Where the numbers
come from.
102,176 school profiles assembled from federal and state education data. This page describes the sources, how conflicts are resolved, and what the estimation flags mean — so you can cite the data without trusting us blindly.
Six source families. Vintages disclosed.
Government education data publishes on a lag, and different collections run on different cycles. We do not hide that — every field group carries its vintage.
School directory, enrollment, demographics, free/reduced lunch, teacher FTE, virtual status.
Discipline (suspensions, expulsions, arrests, referrals), AP/IB/dual enrollment, gifted programs, staffing (counselors, nurses, psychologists), athletics, ELL/IDEA/504 counts.
Math/ELA/science proficiency (overall and by subgroup), assessment participation, Title I status, CSI/TSI designations, chronic absenteeism, federal graduation-rate fallback.
State accountability ratings (A–F, stars, indexes), 4-year cohort graduation rates, state science assessments. Each rating row carries its source URL.
District per-pupil expenditure, attributed to each school in the district.
Geocodes, locale classifications, county assignment, attendance boundaries; private-school universe.
Six rules, applied everywhere.
Never overwrite, always attribute
Ingestion is COALESCE-only: a state-reported exact value is never replaced by a federal estimate. When both exist, the exact value wins and the field carries the year it was reported.
Estimates are flagged, not hidden
Federal files suppress small counts into bands (“90–94%”, “≥95%”). Where we resolve a band to its midpoint, the field is marked is_estimated — in the deliverable files, the bulk manifest, and the field catalog.
Mixed vintages are explicit
A school profile combines sources with different reporting years. Every field group carries its vintage; per-row marker fields (graduation_source_year, state_rating_year, assessment_school_year) resolve it per school.
Zeros are data
A school reporting zero expulsions is different from a school that did not report. We distinguish reported zeros from suppressed or not-collected values, which stay null.
No composite scores
We do not blend metrics into a proprietary “school score.” Buyers get government-reported facts and derive their own views. State ratings are shown in each state’s own scheme and are not comparable across states.
Versioned releases
Data ships as releases (current: v2026.07) with a machine-readable manifest of every field’s source, vintage, and coverage. Pin a release; diff the next.
Where the data is thin, we say so.
Not every state reports everything
A handful of state DOEs publish no usable school-level ratings or block automated access. Those fields stay null for those states and the coverage figure is disclosed per field in the manifest — we never impute a state’s rating from federal proxies.
Non-reporters are excluded from summaries
Any headline figure we publish (like the field notes on the home page) excludes schools with no reported data for that metric. A school that is silent is not a school with a perfect record.
Boundaries are partial by nature
72,872 attendance boundaries cover the districts that publish them. Schools without a published boundary are marked as such — a missing polygon is not an open-enrollment claim.
Feeder relationships are derived
The 1.32M feeder relationships are derived from boundary overlap and grade progression, with a documented confidence tier on every row. Derived fields are labelled as derived — never presented as government-reported.
Schools — not neighborhoods, not people.
School data are informational, drawn from government sources as of the vintages listed above. They describe schools — not neighborhoods, and not the people who live in them. We do not publish composite quality scores, and licensees integrating this data into housing surfaces are contractually required to comply with the Fair Housing Act.
If you cite K12 Atlas publicly, cite the release: K12 Atlas (2026), U.S. School Dataset, release v2026.07. If we publish a factual error, it goes in a changelog and affected buyers of the prior 30 days are refunded proactively — we do not silently correct.
Read a report built this way.
The free Austin ISD sample carries this methodology in its appendix.
Download the free sample →