The three streams of MR-81

“Child nutrition connector” wired eleven Octobers and named two blockers, both of them scope decisions rather than obstacles: the 1998-2000 files were “a different, school-centric column set whose unquoted district names push three rows past the last column”, and from 2012 “the Traditional file excludes Community Eligibility sponsors, which are the highest-poverty districts”. Its own alternatives section recorded the choice: “Eleven unbroken years with two named gaps is worth more than seventeen with a silent discontinuity in the middle.”

That was right, and it deferred the question rather than answering it. This phase takes both ends.

Neither blocker survived contact with the files, and one of them was never true.

The early files are fixed-width. Every row is exactly 170 bytes, in all three years, with the thirteen separators at identical offsets. The comma inside Graham School, The shifts nothing, because nothing splits on it. The blocker described a defect of the reader that would have been written, not of the file.

The late files are joinable and should not be joined. All three streams carry the same columns and the same identifiers; concatenating them is a morning’s work. What that produces is a poverty share falling from 46.3% to 33.3% in three years, because community-eligibility schools feed every child and collect no applications at all, so their approval columns are zero by construction. The blocker anticipated the Traditional-only version of this trap. The all-three version has the same shape and is easier to fall into, because it looks complete.

Wire all twenty-one published files: 1998-2000 by offset, 2001-2014 delimited, and 2013’s community stream from the printed report, which is the only form it was posted in.

Carry stream on every row, and carry identified and claimable beside the approvals. Emit Unknown for the three Octobers that state no sponsor type, and resolve it from FY2001 in the consumer rather than writing an inference into the fixture.

Publish no share for FY2012-FY2014. PovertyYear::share and the feed’s meal_program.share become optional and are None there. Those years carry a floor and a ceiling instead, and the chart on /history stops at FY2011 while the table runs to FY2014.

Check every row against the percentage printed beside it, in every year, and fail the rebuild rather than the review.

FY2001 was wrong, and everything downstream of it was. The file is the only comma-delimited one in the series and nine of its rows carry a comma inside a school name. Read positionally, two Cleveland City SD rows put the site identifiers 00026450 and 00093153 in the enrollment column — 119,603 pupils added to a district of 73,562. Statewide public enrollment read 1,921,522 against a true 1,802,937, and the poverty share read 27.7% where the file says 29.5%.

It survived four phases because every check it had to pass, it passed. The row count was right. The sponsor count was right. The share was a plausible number in a rising series. And implausible_sponsors, the filter built for exactly this class of defect, missed it: 192,147 across 118 sites is 1,628 a site, well under a threshold set to catch 342,332 in one.

The repair anchors the row’s nine trailing fields from the right and finds the sponsor type in what remains, from a four-word vocabulary no school name collides with. After it, free and reduced approvals match the department’s printed report exactly for all 1,227 sponsor blocks; twelve sponsors differ on ADM by sixty-eight pupils between them, which is the two publications disagreeing rather than the parser.

Three of the four “published defects” were ours. implausible_sponsors named four sponsor-years and the decision record above described three of them as the department leaking identifiers into the count column for residential and juvenile sponsors. They were the same comma shift. One published defect remains — Portsmouth City’s 342,332 in FY2005 — and it is genuinely in the file, because it reproduces the percentage printed beside it.

A second corpus figure was wrong in the other direction. The DPIA node’s series gave 2005 as 29.9%. That is the share including the Portsmouth defect: the node quoted the number the code refuses. Corrected to 35.7%.

The guard that would have caught both. Every delimited row prints free-and-reduced as a share of enrollment beside the counts, so the columns can be checked against each other. Applied across the series it finds nine violations, all of them in FY2001, and none in the other twenty files. Rows reporting no applications at all are exempt — and the exemption is the finding, not a loophole, because those are exactly the community-eligibility rows whose printed percentage is a claiming rate rather than a measurement.

The series now runs 1998-2011 on applications and stops. 28.9% to 46.3%, seventeen points in fourteen years, with the denominator redefined once at 2010 and drawn as two lines. Three more Octobers exist and carry a band instead: for FY2014, 43.7% to 48.4%. The floor adds the directly certified children to the approvals; the ceiling is what those schools may claim for, the directly certified count times USDA’s 1.6, capped at enrollment school by school. The band brackets FY2011’s 46.3%, so what the source will not settle is not the level but the direction.

And the size of what left the measure is now a number. Enrollment under a sponsor that collects no applications: 7.0% in FY2012, 9.9% in FY2013, 16.6% in FY2014. Community eligibility is open to schools whose poverty is already high, so the population leaving is not a random sample of the one that stays — which is why the naive series falls rather than merely becoming noisier.

What is still not here. The archive stops at October 2014, nine years before the corpus’s FY2024 observations, and nothing in this directory fills the gap. The 1998-2000 files borrow their sponsor types from FY2001 and some thirty-five sponsors a year predate it, so they are excluded rather than typed. Contract 32.0.0.

Append the Traditional file and note the caveat. Rejected, and it is what the blocker was written to prevent. A footnote under a falling line does not stop the line from being read.

Append all three streams’ approvals. Rejected for a stronger reason. It looks like the careful version of the previous option — every sponsor is present, nothing is excluded — and it is worse, because the caveat is no longer visibly needed. The resulting FY2014 share is 33.3%, thirteen points below FY2011, and nothing about the row set says why.

Publish the floor as the series, with the ceiling in a note. Rejected. The floor is a real quantity and it is not this quantity: direct certification reaches families already on SNAP, TANF, foster care or a homeless roll, and an application reaches anyone under the income line who files one. Drawing the floor asserts poverty fell after 2011. Drawing the ceiling asserts it rose. Drawing their midpoint asserts a precision neither supports. The chart draws none of them.

Resolve the 1998-2000 sponsor types from the whole panel rather than from FY2001. Rejected for twenty sponsors. It would reach 1,113 of 1,132 rather than 1,094, at the cost of inheriting an answer from up to fourteen years later for an IRN that may have been reused — and five IRNs in the panel are typed two ways across the window. FY2001 is adjacent, contemporaneous, and numbers districts the same way.

Write the resolved type into the fixture. Rejected. The files state no type; writing one into the committed CSV would make an inference indistinguishable from a reading at the point where every consumer meets it. Unknown is what the file says, and the resolution is code with a test on it.

Repair the FY2001 file rather than the reader. Never on the table, for the reason the previous record gave about the published defects: the digest manifest pins what the department published. What changed is the reader, and the department’s bytes are byte-identical to what was cached before.