Methodology

The data layers

1. Source files are fetched from the publishers, recorded with a SHA-256 hash and retrieval time, and inspected. Registry: sources.yaml; first-fetch findings: findings.md. The state’s public host refuses automated requests from cloud networks; the same files are served from its origin host, which the fetcher records.
2. PDF tables(the state’s facility inventory report and the board’s attendance-zone book) are parsed twice by independent code paths and loaded only when both agree on every number and the publisher’s own arithmetic reproduces (fish_los.py, abc_book.py; parsed tables under reference/parsed).
3. Loadersnormalize every file to long-format facts (school × year × metric). School numbers are zero-padded to four digits at the boundary because the state pads them in some files and not others. A grade cell marked with an asterisk is kept as suppressed unless the numeric grade cells already sum to the school’s published total, in which case the asterisk marks a grade the school does not serve.
4. Derived metrics (utilization, five-year change, kindergarten-to-grade-5 ratio, students per teacher) are recomputed from the served inputs by the validation harness and must match exactly.

What “validated” means

Each served metric runs through the four-check taxonomy used for New York and Philadelphia:

• (i) Completeness — every school row in the source file is served and every served cell is in the file.
• (ii) Base case — the served value equals an independent re-read of the exact ingested file. Exact tolerance: these are counts and published percentages.
• (iii-b) Cross-publisher — the state’s October membership against the federal copy in the NCES Common Core of Data.
• (iv) Computed values — derived metrics recomputed from validated inputs.

Latest run: PASS — base case 13 components, 19,596 matching cells, 0 disagreements; cross-publisher 98.9% exact on 1,533 school-years; computed values 0 mismatches. Check report · per-metric definitions and known gaps

Analyses

Enrollment, capacity and the 2026 closure lists— ten years of October enrollment, building capacity from the state inventory, per-pupil spending, and the board’s closure and review lists. Definitions and limitations: the metric methodology page.

Residual-z: outcomes adjusted for who enrols— a weighted regression of each outcome on same-year composition, size, level, sector and magnet status; the standardised residual is the second method on the outliers page. The page reports the pre-registered checks against the state’s learning gains (one of which fails), a covariate ablation and the disagreement with the peer-percentile method.

The school climate survey— seven years of the district’s student, parent and staff surveys: coverage, how the composites are built from items asked every year, and a pre-registered test of whether parent and student sentiment predicts later enrollment loss beyond utilization and school grade.

Metric catalogue

Every served metric has a page giving its definition, the source file and loader, what is served, what it cannot show, year-to-year comparability and its validation status. Ratings use the performance and climate metrics; the enrollment, capacity and spending metrics feed the closure analysis.

What is not here

The school-level special education share, mobility and prior achievement the state does not publish are absent from the peer-group features and the residual-z covariates, so both adjust for less than a value-added model would. The plan and its sequencing are in the integration plan; the decisions behind the universe, identifiers and definitions are in the overview.