How This Works — Data, Metrics, Limits

All analysis derives from publicly-sourceable datasets plus live Isle of Man vacancy data. No opinion polls, no hand-weighted scores, no proprietary magic. Every figure should be traceable to a specific row in one of these tables:

• UCM Curriculum Master File (CMF) — 1,891 raw unit rows reduced to 694 genuine parent programmes after suppressing 86 module-code pseudo-courses and 36 APEL (Accreditation of Prior Experiential Learning) placeholders.
• Isle of Man 2021 Census — occupational headcounts at SOC 2020 four-digit level. Not refreshed until the next census; readers should treat distributions as broadly correct but not current-year-live.
• Live IoM job listings — scraped from major employer sites nightly, classified to SOC 2020 via the O*NET crosswalk. The activeVacancies on every page is "live right now", not a YTD total.
• Anthropic Economic Index (AEI) — task-level automation and augmentation shares per SOC. Released Feb 2025. Task-level, not role-level — a task being "automation-exposed" doesn't mean the whole job vanishes.
• Frey-Osborne (2013) computerisation probability — the classical "will this occupation be displaced" score. Older than AEI but still widely cited. Often disagrees with AEI at the task level (e.g. teaching).
• ONS Annual Survey of Hours and Earnings (ASHE), Table 14 — UK national median annual pay by SOC 2020. Low-reliability rows (CV ≥ 20%) filtered out.
• Live IoM salary medians — computed from active advertised salaries when we have ≥3 live postings with salary data for a SOC.
• Azure OpenAI (gpt-4.1) — used to draft narrative paragraphs (Career Outlook per SOC, weekly Insight, Missing Courses gap analysis, Workforce Summary). Prompts force the model to cite specific numbers from the above datasets so outlooks are grounded, not fabricated. Full prompts are in packages/scraper/src/pipelines/*.ts.

• 4% churn is a UK average. Healthcare and hospitality are higher, long-tenure professions are lower. The capacity page uses this across every SOC — be sceptical of any ratio you're about to quote in isolation.
• Graduates don't all stay or enter the matched occupation. Some emigrate, some career-pivot, some choose not to work. The analysis can't see beyond the award date.
• Census is 2021. Four-year-old headcounts. Directionally correct for most SOCs but can't capture post-pandemic shifts.
• The SOC keyword matcher is broad by design. Trust undersupply over oversupply — a keyword-rich umbrella rule will over-credit courses to popular SOCs. Rewritten April 2026 with specific-before-general ordering, but imperfect.
• AI outlook paragraphs are drafted by a language model. Prompts force citation of data numbers but the model can still produce awkward phrasing or stretch an analogy. Every outlook shows the SOC it was generated for — if it doesn't match, the cache is stale (run education-career-ai --force).
• Not every SOC is scored. A handful of SOCs (Advertising/PR at time of writing, SOC 2473) have no Anthropic Economic Index data; they're placed in a separate "Data Unavailable" quadrant rather than being falsely classified.
• Salaries are mixed provenance. Three tiers: live IoM-advertised (best) → UK ONS ASHE (good, clearly labelled) → US BLS via Anthropic EI (weakest, clearly labelled). Chip on each Target Salary tells you which.
• Training lag is an optimistic minimum. Real-world completion rates are under 100%. A "2-year training lag" is the fastest possible response, not the expected response.

In a policy document or report:

"Smart Island Education Analysis, Manx Technology Group (smartisland.im/education), accessed [date]."

For reproducibility, note the data sources cited in Section 1 rather than the Smart Island page alone. If you're contesting a specific number, the pipeline code is open — see packages/scraper/src/pipelines/*.ts on GitHub.