← Part of our AI marking platform
Mark learner work against your rubric
Apply consistent criteria across essays, coursework and structured tasks, then review AI-drafted marks and feedback before students see anything. Build or import the rubrics your department already trusts.
- Rubric builder & import
- Criterion-level feedback
- Batch cohort marking
- Moderation-ready exports
Criterion, evidence, suggested grade, assessor change
Criterion: Analysis of sources (Merit band: “selects relevant evidence and explains how it supports the argument”). AI quotes two sentences from the essay that name a source and link it to the claim, suggests Merit, and drafts a next-step comment. The assessor moves the band to Pass because the second source is misread, keeps the quoted evidence, and rewrites the comment before release.
Weak submissions still get criterion-level notes rather than a blank page; ambiguous criteria produce lower-confidence drafts that teachers should sample first; low-confidence OCR or incomplete uploads are flagged for human review rather than silent guessing.
Rubric marking is a core capability of our AI marking platform. Build criteria in the rubric builder, or read how to write a marking rubric.
What matters for rubric workflows
Rubric builder
Create or upload criteria, weightings and level descriptors your team already trusts.
Criterion-level feedback
See which bands were applied and edit comments per criterion before release.
Batch cohort marking
Run one rubric against a whole class upload in a single session.
Question paper mode
Cross-reference structured Q&A papers with mark schemes.
Targeted follow-up
Generate worksheets from rubric gaps for individual learners.
Moderation exports
Evidence packs for internal QA, MAT reviews and sampling.
Design criteria once, apply them everywhere
Build rubrics in Feedback Flows or import the schemes your department already uses. Shared libraries help MATs and federations keep descriptors consistent without forcing every subject into the same mould.
Marking extended writing against those rubrics? See AI essay marking. School rollout and DPAs: AI marking for UK schools.
Rubric builder: criteria, weightings and level descriptors.
Moderation that starts from shared criteria
When every marker works from the same rubric library, moderation meetings focus on genuine borderline decisions rather than reconciling incompatible comment styles. AI drafts reduce the fatigue-driven drift that appears late in a marking pile; teachers still own the final band.
Most teams begin with one assignment type (mock essays, unit coursework, or a vocational portfolio) and compare AI rubric drafts to a human-marked sample set. Once moderators are satisfied, batch marking scales across cohorts.
Worked marking evaluation: 4 scripts, 33 questions, real overrides
Rather than claiming a universal “99% accuracy” figure from a cherry-picked sample, we publish how our marking and verification pipeline behaves on real anonymised scripts, including where the AI flags uncertainty and where a human assessor overrides the draft.
Method & sample limits: 4 anonymised Level 3 Engineering scripts (132 question responses across 33 scheme items, 44 marks max per paper, mix of short-answer explanations, multi-step calculations, component sketches and plotted graphs). Each script was compared against experienced human assessor judgement. Because n = 4 scripts from one technical unit, this illustrates failure modes and review workflow rather than a universal statistical benchmark.
| Criterion / question type | Human assessor judgement | AI draft & confidence behaviour | Required assessor changes & review time |
|---|---|---|---|
| Defined calculations & short-answer theory (29 of 33 items per script) |
Full or partial credit awarded against explicit scheme values and tolerances. | Matched human marks across standard prose and numerical items; cited met/unmet scheme points clearly. | Spot-check only (~2–3 mins per script). No mark changes required on clear prose/calculation items. |
| Plotted linear graph with missing axis labels (Q1b, 2 marks) |
1 / 2 marks: credited correct rising line (425 Ω to 479.825 Ω) despite missing axis titles. |
Drafted 0.0 / 2.0 (conf 0.55) and raised graph/diagram_0_with_evidence + needs_human_verification. |
Assessor override (+1.0 mark, ~15s). Flagged automatically so the assessor inspected the cropped graph rather than trusting a 0/2 score. |
| Non-linear NTC thermistor curve (Q2a, 1 mark) |
1 / 1 mark: decreasing curve from ~30 kΩ at 0°C levelling off by 60–70°C. | Vision validator confirmed curve shape, but text-scheme reconciler held score at 0.0 / 1.0 with graph/diagram_0_with_evidence flag. |
Assessor override (+1.0 mark, ~10s). Demonstrates why visual curve questions require a human gate. |
| Component symbol with partial labelling (Q4a, 2 marks) |
1 / 2 marks: parallel-plate capacitor drawn, dielectric unlabelled. | Drafted 1.0 / 2.0 (conf 0.55) with needs_human_verification noting missing dielectric label. |
Confirmed without edit (~8s). AI identified both the present symbol and the missing label accurately. |
Total review time & takeaway: Reviewing a 33-question script took under 4 minutes per paper because the assessor focused on the 4 flagged visual/borderline items while clear text and calculation responses were already mapped to scheme points. Need a starting rubric to run your own 10-script evaluation? Download our UK rubric template (CSV) or read how to write a marking rubric.
Frequently asked questions about rubric marking
See rubric marking in action
Free trial credits; pay-as-you-go thereafter. No subscription required.