Skip to content
ExamGrader.aiPowered byPOINTS

Benchmark · as of 14 October 2026

How well does ExamGrader grade real papers?

We graded 52 real handwritten IGCSE and IB scripts on the live ExamGrader system, then re-ran a 12-paper pilot 3 times while we fixed what it exposed. Here are the results, by evaluation type, including what we are still improving.

  • 52

    real handwritten scripts graded

  • 100%

    of evaluations completed

    36 of 36 pilot runs

  • 12/12

    correct letter grades

    was 3/12 before fixes

  • 0

    garbled handwriting readings

    was 14 before fixes

Scores by evaluation type

Results for each type of paper

Only types that were part of this benchmark have scores. Green is met or fixed; amber is still being improved.

QPA

Handwritten answers on the question paper

52 live scripts · 12-paper pilot run 3 times · mark scheme included where one exists

About this type

Evaluations completed

100%

36 of 36 across all 3 runs

Correct letter grade

3/12now12/12

Pilot papers, before vs. after fixes

Garbled transcriptions

14now0

Handwritten answers, before vs. after

Answers left unreadable

4now1

Before vs. after the handwriting fix

Total matches printed paper

5/12

Run 3 · fix shipped, being verified

Score consistency between runs

±17 pts

Largest spread for one script · being improved

Choice

Choice papers (“answer one of”)

One Economics paper, scored before and after the fix

Economics paper score

35%now70%

Now scored only on the questions the student chose

Not benchmarked yet

These types are supported but were not part of this run, so we don't publish scores for them yet.

What we fixed

Before and after, on the same 12 scripts

Every run completed 12 of 12 papers. Fewer errors is better.

Errors in the 12-paper pilot

  • Papers with the wrong letter grade

    Before: 9 of 12 papers. After: 0 of 12 papers.

    Letter grades now always match the percentage score.

  • Garbled handwriting transcriptions

    Before: 14 answers. After: 0 answers.

    After the handwriting fix, every answer in the pilot was read cleanly.

  • Answers marked unreadable

    Before: 4 answers. After: 1 answers.

    Far fewer answers are left ungraded because the handwriting could not be read.

View as table
Errors in the 12-paper pilot
MetricBeforeAfter
Papers with the wrong letter grade9 of 120 of 12
Garbled handwriting transcriptions140
Answers marked unreadable41

  • Choice papers scored correctly

    On “answer one of” papers, only the questions the student chose are scored. One Economics paper went from a wrong 35% to a correct 70%.

  • Crossed-out work is ignored

    When a student crosses out an answer and writes a correction, only the correction is graded.

  • Nothing had to be re-uploaded

    Every evaluation finished in every run (12 of 12 each time). In run 3, 7 evaluations were interrupted by a server update and resumed on their own. No one had to re-upload anything.

Built-in safeguards

What you can count on

Whether you upload one script or a whole class.

  • Evaluations always finish

    If something is interrupted, the evaluation picks up again on its own. You never re-upload.

  • Never charged twice

    A paper is graded and billed once, even if it has to be retried.

  • Grades match the score

    The letter grade is always worked out from the percentage, so the two never disagree.

  • No skipped questions

    A second check grades any question part that was missed the first time.

  • Same paper for the whole class

    In a bulk upload, every student is graded against the identical question paper and mark scheme.

  • Flags anything that needs a look

    Results are marked “Needs review” when totals don't add up, question numbers repeat, or an answer couldn't be graded.

Honest limitations

What we are still improving

  • Graded total vs. printed paper total

    In run 3, the graded total differed from the printed total on 7 of 12 papers.

    Fix shipped, being verified

  • Run-to-run consistency

    Scores for the same script varied by up to 17 percentage points between runs.

    Improvements shipped, being verified

  • Agreement with teacher marking

    We have not yet measured how closely ExamGrader agrees with teacher-marked scripts.

    Next benchmark

How we tested

Real papers, uploaded like a teacher would

  1. 1

    52 real handwritten student scripts were graded on the live system, uploaded exactly as a teacher would.

  2. 2

    Each upload had the question paper with the student's answers, plus the mark scheme where one exists.

  3. 3

    The same 12 scripts were graded 3 times: before fixes, after the first fixes, and after the handwriting fix.

  4. 4

    Each run was compared on letter grades, handwriting reading, unreadable answers, completion and totals.

Subjects covered

  • IGCSE Chemistry
  • IGCSE Physics
  • IGCSE Biology
  • IBDP Chemistry (HL/SL)
  • IBDP Biology (HL/SL)
  • IBDP Physics (HL/SL)
  • IBDP Economics (HL/SL)
  • IBDP Business Management (HL/SL)
  • IB Mathematics AA HL

Results as of . Anonymised: no student names or identifying details are published.

Ask us about the benchmark

See it on your own papers

Upload a real handwritten script with its mark scheme and check the result yourself. Start free with 50,000 tokens, no credit card required.

Start free Talk to us