Where the model must hold back
Guardrails
Learners are people, not systems. Six limits follow, and each one sets a rule the lifecycle must respect.
| Limit | Why it matters | Guardrail |
|---|---|---|
| The product is a change in a person | Outcomes depend on the cohort, the teacher and learners' circumstances, not only on the materials | Treat delivery data as signals to investigate, never as automatic verdicts on content or people |
| No changes mid-cohort | Changing assessment mid-delivery threatens fairness and validity | Queue changes for the next release unless they correct a safety, accuracy or compliance defect |
| Experiments involve people | Deliberately giving cohorts different experiences raises ethical and equity concerns | Prefer comparison across successive releases; run controlled experiments only with ethical review |
| The teacher is part of the program | The same program performs differently in different hands | Record delivery context and separate content issues from delivery issues before attributing cause |
| The calendar sets the release cycle | Intakes, terms and enrolments fix when changes can be released | Plan releases against intake dates |
| Compliance is not quality | Passing an audit shows conformance, not that learners thrived | Report strengths as well as gaps, and measure learning outcomes alongside conformance |
Two further cautions
Observation can slide into surveillance. Data should be collected for program improvement, with clear purpose limits, minimisation and transparency to learners.
Measures invite gaming. A program tuned to reduce resubmissions might simply lower its standards. Indicators should be read together and reviewed by people, never optimised in isolation.