Generated from cat-harness/skills/authoring/authoring-core/milnor-exposition-standard.md — do not edit here.
✎ Edit this page’s source

The Milnor exposition standard

Three files sent readers here for this section and it did not exist. skills/authoring/authoring-core/exposition-swarm-drain.md cited “one-voice style guide → ‘The Milnor exposition standard’” three times, one of them naming a .claude/skills/local/one-voice-style-guide.md that is not on disk. The eight hallmarks lived only inside the expo-milnor-clarity criterion’s description — eight words, no rubric — while that criterion is a strict gate passing only on a perfect 16/16. Scoring against eight bare words is how eight reviewers produce eight different sixteens. Bean bfmf.

It is now derived from the paper, not from the name. The source is the milnorlink library entry — John Milnor, Link Groups, Annals of Mathematics, Second Series, 59(2), March 1954, pp. 177–195 (JSTOR 1969685), ingested at journal-page granularity so a citation to “p. 179” resolves to its sections/page-179.md. The entry is held by the science layer’s library graph and reached through that declaration, never through a path written here: this layer sits below it and does not link up (bean rfuq). The science layer’s skills/voices/README.md links down to this skill instead.

The path moved with the corpus (bean frs5): the platform holds no content, so milnorlink went to the science instance. This link went on pointing at cat-harness/library/ for a day and nothing reported it — check:subgraphs only began walking this directory once main added its zero-dangling assertion, which is what finally surfaced it. Every rule in the milnor voice skill’s voice.json (skill_fetch milnor) quotes a page. 8899 words, 496 sentences; measured 2026-09-19.

What the measurement found, including where WE are wrong

Three findings contradict rules stated elsewhere in this very file or in the voice-* criteria, and the paper wins:

clearly is proof economy, not editorializing. voice-editorializing lists “clearly”, “obviously” and “trivially” and fails a block for them. Milnor uses clearly fourteen times in 8899 words, and every one routes the reader’s effort away from a verification that is routine — “Clearly the relation of homotopy is reflexive, symmetric and transitive”, “The inclusion map … is clearly a homotopy equivalence”, “This is clear for the case n = 0”. Not once does it mean “this is remarkable”. Editorializing spends the reader’s attention on the author’s opinion; this spends none and saves some. In a mathematical block, clearly before a check the reader can perform is correct, and a finding against it should be overruled with a reviewer entry. Clearly attached to a claim the reader cannot verify in their head is the real defect — and no phrase list separates the two.

“Never I” is too blunt. §”Author Voice Profile” above says “we” for joint reasoning with the reader, never “I”. Milnor’s split is sharper than a prohibition: exactly one first-person singular in the paper — “I am indebted to R. H. Fox for assistance in the preparation of this paper”, at the end of §1 — against twenty uses of we, every one joint reasoning. Zero our, zero the author, zero you. So: a personal debt is stated personally, in one sentence; the mathematics is done with the reader.

Concrete-before-abstract is a property of the PAPER, not the section. Read §2 alone and it opens on pure abstraction — “Let M be an open 3-dimensional manifold which possesses a regular triangulation. Let C be a circle.” A per-section rubric scores that 0. It is a 2, because §1 already gave the picture in words a reader can hold. Never score H2 on a definitions section in isolation.

Two more numbers worth having: no superlatives at all (zero surprisingly, remarkably, interestingly, beautiful, elegant), and one hedge in the whole paper — “a (possibly redundant) list” — which qualifies a mathematical fact rather than the author’s confidence. Median sentence 17 words, 90th percentile 29, longest 52.

The eight hallmarks, and what 0 / 1 / 2 looks like

Each scores 0, 1 or 2. Sixteen is the only passing total, so a 1 is a finding: “acceptable” is not the bar. Every row’s evidence is in the milnor voice skill with its page.

	hallmark	0	1	2
H1	Economy — the opening is a Summary, not an Introduction	A paragraph could be deleted with no loss	Sentences restate their neighbours, or a closing line summarises what was just said	Nothing can be removed without losing content
H2	Concrete before abstract — scored over the unit a reader reads through	Opens on the general definition and never lands	An example arrives, but after the abstraction it was meant to motivate	The informal instance precedes the formal definition
H3	Why before what — the tool’s purpose precedes the tool	The construction appears with no indication of what it is for	The motivation is present but buried after the construction	The reader knows what problem is being solved first
H4	Uncluttered notation — a symbol is introduced where it is needed	Notation used once; collisions; decoration where a word would do	More indices than the argument needs	Every symbol earns its name and none is overloaded
H5	Linear argument — routine verification is dispatched in a sentence	A forward reference to a result not yet stated	One aside the reader must park and return to	Read once, top to bottom, and it follows
H6	Strategy before execution — say what the proof will build and why that finishes it	The argument exists only in the displays; the prose says “we compute”	Prose narrates the steps without saying why they follow	A reader who skipped every display would still follow
H7	Right-tool framing — say what the method requires, in a clause	A heavy method with no indication why	The right tool, with no sense of what the alternative would cost	The reader sees why this tool and not the obvious other
H8	Respect for the reader — countable: no superlatives, no hedging, no second person	Tells the reader what is beautiful, surprising or important	Hedges a claim the text can support	States what is true and lets the reader judge

H6 is the most transferable of the eight. Lemma 3’s proof opens by saying what will be built and why that finishes the job — “A natural isomorphism … will be constructed for i = 0, 1. Since a homotopy … induces a homotopy equivalence …, this will complete the proof” — and only then constructs it. The reader knows the destination before the first step, so no step is a surprise (p. 179).

Scoring one block

Read it whole, once, without scoring. H5 and H6 are properties of the reading and cannot be recovered from a second pass that already knows the argument.
Score H1–H8 and write the evidence for every score below 2 — the line, and what the 2 would have been. A bare number is not a finding.
milnor-brevity is the companion and its resolve discipline is STRICTER. A repeat may be deleted only where the two occurrences are semantically identical. Where the content differs you may not delete, and you never delete maths — a diagram, an equation or a derivation. Where de-duplication would lose distinct content, keep it, or split the block if that helps explication.

Do not

Do not score from a diff. Every hallmark is a property of the block as a reader meets it, not of what changed.
Do not fail a block for clearly before a routine verification. The checker no longer does either — bean 2t41 narrowed it, and the exemplar now scores exactly one voice-editorializing finding (p194’s “Unfortunately these invariants are not strong enough”, which is real). What remains yours is the distinction no regex makes: clearly before a check the reader can perform is correct; clearly on a claim they cannot verify is the author asserting where they should be proving.
Do not fix H4 by renaming symbols across a chapter without checking what else binds them. Notation is shared.
Do not treat a 16 as a licence. The gate is necessary, not sufficient: a block can be perfectly expounded and wrong.