SchoolsData — Methodology, Sources & Decision Rules

Axis	Inputs
academic	NCEA L2/L3 pass rate + UE rate (Phase 4.7)
trajectory	ERO diff direction × magnitude + roll growth %
wellbeing	ERO wellbeing rating + attendance % + stand-down rate
equity	Māori/Pasifika achievement parity + EQI-adjusted outcomes
governance	Financial health (surplus consistency) + audit opinion + board stability

Parent weight-shifting on the site: each axis renders as a slider, composite recomputes live. Methodology shown on every page.

---

4. School↔suburb matching

Schools join to Crimestats suburb_id via a 4-tier strategy (see match-schools-to-suburbs.cjs):

1. A confidence — exact name + region: norm(school.suburb) + norm(school.region) == norm(suburb.name) + norm(suburb.region). 34 of 2,576 schools.

2. B confidence — name match disambiguated by geo: when school suburb name appears in multiple Crimestats suburbs, pick the geographically nearest (within 15km). 266 schools.

3. B confidence — geo nearest <3km: lat/lng nearest-neighbour to a suburb centroid. 1,737 schools.

4. C confidence — geo nearest 3-10km: same as above but wider radius. 379 schools.

5. Unmatched: rural schools without lat/lng or with no nearby suburb. 160 schools.

Confidence rendered on the public site — A and B are surfaced freely; C surfaced with a "best-fit" disclaimer; unmatched shows no suburb panel.

Known limitation: Crimestats uses StatsNZ SA2 boundaries (small statistical areas), so commercial-fringe SA2s (e.g. "Mount Eden North East") can have inflated harmper1k metrics from non-resident offences. Refinement planned: prefer name-overlap match within 1km even if not absolute nearest.

---

5. School↔ERO institution matching

ERO uses Ministry of Education schoolid as their canonical institutionid. URL pattern: ero.govt.nz/institution/{school_id}/{slug}.

Wayback Machine bypass strategy (see proof-slice.cjs):

1. Try archive.org/wayback/available?url={canonical} with multiple slug variants

2. Fall back to CDX search if wayback/available returns empty (their two APIs sometimes disagree)

3. Pick the newest 200-status snapshot across all variants

Footguns documented in this pipeline:

• encodeURIComponent breaks wayback/available — pass URL literally
• CDX matchType=prefix strips trailing slashes so "54/" text-matches "5499/" — use exact-URL lookup instead

---

6. Defamation + privacy rules

School-level tags — LOW risk

Tags about institution behavior (e.g. turnaroundgovernance, surplusconsecutive3y, improvingmaori_achievement). All deterministic from data; every tag links to its evidence row.

Person-level tags — HIGH risk, restricted

ALLOWED (factual + public-role only):

• prolificerodirector (≥100 reports written) — verifiable count, public role
• regional_specialist (≥80% of reports for one ERO regional office)
• longtenureprincipal (named principal ≥10 years at same school in school_leadership)

NOT ALLOWED (evaluative judgments on named individuals):

• strictreviewer / lenientreviewer — even if statistically supported
• underperformingprincipal / consistentlynegative_reviewer
• Any tag combining named individual + negative outcome

GREY ZONE (allowed only after legal review):

• "Director X averages 73/100 composite score (national mean 71)" — render as factual stat line, not pill
• "Principal X led the 2017-2019 turnaround period at School Y" — positive attribution to public role, date range tied to documented improvement
• Stored with safetyclass='greyzone' + legalreviewedat IS NULL — not surfaced publicly until reviewed

Defamation guard at extraction time

Every verbatim quote inserted into schooleroquotes must be an exact substring of the source markdown. Quotes that fail substring match are silently dropped by the extractor (Haiku occasionally smooths punctuation/connectives — those are caught and filtered, with the drop count logged on the parent report row).

Privacy Act

• Public roles + verifiable facts: OK
• Cross-aggregating named individuals with personal data from other Pangaea sites: requires Privacy Impact Assessment first
• Principal/board chair names from ERO reports: public roles, OK to surface
• Photos of named individuals: only school-published photos with explicit clearance

---

7. Year-stamping + freshness

• Every aggregation page (/schools/{state}, /suburb/{slug}, /tag/{slug}) carries the current year in title + H1 ("Best Schools in Auckland | 2026 Rankings")
• Tier-A source-doc freshness cap: 36 months. If a school's most recent ERO review is older than 36 months, the trajectory panel renders a "data may be stale" banner
• ERO reports from Wayback include sourcesnapshotts — visible on every quote citation

---

8. AEO + LLM citation surface