How Lumen Verifies Its Own AI Marking

Large language models rarely signal uncertainty. When asked to mark an IB extended essay, Claude or ChatGPT will return a band, a grade and a paragraph of feedback, all delivered in the same confident register regardless of how much genuine reasoning sits behind them. In practice, this is what hallucination means for a student: an ordinary-looking, confidently stated piece of feedback that happens to be wrong, with no obvious way to notice unless the correct answer is already known.

This matters for a tool whose purpose is to identify what should be corrected before a real exam. Several specific mechanisms therefore sit between a model's first response and what a student ultimately sees, each addressing a different way an AI mark could go wrong without anyone noticing.

Marking is grounded in the uploaded rubric

When a student uploads a task sheet or rubric, Lumen marks against those exact criteria rather than a general impression of what an IB assessment is likely to require. Where no rubric is available, Lumen falls back to standard IB criteria and states this plainly, rather than presenting the resulting mark as though it were grounded in something more specific.

The distinction matters in practice. A mark grounded in an actual rubric can point to something concrete, such as missing linkage to the command term in a particular criterion. A mark produced without one tends toward generic advice, such as a suggestion to explain reasoning more thoroughly, since the model has no specific standard to measure against.

The confidence label is calculated independently of the model

Every AI mark is accompanied by a reliability rating of high, medium or low. This rating is not generated solely by the model's own assessment of itself. Lumen first calculates a ceiling from objective factors unrelated to how confident the model's language happens to sound:

• the absence of an uploaded rubric limits the rating to medium at most, regardless of how confident the model's wording appears;
• a short submission limits the rating to medium, and a very brief submission of only a few lines limits it to low;
• the absence of a second model's review keeps the ceiling where it already stands.

The model's self-reported confidence can register at or below this ceiling, but never above it. A model that expresses strong confidence despite thin evidence will still be shown to the student as low-confidence, because the ceiling is set by the available evidence rather than by the tone of the model's language.

A second, independent model reviews the first mark

On paid plans, a second examiner reviews the first mark, drawn from a different provider rather than the same model reviewing its own output. Work marked by Claude is cross-checked by GPT, and work marked by GPT is cross-checked by Claude. The second model reviews the same rubric and response, and must explicitly record agreement or adjustment, together with a stated reason and, where it disagrees, the specific criterion being changed.

This differs meaningfully from asking a single model to check its own answer, which typically produces the same conclusion restated with greater apparent confidence. An independent opinion from a different provider, required to justify any disagreement in writing, is considerably harder to produce without genuine review taking place.

Diagrams are checked against the mark scheme automatically

Mock exam questions that include a figure, such as a triangle with labelled sides, a graph or a circuit diagram, are a common point at which AI generation contradicts itself: the figure states one value while the mark scheme was solved using another. After a paper is generated, Lumen runs a separate audit pass whose sole purpose is to check every labelled value in the diagram against both the question prompt and the mark scheme.

Where a mismatch is found, the surrounding text is corrected to match the figure, since the figure is what the student actually sees. Where the discrepancy cannot be resolved, the diagram is removed entirely, rather than allowing a paper to be issued that visually contradicts its own answer key.

Chat responses draw on a student's actual account data

When a student asks Lumen's chat what is scheduled for a given day, or which topics exist for a subject, the response is not generated from general knowledge or a plausible-sounding guess. The system is configured to retrieve the student's actual subjects, timetable, topics, notes and resources from their account, and is instructed to use that data rather than infer an answer. A general-purpose assistant given the same question about a student's biology topics has no access to that information, with nothing preventing it from producing an answer that merely sounds plausible.

Identical submissions receive identical marks

Submitting the same question to a general-purpose model twice can produce two different answers, a consequence of how these models sample their output. An identical resubmission to Lumen instead replays the exact verdict already stored for that submission, so a mark does not vary between attempts for reasons the student cannot see.

The practical difference from using Claude directly

Submitting an essay directly to Claude or ChatGPT produces a single model's opinion, with self-reported confidence that has no independent check, no memory of the student's actual rubric unless it is supplied again with every message, and an answer that may differ if the same question is asked twice. Lumen introduces a second, independent model, a confidence cap derived from objective evidence rather than the model's own tone, grounding in the student's uploaded rubric and account data, and a consistency check on anything the model produces, all applied between the first draft response and what the student ultimately sees.