Every figure says its year

A page on this site puts the FY2027 formula beside a 2024 tax year, a 2024-25 report card, an FY2022 Census survey and a five-year forecast reaching back to FY2020. The header said FY2027 — the model’s year, and the year of nothing else on the page.

Three things made that worse than an omission.

The reckonings are not the same kind of thing. A tax year is a calendar year whose revenue reaches a district in the following fiscal year. A school year straddles two calendar years and is published as 2024-25. A fiscal year runs July to June and is named for the June. Every one of those is “2024” to somebody, and rendering all three as a bare number invites a reader to set them side by side as though they were the same period.

The years were prose. They lived in Rust doc comments — /// Performance Index, 2024-25 — and in about 190 four-digit literals across 25 files in the web layer. One of them, 2024-25, was a string in an Astro <meta> description with nothing connecting it to the download it described.

And one of them was already wrong. Bundle::provenance is a hand-written paragraph naming every year in the feed. It said “millage is TY2023” while all 609 districts carried tax_year: 2024. Nothing could have caught it: a sentence and a column cannot disagree in a way a compiler or a test notices, and the sentence is the half a reader sees. This is the same failure shape connect::index’s node count had when it was the literal 58 — a regenerated constant produces no diff.

Separately, the site had no tooltip mechanism at all. Every title= in web/src/ was the page-title prop; no <abbr>, no aria-describedby. The terms most likely to be misread — “per equivalent pupil”, “effective millage”, “on the guarantee” — had no affordance saying they meant something specific.

The year is a datum, derived, before it is a label. crates/bundle gains a series_years index: one row per block, each carrying series, kind (fiscal | tax | school), a ready-to-render label, and its source. Every label is read off the block it describes — the maximum tax year actually present, the ends of each series, MODEL_YEAR — so a year that moves in the data moves here.

Bundle::provenance stops restating years, and a test forbids it growing them back. One place, or the two drift again.

The report card is the exception that proves the rule: a school year has no number to read off a row, so it is a named constant — checked against the fixture path, report-card-2425-…, so swapping in next year’s download without moving the label fails the build.

Every card showing figures displays its own year, as a chip in its h2, driven from the index. Where a card genuinely mixes reckonings it says so — 2024-25 · spending FY2025 — rather than picking one. A card with no figures carries no chip: dating an absence says nothing.

Tables follow one rule. Where the rows or columns are on one year, the chip carries it and the table does not repeat it. Where they are on more than one, every row says which. Half-labeling is worse than none, because it implies the unlabeled rows share the labeled one’s year.

A defined term is the trigger, with a dotted underline — not a (?) beside it. It renders as a real <button> with aria-describedby pointing at a sibling that is always in the document, revealed by :hover and :focus-within. No script, no hover dependency, and deliberately no title, which would be announced alongside the description.

The stale year was the argument. Nothing about the plan changed when TY2023 turned out to be wrong, but the priority did: the derivation stopped being tidiness and became the fix for a live defect. It is also the reason series_years is derived rather than a second hand-written table, which would have been the same mistake with better formatting.

Zero cards with figures now lack a year, asserted on the built pages rather than per renderer — a card added in an .astro file is as unchipped as one added in src/lib, and only the rendered page sees both. That let the rule be a hard zero rather than a ratchet.

The scan found the wrong branch chipped, three times. A script that put the chip on the first h2 after each data-part attached it to the empty-state card in renderNationalPosition, renderOutcomes and renderSpendingByFunction — the branch that says a district has no data. Reading the built HTML rather than the source caught it; reading the source would not have, because both branches look identical there.

A second chip check was needed, and it is the one that matters. A test that every card has a chip passes just as happily if every chip on the page reads FY2027, which is the state the page was already in. So there is a second assertion: the chips on one page name more than one data-kind, and the bare 2024 carries the long form saying it is a tax year.

The glossary is short on purpose. Twelve entries, and the bar is that a reader who knows ordinary English gets the term wrong — not merely that it is jargon. A page where every third word is underlined has taught the reader to ignore the underlines.

term() throws on an unknown slug rather than falling back to bare text. A silent fallback would let a renamed entry strip the definitions off a page with nothing to notice, which is the same class of failure as a link resolving to a plausible 404.

Contract 34.0.0. Breaking rather than additive because series_years is required: a consumer that renders figures from this feed and has never had to say what year they are on is a consumer whose pages were making the mistake this block exists to stop.

A year field on every numeric field. Rejected. The year is a property of the source, not of the number: the report card publishes one year at a time, the Census survey publishes one year at a time. Hanging a year on each of the two hundred-odd numeric fields would repeat the same string two hundred times and still not say which of them moved together.

One year on the page header, stated more loudly. Rejected, and it is what the page already did. The problem is not that FY2027 was quiet; it is that it was true of one card in six. A louder wrong label is worse than a quiet one.

Keep the years in the provenance paragraph and render it prominently. Rejected on the evidence — the paragraph was wrong, and prose is exactly the form in which it could be wrong without anything noticing. The paragraph survives as narrative and no longer carries a year.

<abbr title> for the defined terms. Rejected. Invisible on touch, inconsistently announced by screen readers, and it puts the definition somewhere a reader cannot select or follow a link from. The corpus link inside each definition is half the point.

A (?) affordance beside each term. Rejected on density. On a page with forty figures that is forty pieces of furniture; the dotted underline is the older convention and carries its own discoverability without occupying space.

A JavaScript tooltip library. Rejected on the house rule this site already follows — the district filter and the basis toggle both work with script off — and on a stronger version of it: a definition that needs a bundle to appear is a definition most crawlers and some readers never get.

Color the chips by reckoning. Rejected. A tax year and a fiscal year are equally authoritative, and coloring one would imply a hierarchy that is not there. The distinction is carried by the label and spelled out in the title.