For professors, admissions & large-scale evaluators
CERENO Eval
Assess the argument.
Inspect the evidence.
First-pass assessment designed around your marking scheme or criteria. A reviewable marking profile, evidence-linked proposals and attention directed to the cases that need your judgement.
Discuss an institutional pilotDifferent settings. The same evidential standard.
University professors
Review reasoning against the marking scheme, identify gaps and focus expert attention on answers that need closer examination.
Admissions teams
Organise evidence against your stated criteria and surface uncertainty. The admissions decision remains with your institution.
Large-scale evaluators
Examine consistency across reviewers, languages and cohorts, with an evidence trail that makes disagreements inspectable.
The institutional assessment tension
Four demands.
Fairness binds them together.
A consistent system can still measure the wrong thing. A deep reading can still be unfair. Credible assessment has to hold validity, reliability, scale and depth in tension—not maximise one and assume the others follow.
- Validity
- Are we measuring the intended capability—or its polished proxy? Pilot criteria must distinguish reasoning, knowledge and expression, and make clear whether the task assesses independent or assisted work.
- Reliability
- Would another reviewer reach a comparable judgement? Explicit criteria and evidence-linked proposals make disagreement visible for expert-marker comparison.
- Scale
- Can review remain workable at institutional volume? A first pass should help people focus their attention, but workload benefits have to be measured rather than assumed.
- Depth
- Can an unconventional but sound argument receive a fair hearing? Complex work needs room for context, uncertainty and human escalation—not just resemblance to a model answer.
Fairness is not a fifth box to tick.
Language and cohort disparities must be examined across all four demands. Reliable marks are not necessarily equitable marks, and a strong average can conceal a systematically worse result for a smaller group.
Break the synthetic loop
Assess the person’s reasoning.
Not the machines’ agreement.
When applicant AI generates a persuasive answer and evaluator AI rewards its familiar shape, a high mark can reflect alignment between machines rather than understanding. Institutions need to define what the work can legitimately demonstrate before deciding how to mark it.
Eval is designed to show the reasoning evidenced in the answer, not reward polish as a shortcut. The reviewer can inspect the chain from claim to evidence to proposed mark and decide where the record is insufficient.
Where independent thinking is the target, authentic process evidence matters: a supervised task, visible revision or a live explanation can help establish what the student can do. These are institutional assessment choices, not capabilities CERENO claims to infer from static text.
Where tools are permitted, the question changes: can the student judge, challenge and take responsibility for the output? A sound assessment makes that distinction explicit instead of treating every AI-assisted answer as either mastery or misconduct.
The evaluator’s view
A marking profile.
Not just a total.
The intended review experience separates the dimensions of an answer, so a single overall impression does not conceal where the work succeeds or falls short. Your marking scheme defines the weighting; admissions teams use their institution’s criteria.
Expression is assessed explicitly, rather than quietly determining the other dimensions. Reviewers can distinguish a strong argument written plainly from fluent prose with unsupported claims.
- Reasoning
- Inspect the connection between premises, steps and conclusion.
- Problem framing
- See whether the answer addresses the task that was set.
- Evidence & knowledge
- Review support for claims and the accuracy of the material used.
- Expression
- Assess clarity separately against the expectations of the task.
- Calibration
- Consider whether the student’s confidence is warranted by the evidence.
Evidence-linked scoring
See what supports the proposed mark.
An unexplained impression
Explicit marking rules
Start with your rules
The marking scheme or admissions criteria define what counts. The model’s preferences should not become the standard.
Inspect the supporting passage
For each proposed mark, reviewers should see the relevant criterion, the student’s exact words and the reason those words do—or do not—meet the requirement. A located passage is not necessarily sufficient evidence.
Challenge the proposal
The intended record connects the answer, criterion, evidence and reason. Reviewers should be able to locate the disagreement, record their judgement and change the proposed result.
Bias reports · language & cohort
An average can hide a disparity.
Translated first
Read as written
The pilot reporting agenda examines agreement with expert markers separately by validated language and cohort. A useful report should show the size and composition of the sample, where disagreement occurs, and whether particular groups experience systematically different proposed marks.
Reading the original answer avoids adding a translation step to the evidence trail. That does not establish fairness by itself: each language must be validated against agreed criteria before its coverage expands.
Reviewers need visibility into gaps and limitations, not a claim of zero bias. These reports are planned validation outputs; CERENO does not present completed performance results.
Language as a barrier
Validated language access
Human escalation
Know which cases need your attention.
Uncertain, contested and sensitive work belongs with a reviewer. The intended escalation record preserves the answer, the relevant evidence and the reason for the flag, so a person can investigate rather than simply accept a warning.
For the volume
For the hard calls
Taking over
Supporting judgement
Reviewers should be able to challenge the proposed reasoning, record their decision and change a mark. Unusual but potentially valid work deserves examination, not an automatic penalty.
Eval supports first-pass assessment and triage. Final grades and admission decisions stay with the institution. It is not an AI detector or a measure of potential from one answer.
Oversight and deployment responsibilitiesPilot onboarding
Start with a bounded assessment question.
Pilot partners are in discussion. Onboarding is intended to establish a controlled evaluation with your experts before any wider deployment—not switch on automated grading across an institution.
01
Agree the scope
Define the task, marking scheme or criteria, intended users, languages and the decisions the pilot will not make.
02
Set the evaluation
Agree permitted sample use, expert-marker comparisons, cohort reporting, human review and measurable acceptance criteria.
03
Review before expanding
Examine disagreements, limitations and reviewer workload together. Extend the pilot only when the agreed evidence supports doing so.