# Status and next constructions ## Where the exemplars stand | Article | Definitions | Claims | Verified | Not read yet | | --- | ---: | ---: | ---: | --- | | Blest (`Papers/Rockel2026ExactBlest/Source`) | 37 | 48 | 48 | Lemmas 2.1, 2.2 and 4.2, the last two sentences of Corollary 4.6 | | Rho-gamma (`Papers/AnsariRockelSteinmassl2026RhoGamma/Source`) | 11 | 8 | 8 | the parametrization `G`, `P` of Theorem 1.1, the lemmas of Sections 3–5 | Every result of the first article whose statement uses only the article's own definitions and the vocabulary is compiled and certified: Theorems 1.1 and 1.2, Lemmas 3.1 and 4.1, Propositions 4.3 and 4.4, Corollaries 3.2 (with its proof, sentence by sentence) and 4.5. What remains needs constructions the grammar does not have yet. A third research article, checked as it was written, showed most of the items below as well. ## Next constructions 1. **Random pairs and expectations under a coupling.** "If $(X,Z)$ follows the coupling $\gamma$, then $\alpha(F)=12\,\E[XZ]-3$", "Every $\gamma\in\Gamma$ satisfies $\E_\gamma[c(X,Z)]\le\int_0^1\varphi+\int_0^1\psi$". Reading: `E_γ[f(X,Z)]` is the integral of `f` against the measure `γ`, and "the coupling associated with `F`" is a definite description. 2. **Results that quantify over functions with properties.** "Let $c:[0,1]^2\to\R$ and $\varphi,\psi:[0,1]\to\R$ be continuous functions such that …", "let $f:[0,1]\to\R$ be bounded and nondecreasing". Binders of function type with hypotheses, and finite sets (`N_1, N_2`). 3. **Arbitrary probability spaces.** "random variables on the same probability space" quantifies over all probability spaces, which needs a vocabulary whose model is parametric in the space. 4. **Anaphora to extremizers.** "For $n=-1$, this NOUN is $W$.", and the identification of a maximizer as `A_w^\top` or `B_b^\top`. 5. **Parametric families defined by a construction**, such as the extremizers `F_\theta` of the second article (Section 3, `def:extremizers`), which Theorem 1.1 uses through `G(\theta)` and `P(\theta)`. 6. **Markov kernels.** A notion such as "We call a NOUN $F$ \emph{NAME} if its Markov kernel can be chosen such that, for every $v\in(0,1)$, the map $t\mapsto K_F(t,[0,v])$ is non-increasing." is a notion in the running text, which the grammar reads, but its condition needs the kernel `K_F` in the vocabulary, as a kernel determined up to null sets, with "can be chosen such that" read as an existential over its versions. 7. **Partial derivatives.** `\partial_1F(u,v)` in a definition, with its definedness condition (the partial derivative exists almost everywhere). 8. **Attainment and equality cases.** "the upper bound is attained by $S_s$ and by no other NOUN", "with equality if and only if …", "occur only at $\alpha=\frac{5}{32}$". 9. **More of the notions that an article does not define.** The list of the preamble and the sentence `Let $\chi(G)$ denote the chromatic number of a graph $G$.` are read. Not yet: a noun that is a kind of another (`a tree is a graph`), an adjective with a hyphen or a formula (`$K_r$-free`), a relation between two objects (`$H$ is a subgraph of $G$`), notation of the subject (`\otimes`, `\dim`), the kind of a quantity from its domain (`$d(G)\in\mathbb N$`), and a sentence that names the source of a notion (`a matroid in the sense of \cite{oxley}`). Each construction gets one documented reading in `Leanify/Grammar.lean` and `Leanify/Doc.lean`, an example in `docs/writing.md`, and an exemplar sentence that compiles and is certified. ## Beyond constructions These points limit what a certificate says. None of them is a construction of the grammar. 1. **Organizing prose.** One sentence of a proof without a reading makes the proof incomplete, and its result is then checked, not derived. A few sentences that only organize a proof are read as such (`We distinguish two cases:`, `We argue by induction on $n$.`); most are not. The check counts the complete proofs. A claim after an unread sentence is read without what the sentence says, and the reports name and count these claims. What is open is to read more of the sentences that end a hypothesis (`In the other case, …`, `which is absurd`), and a way for an author to say that a sentence is only a remark. 2. **The size of a step.** The certificate tries one fixed tactic, `leanify_step`, on every derived step and reports the steps it closes as *standard*; a result with another step is *argued*. What a careful reader accepts as one step is wider than what the tactic closes (a rule for real powers, a known inequality, a compactness argument in a line, the induction principle): a list of named standard facts that a step may use without the label of an argument is open. 3. **Statements and steps that hold for an empty reason.** For every set of hypotheses under which a claim is made the reports say whether they are shown to hold together (`satisfiedK`). The definitions of the article are not covered yet: a notion that no object satisfies makes every statement about it true. 4. **Defaults of the grammar.** Where a limit, a sum, a set or a threshold states no domain, the letters `n`, `m`, `k`, `j`, `i`, `N`, `M`, `K` are integers and every other letter is a real number. A convention sentence declares further letters to be integers (`Throughout, $p$ and $q$ denote integers.`), counting the elements of a set whose letter has no stated domain is refused, and `\leanifyreadings` prints the convention. Declaring an index letter to be a real number, or a letter to be a prime or a positive integer, is not read yet. 5. **Readings in the PDF name the vocabulary of the package option.** The vocabulary an article is compiled with is named in its Lean file, and `\leanifyreport` prints the option `vocab` of `leanify.sty`. The two should be one. 6. **The trust in the compiler.** By default the compiler is run by Lean's evaluator, and `leanify audit` has the kernel evaluate it. What stays trusted in both cases is the grammar: it says what a sentence means. The readings it gives are printed, and a second, independent reader of the same grammar would be the next step.