Skip to content
LEAPDocs

Grounding and guardrails

How the platform stops the AI from inventing your curriculum.

The principle

Generated content must trace back to something you supplied. Every artefact carries the source spans it was drawn from, and generation fails loudly rather than inventing content when it cannot ground an assertion.

Why this matters more here than elsewhere

A plausible-sounding hallucination in a chat assistant wastes a minute. The same hallucination in a syllabus becomes a lesson plan, a set of questions, a mark, and eventually a transcript entry.

The case that made this non-negotiable

Before paper-quiz grading was grounded in the extracted answer sheet, a model could produce responses that were never on the page. A blank answer sheet scored twelve out of twelve. The extraction was hallucinating, not the grader — and the mark looked entirely reasonable. Grounding was added so every mark names the criterion and the extracted text it was matched against.

Where grounding applies

ExtractionGates and outcomes must come from the uploaded document. A thin syllabus produces thin gates rather than invented ones.
AuthoringLesson plans and scripts carry the source spans behind them.
Paper gradingGrounded in the extracted answer sheet and the uploaded marking scheme, with the matched criterion shown per response.
Rubric gradingRecords the detected answer verbatim, so you can check the quote against the page.
AI tutorRestricted to the course’s own material, and to gates the student has reached.

Other guardrails

  • Approval before exposure. Nothing generated reaches a student unapproved and unpublished.
  • Teacher-only generation, enforced server-side rather than hidden in the interface.
  • Strict output validation. A malformed response fails the job instead of being partially written.
  • Override always wins. Any AI-assigned mark is overridable, and the override recomputes everything downstream.
  • Tutor sequencing. The tutor will not explain gates the student has not reached.

What grounding does not promise

Traceable is not the same as correct. Grounding guarantees an artefact came from your document; it does not guarantee the pitch suits your cohort, the terminology matches your textbook, or the emphasis is right. That is what the review step is for.