Nonpublic school location

R.C. 3317.024(E)(1) pays auxiliary services through the district a chartered nonpublic school is located in. program/auxiliary-services carried “the district each chartered nonpublic school is located in” as an unfilled field, and “Nonpublic enrollment connector” recorded why: no October file in any era has the column, and a join on name fails because October 2023 holds 711 schools under 601 names, sixteen of them St Mary.

#759 named three candidate sources in order: the department’s directory (OEDS), geocoding into the Census block geography the legislative crosswalk already reads, and NCES’s Private School Survey as a cross-check.

OEDS has a public extract. It lists every open nonpublic school by building IRN with a mailing address and a parent organization — the diocese or association, never the public district. The directory has no field for that. So the first source supplies the key and the address, and the second has to supply the district.

Declare dew-directory, Wired, feeding school, with one source: the OEDS extract for organization type 5, fetched by POST. Add geocoder-nonpublic-schools to census-geography: the Census Bureau’s batch geocoder’s answer for every address in that extract, against the current address ranges and reporting 2020 blocks.

Build crates/project/fixtures/nonpublic-school-district.csv from three joins, each on a key: building IRN to address, address to 2020 block, block to district through the _SDUNI member of baf-2020-oh, and district code to IRN through the CCD directory. No boundary file is added: the blocks are the ones the legislative crosswalk already apportions.

Two request shapes in the fetch layer. Neither source is a file at a URL. registry::REQUESTS holds them beside the registry — a form POST and a multipart geocoder upload built from another source’s cached bytes — rather than as a field on Source, which would mean writing Get into four hundred literals. A test insists every key there is a registered source and that the geocoder’s input is fetched first.

Every school is a row, placed or not. A school the geocoder could not place carries its outcome (No_Match or Tie) and a blank district, so the unmatched set is in the fixture and not only in a count.

The district is written as the block file found it. A district that has ceased since 2020 keeps its own IRN with the last CCD year that listed it. Rewriting it to the receiving district in the builder would hide the one row where the census and the present disagree.

The reach, against October 2025. 749 buildings reported; 743 are in the directory and 714 are placed — 95.3%. Of the 35 without a district, six are absent from the directory and 29 are in it unplaced: 28 the geocoder could not match and one it matched to two places equally. crates/project/tests/the_district_a_nonpublic_school_sits_in.rs pins the rate and both sets by IRN, so a refresh that changes either fails with the school named. Directory-wide it is 787 schools, 755 placed, in 278 districts.

A snapshot, and the rate decays backward. The directory lists schools open on the day it was generated. October 2024 has 710 of 722 buildings in it and October 2023 688 of 711. Nothing here is projected onto an earlier October: a school that has moved is placed where it is now.

One school sits in a district that has since ceased. St Helen, in Newbury, geocodes to Newbury Local, which the 2020 blocks still record and whose territory went to West Geauga on July 1, 2020. The test resolves it against the territory-transfer record. The same gap affects the legislative crosswalk, which joins blocks to the 2022-23 CCD directory alone and so drops Newbury’s blocks from West Geauga’s apportionment without a word. That is outside this record and is filed separately as #769.

A boundary is a hypothesis about address ranges. The whole extract was geocoded a second time against Public_AR_Census2020. It placed 750 rather than 755; of the 742 both placed, 30 landed in different blocks and two in different districts — 067629 (Mansfield, 04943 against 04945) and 022583 (Youngstown, 05019 against 04516). So a school near a boundary is placed by an interpolation along a street segment, not a surveyed point, and about one school in four hundred moves district with the reference data.

Both digests move on every refresh. The extract’s first line is a generation timestamp, and the geocoder answers in no fixed order against address ranges the Bureau revises. A changed digest is expected; a changed fixture is the thing to read.

Leave the field unfilled, as nonpublic-enrollment-connector decided. That record rejected matching “by name or by geography” because the result “would be wrong for an unknown fraction of buildings and look exactly like a right one”. The objection is to the unknown fraction, and it is answered rather than overruled. No step here matches a name: the school is keyed on its IRN, the district on its code. The fraction is known and pinned — 35 of 749, by IRN — and every row carries the geocoder’s own verdict on itself. The residual error that cannot be pinned, a school near a boundary, is measured by the second benchmark at two schools in 742.

Take the 2020 benchmark’s answer where the current one misses. It matched eight of the 32 current misses. Rejected: one benchmark is one set of address ranges, and a fixture whose block depends on which pass found it cannot say what any single row was placed against.

Cross-check against NCES’s Private School Survey. Named by #759 as the third source. Not attempted: the survey carries an address and a county rather than a district, so it would test the directory’s addresses, not the placement. It stays open as a check, not as a blocker.

A district boundary file. Rejected: the block assignment file is already a district boundary, at block resolution, and is the one the legislative crosswalk uses. A second boundary source would let the two disagree at the edges.