Evaluating AI Marking Accuracy in A-Level Economics: A TeachEdge Pilot Study

Comparing TeachEdge's marking of 67 student answers with published Pearson Edexcel examiner awards

Published by: TeachEdge Software Ltd
Study conducted: 14 October 2026
Subject: Pearson Edexcel A-Level Economics A (9EC0)
Study type: Company-conducted pilot evaluation

Key findings

How closely can AI marking reproduce the judgements of experienced examiners?

To investigate this question, we used TeachEdge to mark 67 student answers drawn from Pearson Edexcel A-Level Economics A examiner reports. We then compared TeachEdge's marks with the published examiner awards.

The answers covered 39 different examination questions, worth between 4 and 25 marks. They included handwritten responses, diagrams, economic analysis and extended evaluation.

TeachEdge awarded exactly the same mark as the examiner on 46.3% of answers and came within two marks on 82.1%.

Agreement was particularly close on shorter questions. Results were more variable on extended responses, where examiners must assess the quality of economic reasoning, analysis and evaluation.

We also compared our findings with Ofqual's published research into human examiner marking consistency. For questions worth more than 20 marks, TeachEdge's marking differences had a standard deviation of 3.11 marks, compared with approximately 4.0 marks for Pearson's human examiners in Ofqual's separate study.

This is an encouraging contextual comparison, although the studies used different answers and cannot establish that TeachEdge is more accurate or consistent than human examiners.

The pilot also identified limitations. TeachEdge tended to award higher marks to weaker answers and slightly lower marks to stronger ones. In some cases, it appeared too generous when assessing underdeveloped evaluation; in others, it applied stricter requirements than those reflected in published examiner judgements.

These findings provide evidence of TeachEdge's ability to reproduce published examiner awards while identifying areas requiring further investigation.

This was a deliberately selected, relatively small sample, not a representative assessment of all A-Level Economics answers. The study was conducted by TeachEdge and has not been independently reviewed.

---

1. Purpose and scope of the study

TeachEdge provides AI-assisted assessment and feedback for UK secondary schools, including GCSE, A-Level and Key Stage 3 subjects.

Its marking instructions are designed to reflect the assessment requirements of different qualifications, examination boards and question types.

We wanted to investigate three questions:

1. How closely do TeachEdge's marks agree with published examiner awards?
2. Does agreement differ according to question length and the quality of the student's response?
3. Where TeachEdge disagrees with the examiner, can we identify possible reasons?

The study focused on Pearson Edexcel A-Level Economics A, using student answers and examiner awards published in official examiner reports.

These published awards provide an established reference against which TeachEdge's marking can be assessed.

However, agreement with a published examiner mark is not necessarily the same as marking accuracy. Examiner judgements can differ, and a published award is not necessarily beyond question.

Throughout this report, agreement means agreement with the published Pearson examiner award.

Who conducted the study?

TeachEdge selected the answers, conducted the marking exercise, analysed the results and prepared this report.

The published examiner awards were established before the study took place.

We did not commission independent teachers or examiners to mark the same answers, and this report has not undergone independent peer review.

These limitations are relevant when interpreting the findings.

2. Methodology

2.1 Selection of questions and student answers

We selected student answers from official Pearson Edexcel examiner reports, using the accompanying question papers and mark schemes.

The questions covered microeconomics, macroeconomics and questions combining the two.

We initially identified 54 suitable answers. Before marking began, we added another 13 to broaden the coverage of lower- and middle-scoring responses.

The final sample contained 67 answers to 39 distinct questions.

The sample included 65 handwritten answers and two printed answers.

We grouped answers according to the proportion of available marks awarded by the examiner.

These are descriptive categories used for this analysis, rather than examination grades or formal examiner levels.

The sample contained no zero-mark answers.

It was not randomly selected. Examiner-report examples are themselves chosen to illustrate aspects of marking, while our selection was designed to provide coverage of different questions, mark allocations and levels of performance.

Consequently, the results should not be treated as representative of all A-Level Economics responses.

2.2 Preparing the answers for marking

We obtained the student responses from the original published scans.

Handwriting, diagrams, crossings-out and other features of the original work were preserved. We did not transcribe, rewrite or improve the answers.

All 147 image sections were checked against the originals.

For each answer, TeachEdge received:

• The examination question.
• The official mark scheme.
• Relevant extracts, data or figures.
• The student's original response.
• TeachEdge's normal marking instructions for that question type.

The published examiner mark and comments for the answer being tested were excluded from the material provided to TeachEdge.

The marking instructions included existing calibration examples drawn from other sources.

2.3 Preventing test contamination

None of the 39 questions had previously been used by TeachEdge to develop or calibrate the marking instructions used in this study, according to the development records available.

We checked the test materials against those records before marking.

However, Pearson's published examination materials are publicly available. We cannot establish whether any of them were included in the underlying AI model's original pre-training data.

The study therefore avoids known reuse within TeachEdge's own calibration process, but cannot rule out exposure during foundation-model training.

2.4 Keeping the marking conditions consistent

Before testing, we fixed the nine relevant sets of TeachEdge marking instructions and the AI model settings.

These remained unchanged throughout the exercise.

Each answer was marked once.

We did not:

• Retry answers that received unexpected marks.
• Change the marking instructions during testing.
• Switch AI models or technical settings.
• Repair individual results.
• Select a preferred result from several attempts.

All 67 marking outcomes were retained.

The precise technical settings and checksums identifying the marking-instruction versions are recorded in the supporting methodology documentation.

The exercise used TeachEdge's marking instructions and software for preparing marking requests in a separate study setup. It was not an end-to-end test of the live TeachEdge website.

The findings therefore relate to the particular marking configuration tested, rather than every TeachEdge feature or setting.

2.5 Measuring agreement

Before testing began, we established the principal measures used to compare TeachEdge's results with published examiner marks.

These included:

Exact agreement: Whether TeachEdge awarded precisely the same total.

Agreement within one or two marks: Whether the absolute difference between the two awards was no greater than one or two marks.

Mean absolute error (MAE): The average size of the marking differences, ignoring whether TeachEdge marked higher or lower.

Mean signed difference: The average of TeachEdge's mark minus the examiner's mark. Positive figures indicate more generous marking; negative figures indicate stricter marking.

Relative absolute difference: The absolute marking difference expressed as a percentage of the marks available.

We also examined performance across different question lengths and the three score groups.

Agreement within two marks is a useful descriptive measure, but it is not a claim that every two-mark difference is educationally acceptable. A two-mark discrepancy matters differently on a four-mark question and a 25-mark essay.

2.6 Comparing assessment levels

After the first 12 results had been recorded, we added an analysis of agreement on assessment levels.

For extended answers, we examined:

• Knowledge, application and analysis (KAA).
• Evaluation.

We included an examiner level only where it was explicitly published or could be established without ambiguity from the published marks and mark scheme.

This analysis was supplementary to the original total-mark comparisons.

It covered 26 answers, rather than the complete set of 67.

---

3. Results

3.1 Overall agreement with published examiner marks

All 67 answers produced a usable marking result on the first attempt.

We checked that the marks were within the permitted ranges and that component marks added up to the reported totals.

No results were missing or unusable.

These checks confirmed that the results could be compared, rather than independently establishing the correctness of the marks.

TeachEdge matched the examiner's total on almost half the answers and came within two marks on more than four-fifths.

The mean signed difference of +0.28 marks indicates a small overall tendency towards more generous marking.

However, this average conceals differences in both directions. Higher and lower marks can cancel each other out, so the mean absolute error provides a better indication of the typical size of disagreement.

The relative measure also matters because the questions carried different maximum marks.

3.2 Agreement by question length

The extent of agreement varied according to the marks available.

Agreement was closest on shorter questions.

For questions worth four, five or eight marks, all 34 answers were within two marks of the published award.

For the 13 essays worth 25 marks, eight were within two marks and one matched exactly.

The average absolute difference on 25-mark essays was 2.46 marks, equivalent to 9.85% of the marks available.

These results suggest that TeachEdge reproduced published awards particularly closely on shorter responses, while extended questions generated greater disagreement.

This is consistent with the greater scope for judgement when assessing sustained economic analysis and evaluation.

However, a larger absolute marking difference is also more likely on questions carrying more marks. The results should therefore be interpreted in relation to the question tariff, rather than treating all absolute differences as equivalent.

The overall agreement figure is also influenced by the composition of the sample. Twenty-four of the 67 answers were worth only four or five marks, where the possible numerical differences are relatively limited.

3.3 Did TeachEdge tend to mark higher or lower?

The direction of marking differences also varied by question length.

TeachEdge tended to mark more generously on most question lengths, but slightly more strictly on 25-mark essays.

Across the full study, 12 answers differed from the examiner by more than two marks.

Six of these differences were positive and six were negative.

The apparent overall balance therefore conceals some systematic patterns, particularly when answers are grouped according to the examiner's award.

3.4 Agreement across lower-, middle- and higher-scoring answers

This analysis reveals an important pattern.

TeachEdge awarded higher marks than the examiner on ten of the eleven lower-scoring answers and matched the remaining answer.

For higher-scoring work, it tended to award slightly lower marks.

All six differences greater than two marks in TeachEdge's favour occurred in the lower- and middle-scoring groups.

All six differences greater than two marks in the opposite direction occurred in the higher-scoring group.

In this sample, TeachEdge tended to draw marks towards the middle of the available range.

This pattern deserves attention because it may affect how accurately the system distinguishes between weaker, competent and outstanding responses.

The closer examination of individual disagreements, discussed below, suggests possible explanations.

The original selection of 54 answers had a mean absolute error of 1.11 marks, with 45 answers within two marks.

The 13 additional answers, selected before marking began to increase the representation of weaker and middle-scoring work, had a mean absolute error of 1.62 marks, with ten within two marks.

TeachEdge's mean signed difference was zero in the original group and +1.46 marks in the additional group.

This reinforces the importance of testing across a broad range of student performance.

3.5 Agreement on assessment levels

Total marks provide only part of the picture.

For extended responses, an examiner's judgement also concerns the quality of different aspects of the student's work.

We therefore examined agreement on KAA and evaluation levels wherever the published material allowed these to be identified reliably.

TeachEdge matched both the KAA and evaluation levels on 14 of the 26 answers (53.8%).

Agreement was stronger for KAA than for evaluation.

For KAA, TeachEdge awarded a higher level on three answers and a lower level on three.

For evaluation, it awarded a higher level on seven answers and a lower level on three.

The largest evaluation-level disagreement involved an answer where the examiner awarded no evaluation credit but TeachEdge awarded Level 2.

This case also contained an inconsistency in the published examiner commentary.

Six answers differed in at least one assessment level even though their total marks were within two of each other.

This illustrates an important limitation of relying exclusively on total-mark agreement: two markers may produce similar totals while reaching different judgements about the underlying quality of the answer.

The results suggest that evaluation is a particularly important area for further testing and refinement.

Of the 26 answers included in the assessment-level analysis, 19 had levels or component marks explicitly provided in the examiner reports. The remaining seven had levels that could be inferred unambiguously from the published totals and the relevant mark-scheme boundaries.

3.6 Sensitivity to unclear source information

Before testing, we identified three answers where the published material contained a potential ambiguity.

Two had inconsistent level labels, while one required the total mark to be calculated from its components.

We retained all three in the main analysis.

As a sensitivity check, we repeated the calculations excluding them.

Excluding these cases made little difference to the overall results.

No additional answers were removed after examining TeachEdge's marks.

---

4. Investigating the marking disagreements

Statistical agreement is useful, but it does not explain why different marks are awarded.

We therefore examined 18 answers in greater detail.

These comprised the 12 answers where TeachEdge's mark differed by more than two marks, plus six additional answers where its KAA or evaluation level differed from the examiner's.

For each, we considered the original response, the official mark scheme, the published examiner commentary, TeachEdge's marking instructions and its feedback.

This closer review was AI-assisted and could see both the published examiner award and TeachEdge's award. It was not a blind or independent re-marking exercise.

Three recurring issues emerged.

4.1 Some strong essays appeared to be marked too strictly

Two of the largest differences involved strong 25-mark essays.

Both answers contained substantial economic reasoning and evaluation.

TeachEdge also identified genuine weaknesses.

However, its marking instructions required a weighed overall conclusion for top-level evaluation.

In the published commentary, the examiners awarded top-level evaluation despite noting weaknesses in the final judgement or the absence of a fully developed overall conclusion.

This suggests that TeachEdge may sometimes have applied a stricter interpretation of evaluation requirements than the published examiner decisions.

The issue is important because a good marking system should reflect how the examination criteria are actually applied, rather than imposing additional requirements that are not consistently expected by examiners.

We have not yet established how much of these differences can be attributed specifically to the wording of TeachEdge's instructions.

That would require further controlled testing.

4.2 Some weaker answers received generous evaluation marks

The opposite pattern appeared in several lower-scoring responses.

In some cases, TeachEdge appeared to award evaluation credit for identifying a limitation without adequately developing or explaining it.

For example:

Infrastructure question: The examiner awarded 8/15, while TeachEdge awarded 12/15. TeachEdge placed the response one level higher for both KAA and evaluation.

Energy-regulation question: The examiner awarded 4/10, while TeachEdge awarded 7/10. TeachEdge identified a weakness in the economic explanation but nevertheless awarded Level 2 evaluation for an argument that relied upon it.

These examples suggest that TeachEdge may not always distinguish sufficiently between:

• Identifying a relevant evaluative point.
• Explaining the point using sound economic reasoning.
• Developing a convincing, well-supported evaluative argument.

This distinction is particularly important in Economics, where effective evaluation normally requires more than mentioning a limitation or alternative perspective.

The relevant responses can be located in the 2022 Paper 2 examiner report, PDF page 41 and the 2018 Paper 1 examiner report, PDF page 39.

4.3 Some published examiner comments were open to interpretation

Not every disagreement could be attributed straightforwardly to TeachEdge.

One answer received 3/12 from the examiner and 9/12 from TeachEdge.

The published commentary stated that the answer did not link its argument to revenue, although the student referred to an increase in sales value.

That reference does not necessarily constitute a sufficiently developed explanation or demonstrate that TeachEdge's higher mark was justified.

It does, however, illustrate the difficulty of interpreting some published comments without further clarification.

In another case, the published commentary labelled four KAA marks as Level 3, although the relevant mark scheme placed four marks in Level 2.

We retained the published numerical mark and disclosed the inconsistency.

These examples reinforce the distinction between agreement with a published award and establishing the objectively correct mark.

They also illustrate why some differences remain open to reasonable interpretation.

---

5. How does this compare with human examiner marking?

A relevant question is whether the differences observed in TeachEdge's marking are large or small compared with differences between human examiners.

Human marking is not perfectly consistent. Even experienced examiners can reach different judgements when applying the same mark scheme, particularly to extended responses requiring qualitative assessment.

Ofqual has investigated this issue through its published research into marking consistency.

5.1 Ofqual's marking consistency research

Ofqual's report, Marking consistency studies: Summer 2016 and 2017 units, examined differences between examiners marking the same scripts.

The research included A-Level Economics examinations from AQA and Pearson.

For the 2017 study, groups of examiners marked examination scripts independently. Their marks were compared with those awarded by a principal examiner, who provided the reference mark.

Ofqual examined both the average direction of marking differences and their variability.

Figure 10 of the report, on PDF page 22, presents these results for individual Economics questions, grouped by the marks available.

The figure separates Pearson and AQA results. For our comparison, we used the Pearson figures because our own study concerned Pearson Edexcel Economics. GOV.UK

5.2 Comparing the marking differences

We calculated equivalent measures of the direction and spread of TeachEdge's differences from published examiner awards.

The Ofqual values below are approximate readings from Figure 10, rather than exact figures supplied in a data table.

Differences are calculated as the marker's award minus the reference examiner's award. A positive figure indicates more generous marking. Standard deviation measures how widely individual marking differences vary around their mean.

The findings are mixed but encouraging.

For questions worth 7–10 marks, TeachEdge's marking differences had a smaller standard deviation than those reported for Pearson's human examiners.

For questions worth 11–20 marks, TeachEdge's differences showed greater variability.

For questions worth more than 20 marks, TeachEdge's standard deviation was 3.11 marks, compared with approximately 4.0 marks for human examiners in the Ofqual study.

For these longer answers, TeachEdge's differences from published awards were therefore less widely dispersed around their average than the human examiner differences in Ofqual's separate research.

This is a potentially important finding, given the judgement involved in assessing extended Economics essays.

However, it requires careful interpretation.

5.3 Why these results cannot establish that AI marking is better than human marking

The studies were not conducted on the same answers.

Ofqual's research involved human examiners marking a common set of scripts, with differences measured against the principal examiner's marks.

Our study involved TeachEdge marking answers selected from published examiner reports, with differences measured against the published awards.

The samples may have differed in their difficulty, quality, composition and assessment requirements.

The studies also differed in scale and marking arrangements.

Furthermore, a smaller standard deviation does not necessarily indicate greater accuracy.

A marker could award consistently higher or lower marks than the reference examiner and still have a relatively small standard deviation.

Both the average difference and its variability therefore matter.

Ofqual's report also notes that its marking exercise differed from live examination marking. Examiners were working in a separate research environment without all the monitoring and support arrangements associated with live marking. GOV.UK

For these reasons, the comparison should be regarded as a contextual benchmark, not a controlled test of TeachEdge against human examiners.

Nevertheless, it provides useful perspective.

TeachEdge's marking differences, particularly on 25-mark questions, were of a magnitude that merits further investigation alongside published evidence of human examiner variability.

A direct comparison would require TeachEdge and a group of human examiners to mark the same previously unseen answers independently, using the same assessment criteria.

---

6. Limitations of the study

The results are encouraging, but several limitations affect the conclusions that can reasonably be drawn.

6.1 Sample size and selection

The study involved 67 answers across 39 different questions.

This is sufficient to provide preliminary evidence and identify possible patterns, but it is not a large-scale validation exercise.

The answers were deliberately selected rather than randomly sampled.

Published examiner-report examples are designed to illustrate marking decisions and may not represent the full range of responses encountered in schools or live examinations.

The sample also contained substantially more higher-scoring answers than lower-scoring ones.

Some questions were represented by multiple answers, meaning that the 67 responses were not independent tests across 67 different questions.

Several subgroups were small. For example, only five answers were worth 15 marks.

Apparent differences between groups should therefore be treated as exploratory rather than definitive.

6.2 No independent marking of the same answers

We used existing published examiner awards as the reference.

We did not arrange for additional human examiners or teachers to mark the same responses.

This means the study measures agreement with the published awards but does not establish whether TeachEdge would perform better or worse than human examiners assessing the same material.

It also cannot conclusively resolve disagreements between TeachEdge and the published examiner.

6.3 Possible exposure to published examination materials

We checked that the questions had not previously been used to develop or calibrate TeachEdge's marking instructions.

However, we cannot verify the complete contents of the underlying AI model's pre-training data.

Published examination materials may have been available during that training.

Although we excluded the relevant examiner marks and comments from each marking request, we cannot rule out the possibility that the model had previously encountered some of the material.

This is an inherent limitation when evaluating general-purpose AI models using publicly available examination content.

6.4 One marking attempt per answer

Each answer was marked once.

This prevented selective retries from influencing the results, but it also means we did not measure repeatability.

AI systems can sometimes produce different outcomes when given the same input on separate occasions.

Further testing would be needed to establish how consistently TeachEdge awards the same marks when an answer is submitted repeatedly.

6.5 Assessment-level information was incomplete

The comparison of KAA and evaluation levels covered only 26 answers.

For the remaining responses, published information was insufficient to establish the examiner's level unambiguously, or the question did not provide a comparable level breakdown.

The assessment-level results should therefore be interpreted separately from the overall mark comparisons.

6.6 The study did not assess feedback quality or educational impact

TeachEdge generates feedback as well as marks.

This study principally assessed the agreement of numerical marks with examiner awards, with supplementary analysis of assessment levels.

It did not systematically examine whether TeachEdge's written feedback:

• Correctly identified the student's principal strengths and weaknesses.
• Provided appropriate and actionable advice.
• Helped students improve their subsequent work.
• Reduced teacher workload.
• Improved learning outcomes.

These are important questions, but they require separate evaluation.

6.7 The findings apply to the configuration tested

The results concern the particular marking instructions, AI model and technical settings used on 14 October 2026.

Changes to instructions or model behaviour may alter future marking outcomes.

The study used a separate marking setup rather than testing the complete live TeachEdge platform.

Although the source answers retained their original handwritten and visual features, published examiner-report scans may be clearer than some images uploaded by students or teachers in everyday use.

The findings should not be generalised automatically to other subjects, examination boards, question types or TeachEdge tools.

---

7. Implications for schools

The findings provide preliminary evidence that TeachEdge can reproduce published examiner marks with considerable agreement across a range of A-Level Economics questions.

In this sample, 46.3% of answers received exactly the published mark and 82.1% were within two marks.

Agreement was particularly close on shorter questions.

Extended answers produced greater differences, although the variability observed on 25-mark essays compared encouragingly with figures from Ofqual's separate research into human examiner marking.

For schools, three implications are worth considering.

7.1 AI marking can provide a useful starting point for assessment

The results suggest that TeachEdge's marking can provide a useful initial assessment of student work, particularly where teachers need to assess substantial numbers of answers regularly.

This could support more frequent practice and individual feedback without requiring teachers to mark every response from scratch.

However, the study does not establish that TeachEdge marks are interchangeable with those of experienced teachers or examiners.

Teacher oversight remains important, particularly where marks inform significant decisions about student performance.

7.2 Total marks are not the whole story

The study found stronger agreement on knowledge, application and analysis levels than on evaluation levels.

This matters because a useful assessment must do more than produce a plausible total.

It should also identify where the student has demonstrated understanding, developed analysis and evaluated effectively.

A system that produces a similar total through different assessment-level judgements may still provide feedback that requires teacher review.

Our findings suggest that evaluation deserves particular attention.

7.3 Performance should be understood across different kinds of work

The study identified differences according to question length and the examiner's assessment of answer quality.

TeachEdge was particularly close to published awards on shorter responses but showed more variation on extended work.

It also tended to award higher marks to weaker answers and lower marks to some stronger ones.

Schools should therefore consider the nature of the work being assessed rather than relying exclusively on a single overall accuracy figure.

---

8. Next steps

This was designed as a pilot study.

Its purpose was not only to measure agreement with published examiner marks, but also to establish a repeatable approach for investigating TeachEdge's performance.

The findings suggest several priorities.

8.1 Investigating evaluation judgements

Further testing should examine how TeachEdge applies evaluation criteria, particularly where students:

• Identify a relevant limitation without developing it.
• Present strong analysis but omit an explicit concluding judgement.
• Demonstrate evaluation throughout an essay rather than concentrating it in the conclusion.
• Offer plausible arguments that are not fully supported by economic reasoning.

These cases appear particularly relevant to the differences observed in this study.

Any changes to the marking instructions would need to be evaluated using new, previously untested answers.

Reusing the existing answers to measure improvement would risk overstating performance.

8.2 Testing a more representative range of answers

Future evaluations should include a larger and more balanced selection of weaker, middle-scoring and higher-scoring responses.

This would help establish whether the tendency towards the middle of the mark range persists across a broader sample.

It would also provide stronger evidence about performance on different question lengths.

8.3 Examining repeatability

Repeated marking of the same answers would provide evidence about the consistency of TeachEdge's outputs over time.

This would complement the present study's comparison with published examiner awards.

8.4 Extending the approach to other subjects

The methodology could be applied to other GCSE and A-Level qualifications.

Different subjects present different assessment challenges.

For example, extended historical arguments, literary interpretation, scientific explanations and mathematical reasoning involve different combinations of knowledge, analysis and judgement.

Separate subject-specific studies would provide a stronger basis for understanding where AI-assisted marking performs well and where further development is required.

8.5 Investigating feedback quality

A future study could examine whether experienced teachers agree with TeachEdge's identification of strengths, weaknesses and improvement priorities.

This would help assess the educational usefulness of AI-generated feedback, rather than focusing only on numerical marks.

---

9. Conclusions

This pilot study compared TeachEdge's marking of 67 Pearson Edexcel A-Level Economics A answers with the awards published in official examiner reports.

TeachEdge matched the examiner's mark exactly on 31 answers (46.3%) and came within two marks on 55 (82.1%).

The mean absolute marking difference was 1.21 marks.

Agreement was particularly close on shorter questions.

On 25-mark essays, TeachEdge's marking differences had a standard deviation of 3.11 marks. This was lower than the approximately 4.0 marks reported for Pearson's human examiners on questions worth more than 20 marks in Ofqual's separate marking consistency study.

Although encouraging, this comparison does not demonstrate that TeachEdge is more accurate or consistent than human examiners, because the studies used different answers and marking arrangements.

The pilot also identified shortcomings.

TeachEdge tended to mark weaker answers more generously and some stronger essays more strictly than the published examiner awards.

The closer review suggests that evaluation judgements were an important source of disagreement, particularly where an answer contained an underdeveloped evaluative point or lacked an explicit overall conclusion.

These findings provide useful evidence of TeachEdge's marking performance while identifying areas where further investigation is warranted.

The central conclusion is that TeachEdge demonstrated substantial agreement with published examiner awards in this selected sample, but further independent and larger-scale testing is needed to establish how closely its marking compares with human examiners more generally.

We regard this study as an initial contribution to that evaluation, rather than a definitive measure of AI marking accuracy.

---

10. Sources, supporting data and reproducibility

Pearson Edexcel examination materials

The study used the following official Pearson examination materials.

Ofqual research

Holmes, S., Morin, C., Cuff, B. and Black, B. (2018). Marking consistency studies: Summer 2016 and 2017 units. Ofqual.

Read the Ofqual report (PDF)

The relevant Economics item-level results appear in Figure 10, PDF page 22.

Supporting study records

The supporting study records comprise:

Download the results dataset (CSV): The examiner and TeachEdge marks for all 67 answers, assessment-level comparisons where available, and identifiers linking each answer to its published source.

Download the methodology record (JSON): The AI model settings, marking-instruction checksums, calibration checks and records of question selection and exclusions.

These records were retained during the study. The complete marking instructions and technical configuration have not been publicly released.

The published examiner awards can be checked against the Pearson reports, and the numerical comparisons can be recalculated from the results dataset.

However, reproducing the AI marking itself requires the complete marking instructions and technical configuration. Even under equivalent conditions, later model outputs may differ.

The individual numerical results are reproduced in the appendix below.

---

Appendix A: Individual results for all 67 answers

The table records the published examiner mark and TeachEdge mark for every answer included in the study.

Each source identifier provides the examination year, paper, question and exemplar number.

Where available, KAA and evaluation levels are also shown.

A dash indicates that the published information was insufficient to establish the examiner's assessment level unambiguously, or that the level comparison was not applicable. It does not represent Level 0.

In the mark columns, the examiner's award appears first, followed by TeachEdge's award.

---

This study was conducted and reported by TeachEdge Software Ltd. TeachEdge is independent of Pearson, Edexcel and Ofqual. The findings do not constitute an endorsement by any examination board or regulator.