GROUND
Work from the course
Answers come from the materials of the class you are actually in — the readings, the slides, the assignment in front of you — and point back to where they came from.
Our mission
Coursework is multimodal. The software students get isn't.
Students work in handwriting, diagrams, notation, and code, then get help from tools that only read text. Scope is building the workspace where the work actually happens, and the models that can read it in the form it was done.
Berkeley, California
UNDERSTAND
Read the work as it is
Handwriting, diagrams, notation, notebooks, and code. Not a text box that treats everything else as an attachment.
RESPOND
Ask before answering
A question that moves you one step is worth more than a finished solution you did not write. The hard part is knowing which step you are on.
The work
What we're building.
The workspace
Somewhere to do the assignment: annotate the reading, run the notebook, write the code, work the derivation. It attaches to the Canvas or Brightspace a school already runs, and students use it free.
Course-grounded answers
Retrieval over your own course materials, so every claim points back to the page, slide, or moment in a lecture it came from — and stops at the edge of your course.
Multimodal understanding
Models that read handwritten and drawn STEM work — notation, structures, diagrams, the arrows between them — instead of flattening anything that is not text into a caption.
Proposals, then checks
Where a subject has hard rules, deterministic tools decide what holds and the model only proposes. When the reading is uncertain, the software asks. It does not rule.
Local-first, still
The index lives on your devices, handoffs are end-to-end, and on-device models run where they are available. Nothing sold, ever.
Scope Labs
The research arm. Continuous literature review, a graph tying each method to the component it would change and the experiment that would test it, and evaluation specified before anything is trained.
Claim discipline
The limits are part of the work.
This is not a grading system. Scope does not score work. There is no instructor gradebook, no roster sync, and no replacement for the LMS a school already runs.
A model that reads handwriting will misread some of it. A misread mark and a real misunderstanding look identical a step later. We treat that as the central problem rather than an edge case, which is why the checks are deterministic and the uncertainty is shown.
Producing an answer is the easy part. Any general model can hand over a finished solution. Whether a student can then do the next problem alone is a different measurement, and it is the one that matters.
None of the model work is proven yet. The architecture is designed and the evaluation is written down. The numbers come after, and they go out either way.
How a feature earns its place.
STEP 01
Start from a real failure
A feature begins with coursework the current tools handle badly, not with a capability we would like to demonstrate.
STEP 02
Decide what would falsify it
The success criterion, the split, and the baseline are fixed before the build. A demo that works on the example it was built from is not evidence.
STEP 03
Measure the student, not the demo
The test is whether the work got better afterwards — including when the honest answer is that it did not.
Build it with us.
We want to hear from students who will tell us what is broken, instructors willing to argue about what good help looks like, and researchers working on multimodal understanding, tutoring, and evaluation. The shared standard is a claim that survives a hostile read.