Drafts are not legislation

The request was support for authoring draft legislation and extending a draft into a scenario. Three kinds of draft were confirmed in scope: a real bill before the General Assembly, a counterfactual rewrite of an act that did pass, and an invented change nobody has proposed. Authoring is corpus-side — a node and a skill, not a browser tool. Provisions the calculators cannot run are to be recorded as [open] rather than rejected.

The machinery to price a proposal already existed and none of it needed changing. parameter nodes carry a simulation_key that binds to the calculator; project::policy::Policy holds five levers and edfund-project runs them; scenario is already a valid node before it is run; scenario-delta already refuses a total without the count of districts it fails to reach. What was missing was the document at the front: nothing in the corpus could hold a bill that had not passed.

Two prior decisions bore directly on where it should go. “Ontology” fixed legislation as a UFO Kind meaning enacted, and “Advocacy literature” had already faced the near-identical question one document earlier, when a model bill published by a national organization needed somewhere to live and was given model-policy rather than a flag on legislation.

Add draft-legislation as an eighteenth ontology class, a Kind, holding all three draft kinds with provenance distinguishing them.

Nine properties, and one of them is the reason the class is worth having. provenance, designation, general_assembly, sponsors, status, text_read, question, provisions — and unpriced, mandatory on every node including the ones where the answer is none.

A draft produces scenarios rather than being one: simulated-by runs from the draft to one or more scenario nodes, and the reciprocal is declared on that class too. A draft that passes gains an enacted-as edge to a new legislation node and is not edited.

Seeded with one node — a counterfactual H.B. 96 that refreshes the base cost reference year it froze — plus skills/draft-authoring. No Rust changed, and no retrieval was attempted.

The class split is model-policy’s argument at one document’s remove, and it gets stronger as the document gets closer to law. That class’s definition states the hazard outright: holding a template and a statute in one class “would let a reader take a template for a statute, which is the specific error this class exists to prevent.” A model bill published by an advocacy organization is obviously not Ohio law. A bill numbered, sponsored, and pending in the 136th General Assembly looks exactly like Ohio law, so the same reader error is both likelier and harder to catch. The split that was defensible for model-policy is close to compulsory here.

legislation needed no edit at all, which is the check on whether the split was right. Its Kind framing, its properties and its edges are all still true of exactly the things in it.

The label carries the provenance, because the page cannot rely on the reader looking. A node’s label renders under the h1 and into every social card, and a draft shown with a bill number and a list of sections reads as an enactment. So a counterfactual takes a Counterfactual: prefix, an introduced bill states its stage, and a hypothetical draft may take no designation at all — a number is the one thing a reader will believe without checking.

unpriced is the field that does the work, and its ground is already load-bearing elsewhere. scenario-delta cannot construct a total without its Reach: there is no method returning a bare dollar figure, so a caller who wants the headline receives the count of untouched districts in the same value. unpriced is that rule one level up — a cost cannot be stated without the count of provisions it fails to price — and it is needed because the failure is the easy one. The provisions that bind to levers are the provisions that produce a number, so a pass that binds as it reads yields a draft consisting of precisely the clauses the model can reach, reported as the bill.

The runnable surface bounds this class much more tightly than it bounds scenario, and that is a new exposure rather than an inherited one. A scenario is authored around a perturbation the levers can express, because whoever writes it chooses the question. A draft is authored around provisions somebody else chose. Five levers — the guarantee rule, aggregate base cost, the minimum state share and two phase-in dials — against a budget act that also moves special-education weights, transportation reimbursement, the scholarship deduction and the capital channel. The last two are outside the model entirely rather than merely unlevered.

What share of a real bill can actually be priced is therefore an open empirical question, and the seeded node does not answer it: one counterfactual with a single runnable provision has complete coverage by construction, which is the least representative case available. open

NARROWED by the runner phase. A second node, Fund the Plan and Retire the Guarantee, was authored precisely because the first could not carry this: five provisions, of which two price. It does not settle the empirical question — a constructed bill is not a sample — but it does establish that the three unpriced provisions fail for three different reasons, one unlevered and two outside the model, which is a distinction a single ratio would have hidden. A third node since — H.B. 643 of the 136th, a real pending bill — prices at 0 of 1, which is a lower bound rather than an estimate. What the ratio looks like for a budget act stays open.

The interaction between provisions is larger than the decision anticipated, and it changed what the runner returns. This record proposed reporting a residual between the combined run and the per-provision runs, by analogy with regime-diff. Measured on the second draft, the residual is $75.1M against a combined figure of $143.9M — 52%, not a correction term. Its two provisions priced apart say −$219.0M and run together say −$143.9M, because a district the base cost refresh lifts off the guarantee is no longer standing on the floor for the phase-out to lower. So Priced::residual is a first-class result rather than a diagnostic, and the CLI prints it under the attribution with the reason attached.

And the two levers turn out to cover the state, which is not the same as partitioning it. The refresh leaves 253 districts unmoved and the phase-out leaves 315, and the combination leaves zero. What is disjoint is the two unmoved sets — 253 held on the guarantee throughout and 315 paid by the formula, 568 between them — which is the structure Phasing Out the Temporary Transitional Aid Guarantee established.

CORRECTED in review. This paragraph first said the two levers “partition the state” and that “neither provision reaches the other’s population”. Both are false, and the node’s own figures disprove them two paragraphs earlier: 356 moved by the refresh and 294 by the phase-out sum to 650 against 609 districts, so 41 are moved by both — precisely the 41 the refresh lifts off the guarantee and the phase-out would have cut. The test asserting it was worse than the prose: it pinned 253 + 315 == 568, which is the arithmetic disproof of a partition, under a name claiming one. A tidier result was asserted than the numbers supported.

A lever costs three things and the price is now written down. A field on Policy needs a matching entry in bundle’s checkpoint_policies and a mirrored implementation in web/src/lib/policy.ts, or verify.ts refuses to render the scenario page. That refusal is the mechanism keeping the browser’s second implementation of the formula honest. It is recorded in the skill and in the class’s ACTIONS.md so that “just add a lever” is priced before it is promised.

Provisions do not sum, and the skill refuses to let them. Raising base cost costs $47.7M per point at +2% and $72.5M at +20%, because each increment lifts districts off the guarantee onto a formula that must then pay them — measured in Phasing Out the Temporary Transitional Aid Guarantee. So a draft’s cost is one combined run and the per-provision runs attribute it, with the residual reported rather than absorbed, on the pattern regime-diff already uses for a decomposition it cannot fully explain.

The introduced provenance is declared and unused. No connector reaches a pending bill: ohio-laws serves the Revised Code as it stands and ohio-session-laws serves acts as passed. Declaring the value now costs nothing and avoids retrofitting the field onto every node once the connector exists — the same reasoning “Proposals” used when it approved twelve implied edges as enrichments that are “cheap to declare now, expensive to retrofit once instances exist.”

RESOLVED by the retrieval phase. ohio-bills was added on the-acts-themselves’ precedent and H.B. 643 of the 136th has a node. The connector stays at retrievable rather than wired, and that is the decision worth recording: turning a bill into provisions is reading rather than extraction, so there is no parser and there should not be one. What the retrieval is for is the text, pinned, so a node can be checked against the document it claims to describe.

The unpriceable draft was not anticipated, and it was a defect rather than a gap. A bill whose every provision falls outside the model produces a combined policy identical to current law, so Priced::cost computed exactly 0.00 — and $0.0M in a cost column reads as this bill is free, which is the opposite of this bill was not priced. The first real pending bill wired here is exactly that shape, and it would have published the wrong reading.

cost() and residual() return Option now, the CLI prints the absence rather than the figure, the feed writes null, and the scenario page says outright that nothing on it is the bill. The shape is not exotic: it is the ordinary one for a bill about vouchers, transportation or facilities, which is most of what touches school funding outside a budget act.

A status or enacted field on legislation. Rejected, and it was the obvious option. The class is a Kind whose identity is fixed at enactment; a draft has no signing date, no effective date, no vetoes, and its General Assembly is a hope rather than a fact. Four of seven properties would be permanently empty, which is a description of a different class. And the reader hazard is the one model-policy was split off to prevent, in a sharper form.

Three classes, one per draft kind. Rejected. The three differ in provenance, not identity: they hold the same provisions, bind to the same parameters, and enter the graph at the same places. Three classes would triple every edge declaration to encode one string.

Make the draft a scenario with a text field. Rejected, and this was the closest call. A draft and a run are different objects with different lifetimes: a six-provision bill needs one run per provision plus a combined one, the same run can price two drafts, and re-running against a new fixture year produces a new scenario while the bill’s text is unchanged. Folding them would make “the draft was superseded” and “the run was superseded” the same event, when the second happens routinely and the first almost never.

Reject provisions the levers cannot run. Rejected by the scoping question, and it would have been wrong anyway: it rules out substantially every real bill, since Ohio legislates school funding inside the operating budget. What it would produce is not a stricter corpus but a corpus of drafts that happen to be five levers wide, reported as bills.

Structured provisions as a real data type on the node. Deferred rather than rejected. corpus-node.json allows properties as a flat string map, so provisions is prose in a fixed form for now. When a runner needs to read it mechanically the binding belongs in a committed fixture under crates/ with a test reconciling it against the node — the shape web/tests/unit/year.spec.ts already uses for statewide constants, where the feed is authoritative because crates/foundation computes it and the corpus may not contradict it. Widening the node schema for one class is the larger change and it can be made later on evidence.

RESOLVED as described. crates/project/fixtures/draft-provisions.tsv is authoritative and crates/project/tests/a_draft_cannot_hide_what_it_did_not_price.rs holds the node to it — on the parameter slug and the statutory authority rather than the title, because a title is prose and will be reworded while a binding that changes is a different provision. The node schema was not widened, and nothing has yet wanted it to be.

Emit amendatory legal text — the section reprinted with strikes and inserts. Rejected for now. “Reading an amending act” establishes how Ohio prints one, so the format is known, but nothing here could check that generated language says what the provision means, and unverifiable legal text next to verified figures is the worst pairing available. The plain-English analysis carries the same information and can be checked.