How Our IB Graders Work - Methodology & Accuracy

Frequently asked questions

How accurate are the MyRevisionAgent IB graders?

On held-out essays the Extended Essay grader lands within one grade band 96% of the time and matches the examiner's exact band 47% of the time, with a mean absolute error of 3.6 marks out of 34, on essays marked under the criteria in use until November 2026. The Internal Assessment grader lands within one band 92% of the time and exact 43% of the time, with a mean error of about 2.3 marks. Both figures come from 4-seed held-out cross-validation, so they measure performance on essays the models never saw during training.

Is a 47% exact-match rate good for grading?

Yes - it sits inside the range at which trained human examiners agree with each other. Independent studies of essay marking put examiner-vs-examiner exact-grade agreement at roughly 41-57%, and examiners routinely differ by a full band. Our exact-band rates (EE 47%, IA 43%) fall inside that human range, and being within one band 92-96% of the time is the stronger, more robust signal for how a grade is likely to land.

How were the models trained and validated?

The graders were trained on 593 examiner-marked Extended Essays across 10 subjects and 1,678 examiner-marked Internal Assessments across 26 subject-levels - over 2,000 real, IB-awarded essays in total. Every headline number is measured with 4-seed held-out cross-validation, stratified by subject: the corpus is split 80/20 four times with different seeds, the model is trained only on the 80% and scored on the unseen 20%, and the results are averaged. Nothing in the reported accuracy comes from essays the model trained on.

How do the graders compare to ChatGPT or Claude?

On the same held-out, IB-marked essays, our graders beat every frontier model we tested. On 116 held-out Extended Essays our model scored 49% exact and 97% within one band, versus 46%/93% for Claude Opus 4.8 and 28%/79% for GPT-5.5. On 376 held-out Internal Assessments our model scored 42% exact and 90% within one band, versus 32%/77% for Opus and 16%/54% for GPT-5.5 - and each model was prompted as a fair examiner with the subject and the maximum mark. A general-purpose LLM prompted as an examiner drifts off the IB scale; a model calibrated on real examiner marks does not.

Which model does each grader use?

The Extended Essay grader uses a gradient-boosted ensemble (XGBoost, random forest and gradient-boosted regression) over a selected set of features, because the examiner-marked EEs it learned from share a single /34 scale where non-linear feature interactions help; on the 30-mark criteria first examined in May 2027, its prediction is rescaled and averaged with a frontier LLM's marks on the new criteria, and Reflection is marked from the student's reflective statement. The Internal Assessment grader uses a regularised linear (Ridge) model with subject one-hot encoding across 50 features, because the IA spans 26 very different mark scales (from /14 Economics to /40 English) where a linear map generalises better. In both, a frontier LLM reads the essay against the rubric and extracts scored features, and the trained model turns those features into a predicted mark.

Is the predicted grade an official IB mark?

No. The predicted grade is a well-calibrated estimate to guide your redrafting, not an official or guaranteed IB result. Every prediction is shown with its own confidence range (for example '24 plus or minus 2') rather than a single false-precise number, because grading has genuine variance and re-submitting the same essay can produce a slightly different figure within the same range.

How accurate is the TOK essay grader?

The Theory of Knowledge grader uses a custom-trained model with TOK-specific feature extraction and scores your essay out of 10 against the current IB TOK assessment criteria, first examined in 2022. It has its own smaller corpus and does not yet carry the same published held-out cross-validation figures as the EE and IA graders, so we report it honestly as a well-calibrated estimate rather than attaching a specific accuracy percentage to it.

What are the limits of the graders?

The graders estimate a mark and give rubric-aligned feedback; they do not replace your teacher or the official IB assessment. Per-subject accuracy varies - subjects with fewer training essays (for example some English A and small-cohort IA subjects) are more variable and should be read as indicative. The model scores the writing in front of it, so it cannot judge authenticity of data or sources beyond an automated AI-generation and wrong-subject check. Predictions are estimates with error ranges, not certainties.