0003 · The identifier prefix is the reference's demand

Status: accepted, 2026-08-29. Decided by the maintainer. Issue: #11

---

The gap this closes

A type system needs two sides. ExactTeX had one.

\figure(fig:main) declares a Figure — the class is known the moment the scanner reads the entry token. But @ref(fig:main) demands nothing, so there is nothing a declaration can contradict, so nothing can be checked. The class was computed and discarded.

In TypeScript the demand is written at the use site:

function f(x: number) { … }
f("a")                       // error: the signature says what it wants

The equivalent question here is how @ref says what it wants.

The decision

The prefix before the first : is the demand. @ref(fig:main) demands a Figure. Pointing it at a \table(fig:main) is XT1004.

Three options were on the table. This one was chosen because it costs nothing in the corpus it has to work on, and because the syntax was already frozen — it adds no new form.

Option
The prefix is the demand	chosen
@ref:figure(x) states it explicitly	rejected: more to write, and it changes frozen syntax
No demand; the class only powers hover and rename	rejected: leaves the check that motivated the type system unbuilt

Where the default map comes from

From the published convention, not from a corpus. LaTeX has no specification for label names, but the convention is documented, and two sources agree. Both transcribed on 2026-08-29:

Source	Prefixes it names
LaTeX2e unofficial reference manual	ch, sec, fig, tab, eq
Wikibooks, LaTeX/Labels and Cross-referencing	those five, plus subsec, lst, itm, alg, app

The reference manual: “A common convention is to use key names consisting of a prefix and a suffix separated by a colon or period.” Wikibooks: “It is common practice among LaTeX users to add a few letters to the label to describe what you are referencing”, and “You are not obligated to use these prefixes.”

So the default map is:

[prefixes]
figure    = ["fig"]
table     = ["tab"]
section   = ["sec", "subsec", "ch"]
appendix  = ["app"]
algorithm = ["alg"]
equation  = ["eq"]

lst and itm are documented but have no admitted entity class, so they stay unmapped and demand nothing.

What the corpus is used for, and what it is not

One author’s corpus cannot show that a convention is universal. It can show that something occurs, and that is the only claim made from it here.

Measured across 1,374 labels in that corpus: the documented prefixes are used, and so are six spellings the documentation does not name — appendix, ssec, subsubsec, cap, algo, def. Together they account for 55 labels.

That is an existence proof and it settles one thing only: the map has to be replaceable. A fixed map built from the published convention would report a type error on those 55 correct labels, in the first real project it met.

It settles nothing about how common the convention is. tests/experiments/label-prefixes/ says the same in its limits section.

An earlier version of this record over-claimed

It said the decision “costs nothing in the corpus it has to work on” and that the convention is “already there”, citing agreement rates from that one corpus as though they measured LaTeX practice. They measured one author’s habit. The decision stands, but it stands on the documented convention above and on the fallback in the next section, not on that sample.

Where it does not fire

An unmapped prefix demands nothing. def:sixtuple and lst:algorithm are not in the map, so the reference is ?O and cannot fail. Adding a prefix to the map is how you opt a class into checking, and not adding one is a valid permanent state.

No prefix demands nothing. The 28 unprefixed labels keep working exactly as they do today.

A target of unknown class cannot fail. @ref(fig:x) pointing at an @id attached to a \newtheorem ExactTeX does not model is Known(Figure) ~ ?O, which is consistent. Both sides must be known, per checking.md §3.

This is the same renunciation TypeScript makes with any, and in a document most entities are any.

What it catches

The failure that motivated it, and that no LaTeX tool reports:

\table(fig:results) { … }          % renamed from a figure, label kept
As shown in Figure~\Cref{fig:results}

LaTeX compiles it. \Cref reads the counter and prints “Table 4”. The prose says “Figure”. The PDF is wrong and silent. The author uses \Cref 110 times across 4 projects, so this is a defect they can actually hit.

What would reverse it

Evidence that the published convention is not followed widely enough for an unmapped prefix to be the rare case — say under 90% agreement between prefix and environment across many authors. Then most references would demand nothing, the check would rarely fire, and the explicit form would be worth its verbosity after all.

That evidence does not exist yet in either direction. tests/experiments/label-prefixes/ runs on any corpus; pointing it at a broad sample of published sources is the measurement that would settle it, and it has not been taken.