Education & Assessment
AI-text detectors misflag non-native English writers at a documented 61.3% false-positive rate. Six governance patterns that catch an exam-scoring agent before it fails someone wrongly.
Source: Liang, W. et al., “GPT detectors are biased against non-native English writers,” arXiv:2304.02819 (also published in Patterns, Cell Press); The Markup, “AI Detection Tools Falsely Accuse International Students of Cheating,” 14 August 2023. EU-wide, GDPR Article 22 already gives a student a right to human review of a fully automated integrity or scoring decision, live today, not a future deadline.
Who this pack is for
Any organisation running an AI agent that scores, grades, or evaluates learning outcomes, essay or exam scoring, admissions or enrolment screening, academic-integrity or plagiarism/AI-generated-content detection, automated proctoring flags, or similar.
What's in the pack
Set it up once. After that, the decision table runs automatically every time your exam-scoring agent makes a call, and the Agent 365 artefact plugs straight in, covering every agent of that type you run, now and in the future. Under the EU AI Act, a high-risk system needs to be able to produce this documentation on request. This pack generates it automatically, so it's already there when you need it.
Classifies each agent's inherent risk tier from what it touches, what it can do and the worst realistic outcome. Risk sets governance depth, never autonomy or access.
Registers agents as governance-grade, with risk tier, operating mode and a real revocation path.
Catches a capability added after deployment that governance never re-checked against the agent's authorised actions.
Defines the promotion path from supervised to autonomous, with the two safeguards that block a premature promotion.
Catches an action logged under a shared service account instead of the agent's own identity.
Flags both a runaway-loop cost spike and a slower drift tied to a model version change.
Does not cover the accuracy or fairness of the underlying scoring or integrity-detection model itself, whether a given detector's false-positive rate is acceptable, or how to remediate a documented bias like the one cited above. That's a different discipline (model fairness auditing, detector validation) this library doesn't claim to provide. The six patterns govern how the agent built around that logic is deployed, authorised, monitored, and audited, not whether the logic itself is fair.
€499 / $599 / £449
Fixed price, checkout shows the currency you're billed in. Other currencies convert for a small fee. Instant download after payment. 30-day money-back guarantee.
Buy this pack