Readiness needed more than an agreement on totals.
AI-generated grading preparation needed to preserve the original marking policy. A converted rubric could produce the expected total while missing a score boundary, omitting relevant input or attributing marks to the wrong skill.
The product question was how to make those failures visible and prevent unverified grading from being automatically accepted.
What I owned.
I added validation against original marking instructions, criterion marks and per-skill evidence, using the same resolved grading configuration intended for student delivery. I also added recovery actions for failed preparation.
My contribution covered validation and recovery. The grading engines were existing platform capabilities built with colleagues.
Check the policy, the boundary and the evidence.
Validate the configuration that will actually be used.
Preparation checks needed the model answer, marking guidance and governed scoring configuration used by the student grading path.
Keep expected judgments independent.
Checks considered original attainable scores, paraphrases and missing or contradictory components. Ambiguous original instructions or incomplete required coverage prevented automatic acceptance.
Give a failed attempt a useful next step.
A repair attempt was checked again against the original cases. When that still failed, staff could review a proposed fix, retain a draft or discard the attempt. A repair could not simply redefine a failing example as correct.
Implemented safeguards, with clear evidence limits.
The delivered workflow exposes validation evidence and blocks automatic acceptance when required checks fail. Tests exercise those guards and the recovery actions.
The checks establish how the workflow behaves under defined cases. A separate provider benchmark would be needed to make an accuracy claim.
Make uncertainty actionable.
An unexplained “needs review” state transfers the problem to the user. A useful quality workflow shows what failed, why it matters and what can safely happen next.
Product visuals and interactive examples are recreated for this portfolio with synthetic content. They are not screenshots of learner records or internal systems.
Next case study
Learning through play