The open item audit

Four consecutive phases had the same shape: a recorded claim turned out to be wrong, and the correction was one curl away. Four stale blockers and five wrong claims, all found by opening a file the corpus had already cataloged.

That pattern says something uncomfortable about the [open] markers themselves. They were being read as a map of what is unknown. They are a record of what somebody did not get to, written at a moment, and never revisited. Some describe problems that are already solved; some describe the wrong problem.

So: enumerate every one and classify it, rather than continuing to pick items off by intuition.

Sweep all [open] claims mechanically — by node, by field — and classify each as answerable from data on disk, answerable from a source now wired, genuinely blocked, or misdescribed.

Prioritize the statutory_basis cluster, because ohio-laws had just been wired and every one of those items was of the form “transcribed from the calculator; the correspondence with statute has not been read.”

152 claims, across 12 node classes. The distribution is itself informative:

description      63     prose claims, mostly genuine open questions
series           10     parameter histories; nearly all need `lsc-budget`
series_path       8     unpopulated agency data paths
established       7     agency metadata
statutory_basis   9     -> 3 after this pass
typology          6
vetoes            6

The statutory_basis cluster is now six, and the six are not what they were. Four closed outright, and the interesting part is that reading the statute did not simply confirm or deny the transcriptions. It sorted them into four kinds of parameter with different durability, a distinction invisible in the calculator because a spreadsheet renders all four as a number in a cell:

• Legislated — written in the section to four decimals. The fourteen multiples, DPIA’s $422, preschool’s $4,000 and its X 0.50, targeted assistance’s 0.008 and its 200/400 brackets, gifted’s $24, $2.50 and the three unit prices. Changed only by an act.

• Delegated — statute names the department. R.C. 3317.03(B)(21) hands it the definition of “economically disadvantaged”, constrained only by “shall not be categorically excluded … based on anything other than family income”. Changed by a departmental decision.

  This slot originally held DPIA’s 65/35 blend, and that was wrong. The audit concluded not in law from not in the Revised Code — the same move it was auditing — and the blend turns out to be set on the face of H.B. 96, in a table. It is uncodified statutory, not delegated. Corrected in the phase that wired lsc-budget, and left visible here because a taxonomy that quietly reassigns its own examples is worth less than one that shows where it was wrong. The distinction the correction produced is sharper than the original: the act sets weights on counts whose definitions it does not control.

• Measured — statute specifies a computation over reported data. Transportation’s two rates are trimmed means of what districts themselves spent last year, the ten highest and ten lowest excluded; R.C. 3317.0212 contains no dollar figure at all. The 40th-highest-district benchmark in local capacity is an order statistic. Changed by districts’ own behavior, with nobody deciding and no bill required.

• Uncodified — nowhere in the Revised Code. The base funding supplement, the enrollment growth supplement, the performance supplement and the formula transition supplement. Changed by the biennial budget, and gone if it is not re-enacted.

$214.8m is paid under provisions that are not in the Revised Code — those four supplements together, more than English learner and career-technical funding combined. Nothing improper follows; Ohio funds a great deal this way. What follows is that a reader who takes them for permanent program funding is wrong in a way no figure on the department’s report would reveal, and this corpus was listing them beside base cost as though they were the same kind of promise.

A second wrong citation. fsfp-gifted-units cited R.C. 3317.053, which does not exist — gifted units are R.C. 3317.051. That is two of nine statutory_basis fields pointing at sections that are not there, after the DPIA node’s R.C. 3317.029. Both were plausible-looking numbers in the right chapter, which is exactly why nothing caught them: a citation is only checkable against the thing it cites.

And one of my own notes was wrong the same way. The registry described R.C. 3317.019 as “gifted units and their prices”. It is Temporary transitional aid. The note was written from the corpus’s own citation rather than from the section, which is the identical failure being audited.

A caution the audit earned the hard way. “Not in the Revised Code” and “not in law” are different claims, and this pass conflated them for the DPIA blend, for the four supplements, and in the framing of the whole uncodified category. The supplements’ classification survives — they really are temporary law that lapses — but the reasoning that produced it was “absent from the sections we fetched”, which would have been equally satisfied by a provision sitting in a section nobody thought to look at. The check that settles it is reading the enacted act, and until lsc-budget was wired the audit could not do that.

What did not move. The 63 description claims are mostly genuine — questions about intent, about incidence, about why a threshold sits where it does. The 10 series items need lsc-budget and stay blocked. The series_path, established, typology and vetoes clusters are agency and legislation metadata that was never populated; they are honest emptiness rather than open questions, and a future pass should probably distinguish “unknown” from “unentered” in the notation rather than marking both [open].

Keep picking items off by intuition. It had been working — four phases, four finds. Rejected because the hit rate was evidence that intuition was sampling from a badly-mapped space, not that it was good at choosing. A sweep says how much is left and of what kind; picking cannot.

Close every [open] that the statute touches, and stop there. Rejected. Four of the nine closed cleanly and the other five were more interesting as corrections than they would have been as closures: “transportation’s rates are measured, not chosen” is a better fact than “transportation’s rates match statute”, and it would have been lost by treating the audit as a checklist.

Introduce the four-kind taxonomy as an ontology change — a kind field on parameter. Deferred, not rejected. It is the right shape and it would touch every parameter node; doing it in the same pass that discovered the distinction would mean committing to a classification derived from nine examples. Recorded here so the next pass can decide on more.

Change the notation to separate “unknown” from “unentered”. Also deferred, and for the same reason — it is a change to how the whole corpus is written, and this pass found the need rather than the design.