CERENO

The foundation

Inside CERENO

Reading an answer and judging its reasoning are different jobs.

Inspired by Fast and Slow thinking, CERENO separates rapid reading from deliberate reasoning. The foundation is a simple distinction: recognising a plausible answer is not the same as checking what it proves.

The Two Machines dilemma

Preventing two machines from negotiating over a proxy for a human mind.

An applicant uses AI to generate an essay that fits the expected criteria. An evaluator uses AI to screen the essay against those criteria. Each machine can become better at satisfying the other while neither establishes what the applicant actually understands.

The failure is not simply that the text is synthetic. It is that the system treats an artefact as proof of a capability. A student who submits an answer without understanding it, one who rebuilds an argument after challenge, and one who responsibly reasons with tools are doing different intellectual work.

CERENO’s design shifts attention from static polish to the reasoning trace demonstrated in the answer: how claims connect, what evidence supports them, where assumptions enter and whether confidence is warranted. Cerebrum locates that evidence; Nous checks what it justifies. A fluent summary cannot substitute for the trace.

That trace is still evidence in a text—not an authenticated record of someone’s cognitive process. To assess independent ability, institutions need appropriate conditions, visible drafting or revision, or a live follow-up where the student explains and defends their choices. CERENO supports inspectable judgement; it does not certify authenticity from a finished answer.

Read the essay · Grade Humans as Humans (Part 1 of 4) ↗

Cognitive science

From recognition
to reflection.

Read and recognise

Check and deliberate

Fast to read. Slow to judge. Never on trust.

Dual-process accounts of cognition distinguish relatively automatic processing from more deliberate, effortful reasoning. Daniel Kahneman’s Thinking, Fast and Slow brought this distinction to a wide audience; researchers including Jonathan Evans and Keith Stanovich examine how intuitive responses and reflective processes interact.

For assessment, the important insight is not that “fast is bad” or “slow is always right”. It is that an initial response can feel compelling without having been checked. Familiar phrasing, confident language and orderly presentation can create a favourable impression before the underlying argument has earned it.

CERENO translates that insight into an engineering discipline: one layer reads and gathers evidence; a separate layer examines whether the evidence supports the assessment. The second stage should be able to reject the first stage’s interpretation, not merely rewrite it more persuasively.

This is an architectural inspiration, not a literal model of two brain systems or proof that the product is accurate. Two AI stages can still share blind spots. Independence of checking must be tested against expert judgement, diverse work and real institutional use.

Evans & Stanovich · Dual-Process Theories of Higher Cognition, 2013 ↗

Why style bias corrupts AI grading

Fluency is a signal.
Not a warrant.

An impressive surface

A sound foundation

Polish isn’t proof.

A general-purpose AI judge receives both the substance of an answer and its presentation at once. Headings, confident phrasing, familiar vocabulary and orderly paragraphs can influence its overall impression even when the underlying reasoning has not changed.

When that impression becomes the mark, the system can substitute an easier signal—what a strong answer tends to look like—for the harder question of whether the argument actually meets the criterion. This is a validity problem: the assessment drifts from the intended skill towards surface features.

Non-native writers can face the reverse effect. Shorter sentences or unfamiliar rhetorical patterns may make valid reasoning look less sophisticated. Translating or polishing the answer first can then change the evidence that should have been assessed.

CERENO’s response is to separate retrieval from judgement. A located passage is only a candidate piece of evidence. Nous must check its context and its logical connection to the criterion. Neither attractive formatting nor an exact keyword match should settle that question.

This design aims to constrain style-driven judgement, not prove that bias has disappeared. The architecture still requires independent validation.

The architecture · Cerebrum + Nous

Every proposed mark has to survive a second reading.

Key one: read

Key two: reason

Every mark needs two keys.
01

The reading layer

Cerebrum

Cerebrum reads the original answer in its own language, where validated, and gathers passages relevant to each part of the marking scheme or admissions criteria. It preserves exact words rather than substituting a fluent summary for the evidence.

Its job is to locate what the student has written—not to infer abilities that never appear in the work. A missing piece of evidence should remain missing.

02

The reasoning layer

Nous

Nous returns to the original answer and checks the evidence. Does the quotation exist as recorded? Does the surrounding context change its meaning? Does it actually meet the criterion, or only mention the right vocabulary?

Only then does the design apply explicit marking logic and attach a reason and an indication of certainty to the proposed mark. A quote without a valid connection to the criterion is not sufficient.

03

The accountable judgement

The reviewer

A teacher or reviewer examines uncertain, contested and sensitive cases. They can inspect the original, challenge the reasoning and change the result. Work that follows an unfamiliar but valid route deserves attention rather than an automatic penalty.

Final grades and admission decisions remain outside CERENO’s role. The institution retains responsibility, including the possibility of follow-up assessment when text alone cannot answer the question.

Illustrative reasoning · not a product result

Mentioning the idea is not explaining it.

Suppose a marking scheme awards a point for explaining why a comparison needs a control group. An answer says, “We used a control group to make the experiment fair.”

The phrase is present. Is the causal explanation?

The reading layer can locate “control group” and quote the sentence. That is useful retrieval, but it is not yet evidence that the student understands how a control helps isolate the intervention’s effect.

The reasoning layer must check the actual requirement. If the scheme requires that explanation, a keyword match or polished sentence should not automatically earn the point. If the surrounding answer supplies it, that context should be considered instead.

A reviewer can then see the criterion, the quotation and the reason for the proposal. This makes disagreement inspectable: the dispute is about an explicit connection between evidence and a rule, not an opaque impression of quality.

The limits are part of the design

Evidence in the text.
Not access to the mind.

CERENO cannot establish who authored an answer, whether assistance was permitted or whether a student could reproduce the reasoning unaided. It is not an AI detector and does not judge a person’s potential from one piece of writing.

The two-layer architecture is designed to make assessment more inspectable. Whether it improves consistency, fairness and reviewer workload is a validation question. Those outcomes must be measured—not inferred from the architecture’s name.

How we approach validation