out how the ideas connect: filling in blanks, assembling supplied pieces, building from scratch, or talking it through. Marked against your expert map.
- Four modes, from scaffolded to blank canvas
- Partial credit for the right pair, wrong relationship
- Keyboard-operable, not a canvas-only feature
- ch4-5-statistical-entropy.pdf0.94
...entropy counts the number of accessible microstates for a given...
- lecture-notes-week3.pdf0.87
...free expansion into a vacuum does no work and exchanges no heat...
- problem-set-4-solutions.docx0.82
...worked example: entropy change for an ideal gas expansion...
Every dialogue, grounded in your course.
Retrieval-grounded in the readings, lecture notes, and materials you assign, never the open internet.
See how retrieval worksgrounded in the materials you assign
Never the open internet
Readings, lecture notes, and source documents you assign, nothing else.
Isolated by tenant
Each institution's data is isolated from every other tenant. Nothing crosses that boundary.
Every answer traces back
Grounding traces to specific chapters and materials, the same way instructors already cite sources.
A guide, not an oracle.
The partner probes the student's reasoning, asks follow-up questions, and helps them think. It does not lecture or give answers.
Apply to the faculty trialDid individual moths change colour, or did something happen across the population?
hands over the answer
Retrieval-grounded
Every response is anchored in the materials you assigned, not the open internet.
Asks the next good question
It probes reasoning and follows up rather than lecturing or handing over conclusions.
Brings itself to a close
Minimum and maximum turns keep every student's experience bounded and comparable.
A puck slides on frictionless ice. What happens to it?
- It speeds up
- It slows down but keeps moving
- It stops immediately
- It keeps a constant speed
Misconception
Believes motion needs a continuing force
Confidence against correctness
A wrong answer that tells you why.
Sofiya drafts items from the material you attached, then asks you to say what each wrong option reveals. From then on a wrong answer is a reading of the learner's thinking rather than a lost mark.
Apply to the instructor trialshows a learner your misconception tags
Every distractor names a misconception
You write what choosing that option suggests the learner believes. Learners never see it: the tag is stripped from what they are served.
Confidence against correctness
A 2×2 that separates sure and right from sure but wrong, the quadrant a percentage score hides completely.
Wrong answers can open a follow-up
A short Socratic exchange about that one item, capped per submission, that never affects the mark.
A hundred conversations, without a hundred pages.
Read the cohort before you read anyone in particular, then walk the queue with one keystroke. Notes on individual replies stay collapsed until you want them, and a comment you write once is waiting for the next learner.
See how the loop works7 of 24 waiting
Assignment · Unit 3
Shared misconceptions
11 learners treated correlation as cause
save and open the next learner waiting
The class before the individual
Run the analysis at the head of the queue and see the misconceptions the class shares, before the first three transcripts form your opinion for you.
Save & next
The queue counts only work still waiting on you for one assignment, which is what keeps the number beside it true.
Bulk return, never bulk assess
Release a set of already-assessed conversations in one action. Nothing here publishes feedback a human did not write.
Learning objective
Distinguish correlation from causation
When the exam and the map disagree, that is the finding.
Sofiya lines up what each modality says about the same learning objective and refuses to average them. A learner who recognises the right answer but cannot explain it is not a middling grade, they are a specific and nameable problem.
See the three modalitiesmodalities, side by side, never averaged
Named patterns, not a blended score
“Recognises, can’t explain.” “Explains, misses on recall.” The label says what to do next; an average would have buried it.
Arithmetic, not another model
The reconciliation is a computation over work that was already marked. No model decides what anyone’s answer was worth.
Silent below five learners
No cross-modality finding is published from a sample too small to generalise from.
Your materials go in. Grounded follow-ups come back.
The readings, lecture notes, and source materials you assign are indexed for one course. Every follow-up question the partner asks is grounded in what you actually assigned, with citations back to the material.
Sources
Course materials
0
chunks indexed
The molecules have more space, so there are more microstates?
Partner
Good. Now connect that to why entropy specifically increases in a free expansion. Grounded from 2 materials.
Privacy and compliance, built in.
Learner data is stored and processed exclusively in Canada, designed for compliance with BC FIPPA, BC PIPA, Alberta PIPA, and PIPEDA.
Multi-tenant isolation
Each organization's data is isolated from every other tenant. Nothing crosses that boundary.
Role-based access
Granular roles so instructors, learners, and admins only see what they should.
Audit logs
Every state-changing action is logged and database-enforced immutable, not just application-level.
LMS-ready sign-in
Single sign-on and roster sync via LTI 1.3, no separate logins for learners or instructors to manage.
Not a chatbot. Not a worksheet. A guided dialogue.
| Capability | Sofiya | Open-ended AI chatbot | Traditional essay / worksheet |
|---|---|---|---|
| Grounded in your assigned materials | |||
| Adapts follow-ups to the student's actual answer | |||
| Bounded, comparable for every student | |||
| Resistant to being outsourced to AI | |||
| Authentic record of the student's reasoning | |||
| Teacher authors all assessment and feedback | |||
| Shows where two forms of assessment disagree |
Pricing that scales with your rollout.
Start with one course. Talk to us when you're ready for your department or organization.
Sofiya is currently available on a free trial basis for interested Instructors.
For an individual instructor trialling Sofiya with their own group.
- Every partner role
- All assessment stays yours
- Grounded in your own materials
FAQ
The questions educators actually ask.
The Socratic method, at the scale of your whole class.
Bring a real assignment. See the loop end to end.