The prose a node holds

yidam’s prose checks named one key. node-too-long counted the lines of description, missing-description reported a node that had none, and yidam embed composed the text an index is built from out of the label and that one field. Meanwhile the claim counter reads the whole file, which is why 780 [verified] tags are counted across prose the length checks could not see.

This corpus has written in four fields since “The four genres of a description” split them: summary leads a page, description is the body, findings is what the corpus established itself, and revisions records what it withdrew. All 129 nodes carry summary and description, 91 carry findings, and 36 carry revisions.

The gap was already measured, from the other side and without a name for it. The kuten record put the median node at 118 lines against a band of 35 to 62, and observed that the median description is 21 lines while all four prose fields together run to 34 — “yidam 0.10.0 narrowed node-too-long to the description for exactly this reason, and this band still reads the whole file, so the two disagree about what a node’s length is.”

A fifth of what any corpus writes is not at the top level at all. Upstream measured 2,763 nodes across sixteen corpora: 68.6% hold a block scalar nested inside another key and 83% of that sits under properties:. Here it is 129 of 129 nodes and 246,194 characters — a sensitivity that explains what moves when a parameter moves, a caveats that says how a metric misleads, a vetoes that is the only place six line-item vetoes are written down. Nothing that read prose could see any of it.

yidam cli/v0.11.0 (RFC-0031) gives the ontology two ways to say so, and both were absent rather than declined: absent both declarations the effective set is [description], which is exactly what 0.10.0 did.

The four top-level fields are declared universally. .yidam/corpus/universal.yml carries prose: [summary, findings, revisions]. The effective set is {description} ∪ universal ∪ class, so description is never listed and naming the other three does not unname it. Universal rather than per class for the reason seeded_because is: these are how this corpus writes any node, and declaring them in eighteen ontologies would be eighteen copies of one decision with a nineteenth class silently missing them.

revisions is a list of records rather than a paragraph, and it is declared anyway. A withdrawal says why the corpus no longer stands behind a claim, in prose, with the test or source that settled it — 67 of them across 36 nodes. That is the node saying something.

A property is prose when the class declared it type: text, with three exceptions. 51 of 54 flagged, and the rule is the corpus’s own vocabulary rather than a judgment made 54 times: text is what this corpus already writes when a value is sentences, string is what it writes when a value is a name, a code or a path. The three excluded are named here rather than left to a reader to notice:

• formula-component/function_tex — LaTeX. \begin{aligned} and \text{...} are markup, and embedding them would put macro names into the text an index searches. The prose that explains the same calculation is in function, which is flagged.
• parameter/written_as — the phrases a formula writes a parameter as, one per line. A naming alias, median 39 characters.
• draft-legislation/sponsors — a roster, median 6 characters, usually “None.”

There is no universal form of the property declaration, and upstream is right that there should not be: a property belongs to the class that declared it, and two classes can use one name for different things. parameter/kind is 761 characters of argument about what kind of parameter it is; fiscal-period/kind is the word “biennium”. They are text and string respectively, which is the distinction doing the work.

web/src/lib/schema/corpus.ts learns the field first. OntologyPropertySchema is .strict(), so a prose: key in an ontology is a validation failure until the Zod schema accepts it, and .yidam/schemas/corpus-ontology.json is generated from that schema and read by the editor over *.ont.yml. Declaring in the corpus before widening the schema would have underlined 51 lines in the author’s editor and failed pnpm test:unit. readOntology is deliberately not changed: it maps a fixed set of keys onto the class page, and how the corpus is measured is not something that page has a reason to render.

OntologyClassSchema is not widened. A class may declare its own top-level prose: upstream, and no class here has a key the universal three do not already cover. Declaring the shape before anything uses it would be a schema saying more than the corpus has decided.

Nothing gates differently, and that was checked rather than assumed. yidam lint holds at 493 findings with no errors, the same six checks at the same counts, before and after. No node-too-long finding appears: the median node’s four prose fields run to 34 lines and the ceiling is well above that, so the number this corpus was measured on being wrong never meant it was over.

The text an index would be built from very nearly doubles. Across the same 174 files yidam embed writes 461,947 characters at this pin with neither declaration, 745,906 with the three top-level fields, and 918,794 with the property flags as well — +99%, and the two halves are 284K and 173K. What stays out of reach is the 30K in string-typed properties that the alternatives below decline to flag, plus prose under links: and under coined top-level keys, which upstream puts at 17% of the nested prose it measured and ships no mechanism for.

That growth buys nothing today. There is no vector index here and .yidam/config.toml says why: index-build needs the index feature, which is not in the released binary’s default set, and [due] index_after is omitted rather than set so no clock reports a debt nobody intends to pay. .yidam/embeddings/ is a derived artifact and git-ignores as one. The declaration is what makes the measurement honest; the index is a separate decision and still not taken.

missing-description can now be satisfied by a property. A node whose substance is a transcription under properties: stops being reported as having said nothing. None is today — all 129 carry both summary and description — so this is a hole closed before anything fell in it.

Two files hold one rule, and now they hold two. universal.yml already restated UNIVERSAL_PROPERTIES and OBSERVATION_PROPERTY from web/src/lib/schema/corpus.ts, with a comment saying to change both together. prose: adds a second such pair — the corpus declares it, the Zod schema permits it — and the same warning applies. The site does not read the declaration, so a drift here is a validator disagreeing about what an ontology may say rather than two validators disagreeing about the corpus.

Declare the top-level fields and stop. The smaller change, and it needs no schema work at all: universal.yml is read by yidam and by nothing in web/, so prose: could have been added to it without touching OntologyPropertySchema or regenerating the editor schemas. Declined because it reaches 462K to 746K and leaves the other 173K — a fifth of what this corpus writes — in exactly the state the change is meant to fix. The property declarations are where the measurement was furthest wrong, and they are the half that needed the schema work.

Flag every property carrying a block scalar. 129 of 129 nodes hold one, so the mechanical rule is available and would have needed no exception list. Declined because it is not a rule about prose: function_tex is 14 multi-line LaTeX values and simulation_key is a single-token identifier that happens to be written as a block in places. A rule that cannot tell a paragraph from a macro is a rule that puts \begin{aligned} into an index.

Flag the string properties that carry prose too. program/mechanism, program/appropriation_line, formula-component/statutory_basis, parameter/statutory_basis and revenue-stream/legal_basis are declared string and hold paragraphs — about 30K characters between them. Declined here, and it is the honest half of the rule above: if text means prose then those five are mistyped, and the fix is yidam migrate retype, not a prose: flag that would make the two declarations disagree on the same property. That is its own question and its own commit.

Restate prose: in the site’s own vocabulary check. web/src/lib/schema/corpus.ts holds the second copy of the universal-property rule and could hold a third of this one, checking that a flagged property really does carry prose. Declined: there is nothing to check it against except length, the site renders these values already without needing to know, and a third validator that measures a different thing is how the first two came to disagree.