The directory cannot say why

nces-ccd was wired for a single year, 2022-23, because that year carried the one column two other connectors were waiting on: ST_LEAID, which is the Ohio IRN behind an OH- prefix. The connector’s blocker said the rest needed “a consolidation-aware long series”, and named two obstacles — per-year zips whose column sets change, and an identifier-change history that “has to be derived rather than read”.

The consumer that made this worth taking up is crates/dispersion/src/ohio_panel.rs. It builds Ohio’s F-33 finance panel across thirteen fiscal years and resolved every agency’s IRN through that one 2022-23 directory, so an agency that closed earlier had no IRN at all: 141 in FY2009, 124 in FY2012, 0 in FY2022. Its own docstring drew the conclusion:

the count going from 124 in FY2012 to 0 in FY2022 is the consolidation history, measured.

That sentence is what this phase set out to check, and it is wrong.

Wire the sixteen school years 2008-09 through 2023-24 — the whole delimited era, and exactly the span the F-33 panel covers. Commit them as one row per Ohio agency per year: crates/dispersion/fixtures/ccd-lea-directory.csv, 17,618 agency-years.

Resolve the panel’s identifiers against the directory for the panel’s own year, falling forward to the nearest later year only where the contemporaneous one is not held.

Publish no reason for any departure. Carry the operational status code verbatim and read nothing out of it. Hold the negative — that Ohio has never filed the consolidation code — as a function with a test on it rather than as a remark, so the day the source starts answering, the build fails instead of the corpus going on saying the question is unanswerable.

Leave the fourteen fixed-width years, 1994-95 through 2007-08, retrievable and unwired, with what they would cost written down.

Every agency in every year is now named. Unnamed agency-years go from 1,046 to zero. And no published figure moved: the feed is byte-identical, because every one of those 1,046 rows was a non-comparable agency the aggregates already excluded. That is the first evidence that the count was not measuring what the docstring said — a hundred and twenty-four agencies whose presence or absence cannot move a district statistic are not a district consolidation history.

The composition settles it. 341 agencies leave the directory across the window: 327 community schools, nine service agencies, five regular districts. Two of the five are STEM schools, which Ohio funds as their own unit under R.C. 3326 and the federal directory types as regular districts. So the genuine traditional-district consolidations in sixteen years are three — Bettsville Local, Ledgemont Local, Newbury Local — and the panel’s own comparable population barely moves across the same span, 612 agencies in FY2009 against 611 in FY2022. The docstring’s number was out by roughly a hundred to one.

The reason for a departure is not derivable, and the source is worse than silent. The CCD defines eight operational-status codes and exactly one marks a consolidation: code 5, “significant change in geographic boundaries or instructional responsibility”. Ohio has filed it zero times in 17,618 agency-years here, and zero times in the twenty-six years back to 1998-99 that were checked and not wired. What Ohio files instead, for all 341 departures with no exception, is code 2: “closed with no effect on another agency’s boundaries.”

For the three real consolidations that is false. Bettsville’s territory went to Old Fort, Ledgemont’s to Berkshire, Newbury’s to West Geauga and Chardon — and all four receiving districts are coded 1 Open in the year before, the year of, and the year after. The departing row asserts the negation of what happened; the surviving row records nothing at all.

The two obvious derivations were tested rather than assumed away. Enrollment absorption on the survivor is confounded roughly fifty to one by ordinary growth — 141 comparable-district year-pairs gain more pupils than Bettsville ever had — and it gets the sign wrong on Newbury, where Chardon lost 110 pupils the same year it received some of them. Name changes on the survivor arrive up to two years out of alignment with the closure they supposedly record, and share a vocabulary with cosmetic re-spellings: ESC of Cuyahoga County becoming ESC of Northeast Ohio looks exactly like Mercer County ESC becoming Mercer County Educational Service Center.

Two columns that look like they would help do not. UPDATED_STATUS is byte-identical to SY_STATUS on every Ohio row from 2014 to 2023. EFFECTIVE_DATE holds one value per release — it is when the file was cut, not when anything took effect.

Three seams a single reader has to cross, each capable of matching nothing silently. The columns carry a year suffix through 2009-10 (STID09, BOUND09) and lose it after. The delimiter is tab through 2014-15 and comma after. And ST_LEAID is a bare six-digit IRN through 2015-16 and OH-043786 from 2016-17. Ohio is selected on FIPST and never on LSTATE, which disagree in both directions: Urban Pathways of Youngstown is FIPST=39, LSTATE=PA, and earlier years file an Arizona agency as LSTATE=OH with an Ohio-shaped state id beside it.

What is still not here. The fourteen years 1994-95 through 2007-08, all retrievable, all fixed-width with no header and the column positions moving in seven of the nine years between 1998-99 and 2006-07. And the reason any district went, which needs Ohio’s own territory-transfer orders under R.C. 3311 — the same answer State Foundation Aid already reached about the department’s disappearing appropriation lines, and for the same reason.

SUPERSEDED by “Before there were service centers”. The fourteen years are wired. Two things said here about them are worth correcting rather than quietly overwriting.

The filenames. The convention recorded here and in the catalog was one year late — pau94datr is 1994-95, not pau95data, and ag981c is 1998-99, not ag991a. The record lengths recorded alongside it belong to the correct alignment, so the two notes contradicted each other and the lengths were right.

The difficulty. “The column positions move in seven of the nine years” is true of the record and false of the five columns this reader takes. Four of them never move at all in thirteen years; the status byte moves four times. What was actually hard was that one of the layout documents states that byte’s position wrongly.

And the negative. “Zero times in the twenty-six years back to 1998-99 that were checked and not wired” is now zero times in thirty years that are checked and wired, under test, over a completely enumerated population of 30,655 agency-years.

CORRECTED by “The order was never the states”. Two things above are wrong.

The custodian. “Ohio’s own territory-transfer orders under R.C. 3311” is right about the statute and wrong about the office. R.C. 3311.22 vests the transfer in the educational service center governing board; the State Board of Education is an appeal body, reachable only if the receiving district objects. West Geauga’s board accepted unanimously, so no State Board instrument was ever made. It is not withheld — it does not exist.

And Chardon. This record and the module both said Newbury’s territory “split between West Geauga and Chardon”. The resolution transfers all of Newbury to West Geauga, and this repository’s own F-33 panel agrees: West Geauga gains 208 pupils in FY2021 while Chardon loses 110. The claim came from a judge’s failure-mode analysis and was carried into a docstring and a test without being checked against the fixture sitting beside it.

Classify from the status code. Rejected on the evidence rather than on principle. If Ohio populated code 5 this would be a direct read and worth having; it does not, and the code it does populate states the opposite of the truth for every case anyone would want to look up.

Identify the successor by enrollment absorption. Rejected at roughly forty-seven false positives per true positive, and on getting Newbury backwards. A district of 400 vanishing the year a neighbor grows by 400 is suggestive and is not proof, and an unproven join in a committed fixture is worse than an absent one because nothing downstream can tell them apart.

Publish the three known consolidations as facts, sourced from county records. Rejected as out of scope rather than as wrong — they are true, and this repository’s rule is that a claim reaches a corpus node through a committed primary source in one hop. The three are named in the module and in the catalog record precisely so that someone can go and fetch those orders.

This rejection has expired. The ground it rested on — that no committed primary source is one hop away — turns out to be false: the Auditor of State’s reports on the receiving districts recite each resolution by date and issuing body, and serve as PDFs to this project’s user agent. See “The order was never the states”, which scopes that as its own phase rather than doing it in passing.

Wire all thirty years while the discovery was in hand. Rejected. The fixed-width era needs a per-year position table for seven distinct layouts, and the F-33 panel it would serve does not reach before FY2009 — so it would be a second reader built for no consumer. The URLs and the layout hazards are written down in the catalog record, which is the cheap half of that work.

This rejection was wrong on its second clause, and the first was overstated. The per-year table is thirteen lines, not seven layouts. And there was a consumer, named in this very record without being recognized: the reason a district departs sits with an educational service center’s governing board, and in 1994-95 Ohio had no service centers — it had 86 county boards, thirty-nine of which disappear over the next seven years with the same code 2 this record is about. The finance panel is not the only thing a membership series can serve, and “no consumer” was a statement about which fixture was in view. See “Before there were service centers”.

Leave unnamed_agencies reporting the gap. Rejected, and it is kept with its assertion inverted instead. The function now says zero and a test holds it there, because the property worth guarding turned out to be the opposite of the one it was written for: not “how much can the directory not name” but “the directory names everything, and a hole here is a retrieval failure rather than a fact about Ohio”.