Moderation of marking: two markers, one standard

Two teachers mark the same written question for two halves of a year group, or one teacher marks sixty answers over three evenings. Moderation of marking is the check that a mark did not depend on who held the pen, or on which evening.

Below is what research and published procedures report, each read at its source on 10 October 2026 (an abstract only where it says so), then a procedure for one meeting of two to six markers.

EdProcto, which publishes this guide, sets your rubric under every written answer as it is marked and can hold every score until you release it. Writing the rubric is in our guide to short answer questions.

What is moderation, and how does it differ from double marking?

Moderation reviews, after marking, whether a whole set of marks meets one standard; double marking gives every script two marks.

• Standardisation, before marking. Every marker marks the same few scripts and the team agrees how the scheme applies. The University of Bristol (n.d.) calls it benchmarking and gives five scripts as an example.
• Double marking. Two markers mark every script, then agree a final mark. Blind means neither sees the other's mark first.
• Second marking of a sample. The same, for some scripts only, blind or with the first marks in view.
• Moderation, after marking. The set is judged from a sample, then confirmed, adjusted or re-marked. In Bristol's code a marker who sees another's marks is moderating, not double marking.

How far do two markers agree on the same script?

Closely on constrained questions, and much less on extended writing. The measure is inter-rater reliability.

• Fox et al. (2026), an unreviewed preprint, report 32,534 answers to 328 questions from mock papers in the style of England's GCSE, each marked independently by two qualified examiners. Marker agreement (a weighted kappa: 0 is chance, 1 identical marks) was 0.84 in maths, 0.69 in the sciences and 0.65 in English (eight questions only). The authors work for the company that ran the mocks.
• Zhao et al. (2017) analysed 67,500 essays from a high-stakes English examination in China in 2011, each marked out of 20 by two of 88 qualified raters. On 17% of essays the two marks stood four or more points apart.

What agreement depends on

On the question, on the scheme and, less surely, on training.

• Ofqual's working paper (Williamson, 2026) notes that the task and the mark scheme both make marking easier or harder, and that the more freedom a task gives, the less consistent its marking tends to be.
• Baird et al. (2004) ran two experiments with UK examiners: neither exemplar scripts with feedback nor discussion at a coordination meeting improved marking reliability (abstract read). A study where calibration did narrow the marks is in our guide to marking a large class. So meet to settle the scheme's wording, then check that it worked.

Is double marking worth its cost, and should the second marker see the first mark?

For every script, only where the stakes are high; and the second marker should not see the first mark.

• Cannings et al. (2005) analysed double-marked case studies from one medical school over two periods. The markers agreed reasonably well, yet a student's final mark was only moderately reliable; the authors suspect variation elsewhere in the process (abstract read).
• Garry et al. (2005) confirmed by experiment, on politics exams, that a moderator who sees the first mark changes it by less than two independent markers differ. They judged blind double marking more objective and dearer (abstract read; no sample given).

So blind-mark a sample, and double-mark only work that decides an award.

How does one marker drift over a long session?

Less by hardening steadily than by wobbling, tiring and following the scripts just read.

• Leckie and Baird (2011), on England's 2008 national writing test for 14-year-olds, found no significant overall drift, yet each rater's severity was significantly unstable from one time to the next (abstract read).
• Ling et al. (2014) compared 72 raters under four shift patterns on about 14,000 spoken test responses. Shorter shifts and sessions scored more accurately and consistently, and three two-hour sessions in a day did best (abstract read).
• In Zhao et al. (2017) an essay read after nine high-scoring essays was expected to score 1.79 points in 20 higher than after none, and 81 of the 88 raters showed the pull. Spear (1997) found the opposite among 336 teachers marking three pieces of differing quality: good work scored higher after weaker work than before it (abstract read).

Our advice either way: mark one question across the class, change the order of students between questions, stop after two hours, and re-mark your first five at the end.

What do exam boards and universities ask for?

A meeting before marking, a checked sample after it, and a record.

• AQA (n.d.) requires internal standardisation for coursework marked in school, normally the head of department's to oversee. As good practice, all teachers mark some sample pieces and discuss the differences.
• The New Zealand Qualifications Authority (n.d.) has a subject specialist verify enough of each teacher's grade judgements, picked from work around grade boundaries and not at random, with the evidence kept to the end of the following academic year.
• The University of Bristol (n.d.) has a moderator judge a sample across the range of marks for each marker. "Adjustments should not be made to individual marks in isolation." Where marker and moderator cannot agree, a second moderator adjudicates.

What happens when a machine is the second marker?

It can check a human marker, and published practice does not let it decide alone.

Williamson (2026) restates that Ofqual's rules for regulated qualifications in England do not allow AI to be the only thing that decides a student's mark. It warns of anchoring and uncritical review where a person reviews marks an AI has already produced; anchoring is avoided where the AI marks in parallel and only discrepancies are flagged.

In Fox et al. (2026) the best model in each subject agreed with the two examiners slightly more closely than they agreed with each other: 0.75 against 0.65 in English, 0.86 against 0.84 in maths. Two examiners are a thin baseline.

So decide your own points first, ask the machine second, and read again each answer where the two differ. Where such readers fail is in our guide to checking the marks a machine proposes.

How do you run moderation in one department meeting?

Agree the scheme on five scripts, mark apart, compare a blind sample, settle the differences by the scheme, and write down what was decided.

1. Write the scheme as elements with points before the quiz. In EdProcto a short answer question holds a model answer and a rubric, and both show under each answer on the grading page. Set "Show the score" to "When released".
2. Standardise on five scripts. Each marker writes points for the same five answers without conferring. Compare element by element, and write each ruling into the rubric.
3. Draw the sample, and let the second marker mark it first, so no first mark is in view.
4. Mark apart.
5. Compare the sample, counting the gaps and their direction.
6. Settle by the scheme, and apply each decision to every script.
7. Write the record, then release. One page, kept with the marks: the date, the question and the markers; each ruling; how the sample was drawn; both sets of marks on it; the decision, and how many marks it changed.

How big a sample should be second-marked?

Eight scripts or a tenth, whichever is greater, and all of them when there are seven or fewer: the guide figure of Bristol (n.d.). Choose across the range of marks, add each script within a mark of the pass mark, and sample every first marker.

A sample this small shows direction, not rare slips. Our arithmetic: if two markers held one standard, a gap would fall either way equally often, so five gaps all one way happen by chance one time in 16, and six one time in 32.

What do you do with a disagreement?

Find the element of the scheme on which the two marks part, and rule on that.

• The scheme decides it. Correct every script on which that marker gave or refused the element.
• The scheme is silent. Write the ruling in, then re-mark that element for the whole class.
• No agreement. A third marker reads the script blind, or the head of department rules.

Large examinations do average: in Zhao et al. (2017) the final score was the mean of two raters. A department in one room can do better, because an average moves only the sampled scripts and leaves the cause in the rest.

How much moderation do the stakes need?

As much as a wrong mark would cost the student.

Recorded answers take a second listener: see the oral exam's recorded form.

What does moderation look like for two markers and one class?

One meeting, eight scripts read twice, and eleven marks changed. The example is invented.

A science unit test runs in EdProcto: twelve rule-marked questions for 24 points and one written explanation, question 13, for 6. Twenty-eight students sit it; the pass mark is 60%, 18 points of 30. The class teacher owns the quiz and marks first; a colleague marks second.

1. Standardise. At one screen each writes points for question 13 on five students' answers. They differ once, and rule that a missing unit loses nothing.
2. Sample. Eight other students, chosen across the rule-made scores already listed. The second marker reads question 13 for each while it shows "to grade", writes points in the department's sheet and saves nothing.
3. Mark apart. The first marker marks all 28, pressing "Save my grade" each time, then "Download the marks".
4. Compare. The second marker's figures go beside the column "Q13 (6)".

script   1  2  3  4  5  6  7  8
first    4  3  5  2  4  6  3  5
second   3  3  4  1  4  5  2  4
gap     +1  0 +1 +1  0 +1 +1 +1

5. Settle. The first marker is higher on six of eight and never lower. All six gaps sit on one element, "names the variable that was kept the same". The first marker had credited "we kept it fair". The scheme asks for the variable by name, so the phrase earns nothing, and the rubric now says so.
6. Apply. The first marker reopens question 13 for all 28. Eleven answers had the phrase credited, six in the sample and five outside it, and each loses the point. One student falls from 18 points to 17, from 60% to 56.7%, so both markers read that answer together.
7. Record and release. A second download, the record, then "Release the scores".

Averaging would have taken half a point from the six sampled students and left the other five the whole point: one sentence, two marks.

What does EdProcto give the markers?

EdProcto is the best tool for a small department's moderation because the scheme travels with the marking: the rubric corrected in the meeting shows under every written answer marked afterwards, and the scores can wait until the markers agree. In the example, that is how one ruling reached all eleven answers.

• The queue. In the list "Students" on the results page, the button beside a student reads "Grade" while answers wait.
• The grading page. One student's submission, headed with the student's name, so a second marker is blind to the first mark and not to the student. A written answer shows "to grade" with your "Model answer" and "Rubric" under it, a "Points" box in half points, "Note for the student" and "Save my grade".
• Who made a mark. A quiz belongs to the account that made it, so two markers mark from it in turn, dividing the class by student. A mark is stored with its points, its note, who decided it (the rules, Procto AI, or a person by hand) and, for a grade, when it was given. Each answer's mark is saved by itself, so two markers on different answers never undo one another. A hand mark is recorded alike whoever made it, and on one answer the latest mark is the one kept, so the department's sheet names the marker and a download before the meeting keeps the first marks.
• Procto AI. On a typed open answer, "Ask Procto AI" sends the question, any passage, the maximum points, your model answer, your rubric and the answer, never the student's name. Points and a note come back, saved at once with the note headed "Procto AI": a machine's reading against your rubric, yours to keep or change.
• Changing marks. "Add a note, or set the points by hand" opens the same boxes on a rule-marked answer. The card for "Re-mark this quiz" says: "A mark you or Procto AI gave is kept."
• The file. "Download the marks" gives a spreadsheet file with a row for each attempt handed in: "Name", "Email", "Sitting", "Started (your time)", "Handed in (your time)", "Score %", "Points", "Out of", "Reduced by %", "To grade", "Minutes", "Security events", "Notes on the record", then a column a question. The answers themselves are read on the grading page, so the sample is marked there.
• Release. With "Show the score" at "When released", the results page says "No student has seen a score or a mark." until you press "Release the scores".

Free takes 30 students a quiz, Tutor 100 and Educator 500, with the full analysis on Educator. Procto AI's grading is from the Tutor plan: see pricing. A summary is moderated the same way.

Quick questions

What is the difference between moderation and double marking?

Double marking gives every script two independent marks. Moderation judges the whole set from a sample, after marking.

How many scripts should a second marker read?

Eight or a tenth, whichever is greater, and all when there are seven or fewer, following the University of Bristol (n.d.).

Can two teachers mark the same quiz in EdProcto?

Yes, in turn, from the account that owns the quiz. In EdProcto the second marker reads the sample while its answers show "to grade" and writes points in the department's sheet.

Can Procto AI be the second marker?

As a parallel check only. Its mark is a machine's reading against your rubric, yours to keep or change, so decide your own points before you press "Ask Procto AI".

References

• AQA. (n.d.). Before submitting marks/grades. https://www.aqa.org.uk/exams-administration/coursework-controlled-assessment-nea/before-submitting-marks-and-grades
• Baird, J.-A., Greatorex, J., & Bell, J. F. (2004). What makes marking reliable? Experiments with UK examinations. Assessment in Education: Principles, Policy & Practice, 11(3), 331–348. https://doi.org/10.1080/0969594042000304627
• Cannings, R., Hawthorne, K., Hood, K., & Houston, H. (2005). Putting double marking to the test: A framework to assess if it is worth the trouble. Medical Education, 39(3), 299–308. https://doi.org/10.1111/j.1365-2929.2005.02093.x
• Fox, M., Samra, K., & Jung, P. (2026). LLM performance on a real, double-marked GCSE benchmark (arXiv:2606.24973). arXiv. https://arxiv.org/abs/2606.24973
• Garry, J., McCool, M. A., & O'Neill, S. (2005). Are moderators moderate? Testing the "anchoring and adjustment" hypothesis in the context of marking politics exams. Politics, 25(3), 191–200. https://doi.org/10.1111/j.1467-9256.2005.00243.x
• Leckie, G., & Baird, J.-A. (2011). Rater effects on essay scoring: A multilevel analysis of severity drift, central tendency, and rater experience. Journal of Educational Measurement, 48(4), 399–418. https://doi.org/10.1111/j.1745-3984.2011.00152.x
• Ling, G., Mollaun, P., & Xi, X. (2014). A study on the impact of fatigue on human raters when scoring speaking responses. Language Testing, 31(4), 479–499. https://doi.org/10.1177/0265532214530699
• New Zealand Qualifications Authority. (n.d.). Internal moderation. https://www2.nzqa.govt.nz/ncea/ncea-for-teachers-and-schools/managing-national-assessment-in-schools/moderation/internal-moderation/
• Spear, M. (1997). The influence of contrast effects upon teachers' marks. Educational Research, 39(2), 229–233. https://doi.org/10.1080/0013188970390209
• University of Bristol. (n.d.). Regulations and code of practice for taught programmes: 21. Marking, moderation and anonymity. https://www.bristol.ac.uk/academic-quality/assessment/regulations-and-code-of-practice-for-taught-programmes/marking-and-moderation/
• Williamson, J. (2026, January 14). Principles of AI use in marking (Ofqual/26/7292/1). Ofqual. https://www.gov.uk/government/publications/principles-of-ai-use-in-marking/principles-of-ai-use-in-marking
• Zhao, H., Andersson, B., Guo, B., & Xin, T. (2017). Sequential effects in essay ratings: Evidence of assimilation effects using cross-classified models. Frontiers in Psychology, 8, Article 933. https://doi.org/10.3389/fpsyg.2017.00933