Student Assessment

A teacher with 150 essays and a weekend will look for help. The moment a model contributes to a mark, the school owns an explanation it may not be able to give. That gap is what this starter addresses. It screens the prompt, masks a national identifier pattern, and refuses a weak answer on the way back.

Assessment is a consequential decision. A grade, a placement recommendation or a comment on a transcript follows a learner for years. A district technology lead needs a defensible account of how each answer was produced. This starter builds part of that account into the request path.

Read the redaction stage carefully. It masks one pattern group only, which is a national identifier such as a national insurance number. Local student numbers, learner names and grade records fall outside that pattern. If your prompts carry those fields, add patterns before the starter matches your real data.

No access control step exists in this chain. Injection screening and redaction are the only request side gates, and both run before the model is called. Everything after that point inspects returned text. A district that needs role separation must add an access rule, because none ships here.

Several checks apply in turn to the answer. A bias signal is present as a configuration point, and this starter sets no custom threshold for it. Source checking flags weak grounding and still delivers the text. Output scoring holds a combined floor of 0.86, with component floors of 0.78 for relevance and 0.80 for accuracy. A response under 120 characters or shorter than two sentences fails that same gate. The review point contains no rule, so no result routes to an educator until you turn it on.

Two closing notes. The final audit step marks an allow decision, never blocks, and creates no stored record on its own. Choose the University starter when the same campus also needs an access rule, student record patterns and term screening across many workflows.