Home education connector

#257 recorded that Ohio holds home-education counts per district, releases them on request, and publishes only the statewide total. It named the one file outside the department that has them — the Johns Hopkins Homeschool Hub workbook, cataloged as Johns Hopkins Homeschool Hub — state and district home-education counts — and concluded it was “not fixture material as it stands”, on four recorded defects.

Three of the four turned out to be answerable from the file and the fixtures this repository already holds.

“No IRN, and 48 rows are unresolvable without one.” The rows are in IRN order, which nothing had checked. 560 of the 611 names are unique in the department’s own district list and 558 of those strictly increase — one monotone run over rows 1 to 606, with four rows appended after it. A repeated name between two unique ones is bounded by them, and the bound admits exactly one candidate every time but four; those four fall out by elimination. 611 rows, 611 distinct IRNs, nothing unresolved.

“A Google Sheets export has no stable content address.” A supposition, and wrong. Four downloads over two days returned the same 645,110 bytes under the same SHA-256. The export is deterministic.

“The year label contradicts the workbook’s own state total.” True, and it is the finding rather than the defect. The catalog entry inferred 2020-21 from it by assuming the 47 censored cells were small. They are not: Huber Heights is 5,631 pupils, Euclid 4,166, Southwest Local 4,150. Against contemporaneous F-33 enrollment, 2020-21 needs a censored participation rate of 0.22% against a reported 3.55% — a sixteenfold discontinuity — while 2021-22 and 2022-23 are arithmetically impossible, and a size-matched estimator on 2023-24 reconstructs 53,100 against the published 53,051, missing by 49 students where every other year misses by 3% to 13%.

Two further things the entry does not record. There is a third censoring marker: 27 <10, 20 . and 5 NC, so 559 numeric rows rather than the 564 the entry gets by subtracting 47 from 611. And 20 districts report an explicit 0.0, among them Cincinnati Public Schools at 34,860 pupils.

What remains of the fourth defect is real and unfixable from here: the workbook’s own citation points at a department page that publishes no counts, and what the count measures — notices received, or children named on them — is undocumented.

Declare jhu-homeschool-hub as a connector, Wired with a recorded remainder, feeding metric and education-agency, with the workbook as its single digest-pinned source, and commit two fixtures from it: the statewide annual series, and the per-district panel keyed by IRN.

Register an aggregator, and say in the module why that is not a rule violation. The prelude’s rule is that figures an aggregator carries support [inference] and never [verified]. That governs what may be claimed, not what may be committed. Pinning the file by digest turns the claim into “this is what the Hub published, and here are its bytes”, which is verifiable; the counts themselves stay [inference] exactly as the catalog entry has them.

Key it by joining to the typology source, not to the typology fixture. The IRNs and counties come from district-typology-2013, read from the cached workbook at rebuild time, so a rebuild joins two published files rather than one published file and one of this repository’s own outputs.

Carry the publisher’s year label verbatim and no other. published_year is SY21-22 on every row, because that is what the file says. The 2023-24 argument above is an inference, it lives in the catalog entry, and writing it into a column would launder it into data — the exact failure the [verified]/[inference] split exists to prevent. The builder asserts the contradiction instead: the numeric rows sum to at least 51,349 against the same workbook’s 2021-22 total of 47,491, and if a revision ever makes those two sheets agree the build fails, because the license for committing undated rows will have gone.

Bound the censored cells, do not smooth them. <10 bounds a cell at [0, 9], the same rule nonpublic reads off the same masking convention. . and NC get a floor of zero and no ceiling, because nothing in the file bounds them above and one of them is a 4,150-pupil district. The published cell text is carried beside the pair, so the three markers stay distinguishable and the twenty explicit zeros stay distinct from all of them.

The only per-district home-education breakdown known to exist is now in the repository, and the denominator question #257 was opened for can be asked: which districts have lost what share of their resident school-age population by a route no enrollment file records. The statewide total could not answer it, and the issue’s own framing is that a population across some 600 districts is thin almost everywhere and concentrated somewhere.

Two independently pinned files now corroborate each other. The Hub’s 2023-24 statewide cell is 53,051 and the department’s own fact sheet — education-landscape-2024, pinned separately under “Facts and figures connector” — prints 53,051. That agreement is the evidence the series is the department’s counts carried forward rather than an estimate, and neither file alone carries it.

That record’s third alternative is falsified and stays on the page. It rejected committing this workbook because “a Google Sheets export has no stable content address” and because of the four defects. Both grounds are gone — the first measured wrong, the second three-quarters answered. The record is not rewritten; a decision record is what was decided and why, and the correction belongs here, in the record that supersedes it.

The year is unresolved in the data on purpose, which means every reader of the district fixture has to consult the catalog entry to know what year it is. That is a real cost, and it is the cheaper error: a wrong year written into 611 committed rows propagates silently, and an absent one does not.

Nine other states use . and Ohio alone uses <10 and NC, which is the evidence for reading the first as the compiler’s blank and the other two as the department’s markers. It is evidence, not documentation — the codebook defines none of the three — so the fixture carries all three verbatim rather than mapping them onto one meaning.

Nobody has looked at the other 38 states. The district sheet holds 12,280 rows and this reads 611 of them.

Leave it cataloged, as the entry and facts-and-figures-connector both concluded. Rejected on measurement. Both grounds for that conclusion were checkable and neither had been checked: the export’s stability took four downloads, and the IRN ordering took one pass over the committed typology fixture.

Fold it into dew-facts-and-figures. Rejected. Same subject, different publisher and a different kind of authority: one is the department publishing its own count, the other is a university republishing a count the department gave it privately. A connector carrying both carries one status for a public file and a records request, which is the argument ohio-bill-versions makes against merging with ohio-session-laws.

Write the inferred 2023-24 into a school_year column. Rejected, above. The inference is strong — 49 students on a size-matched reconstruction — and it is still an inference, and a column is where inferences stop being visible as inferences.

Impute the censored cells from the size-matched estimator and commit the point estimates. Rejected. The estimator exists to date the sheet, which is a question about the file; using it to fill 52 cells would put 1,700-odd modeled students into a fixture indistinguishable from the 559 published ones. The floor/ceiling pair says what is known and what is not.

Commit only the district panel and leave the statewide series out. Rejected. The series is what the year check reads, it is twenty years against the fact sheet’s one, and its 2023-24 row is the cell that ties the whole workbook to a [verified] departmental figure.