Find the first supported issue
A useful tutor should locate the earliest step that evidence shows is wrong, not merely report that the final answer differs.
Research and evaluation
Better tutoring starts with better questions.
Amend is built around a simple standard: feedback should help a student think and continue. This page documents how we turn that standard into concrete evaluation criteria for handwritten STEM work.
Read the open benchmarkWhat good looks like
An answer can be right while the tutoring is wrong.
A completed solution may be accurate and still interrupt learning. We evaluate whether feedback understands the visible work, identifies a supported issue, and leaves the student with a useful next action.
Preserve student reasoning
Feedback should create a productive next move—a question, hint, or correction—without replacing the student's work by default.
Treat ambiguity as ambiguity
Unclear handwriting or incomplete work should trigger clarification or uncertainty, not a confident invented interpretation.
Evaluate the whole interaction
Correctness matters alongside error location, instructional usefulness, calibration, and whether the student can continue.
Evaluation stack
Measure the path, not just the destination.
- 01
Interpretation
What written expression, diagram, or step did the system understand from the page?
- 02
Verdict
Is the visible work correct, partially correct, incorrect, or too unclear to judge?
- 03
Localization
If there is an error, which earliest line contains it and what evidence supports that call?
- 04
Instruction
Does the response offer the smallest useful next move without taking over?
- 05
Calibration
Does the response express uncertainty when the handwriting or mathematical evidence is insufficient?
Published now
Handwritten Math Tutor Benchmark v0.1
A reproducible starter set of written-work cases, expected behaviors, failure conditions, and a scoring rubric. The protocol is public; product performance scores are not yet published.
Explore the benchmarkImportant limitation
AI feedback can be wrong.
Handwriting interpretation and generated tutoring are probabilistic. Amend is a learning aid, not a substitute for an instructor, and students should verify important results against course materials.