Style rules

1. Citations and references

Source: GitHub-flavored Markdown (GFM) footnote syntax and W3C semantic web standards.

Specification:

• Inline citations must use standard GFM footnote markers (for example [^1]).
• The reference block must provide a machine-resolvable URI as the ultimate source of truth. DOIs (digital object identifiers) should be formatted as HTTPS URLs (https://doi.org/…). If a DOI is unavailable, a direct link to a semantic web URI, an archival permalink, or a standard HTTPS URL is required.
• Human-readable bibliographic metadata (author, year, title) may precede the URI for readability, but the URI is the definitive, computable node.

Example

The `go to` statement should be abolished [^1].

[^1]: Dijkstra, E. W. (1968). Go to statement considered harmful. Communications of the ACM. https://doi.org/10.1145/362929.362947

2. Punctuation and quotations

Source: British English standard and “hacker’s style”, as documented in the Jargon File and the Oxford Guide to Style.

Specification:

• Logical punctuation: commas, periods, and other terminal punctuation must be placed outside the quotation marks unless the punctuation is explicitly part of the original quoted string.
• This ensures the exact data string is preserved for the parser without trailing syntax corruption.

• Correct She wrote “the format is stable”.
• Incorrect She wrote “the format is stable.”

3. Words used as words

Source: The Chicago Manual of Style, the MLA Handbook, and the APA Publication Manual. Contrast: AP Style, which marks mentions with quotation marks because wire and newsprint copy historically had no italics.

Specification:

• When a word or phrase is mentioned rather than used — the use–mention distinction — it must be set in italics. That is the Chicago, MLA, and APA convention, and the widely accepted standard in modern formal typography: books, academic papers, formal essays, and technical journals.
• In ce-linear (GFM) italics are emphasis markup: *word*. Underscore emphasis (_word_) is legal GFM, but CE writes asterisks so the mention marker is one token.
• Do not use quotation marks for a mention. AP Style’s quotation-mark convention assumes a medium without italics. CE is Markdown and HTML; italics exist, so that assumption does not apply.
• Quotation marks remain for quotations of speech or text (§2). Backticks remain for program text (identifiers, literals, file names).

• Correct The word set has many senses. (source: The word *set* has many senses.)
• Correct The term is defined as machine-readable.
• Incorrect The word “set” has many senses. (AP Style quotation marks)

4. Hyphenation of compound modifiers

Source: The Chicago Manual of Style, section 7.85.

Specification:

• When two or more words act together as a single adjective modifying a noun that follows them, they must be hyphenated.
• This eliminates part-of-speech ambiguity for the ACE grammar parser, clarifying that the words function as a single modifier unit rather than a sequence of separate syntactic elements.

• Correct It is a machine-readable format. (Hyphenated before the noun.)
• Correct The format is machine readable. (No hyphen when acting as a predicate adjective.)

5. Dates and times

Source: ISO 8601-1:2019 for complete dates and for times; international English / military date format for dates without a year.

Specification:

• Complete dates: a date that carries a year is a complete, sortable value, so it must be written as the ISO 8601 calendar date YYYY-MM-DD. No other ordering, no spelled-out month, no abbreviation.
• Dates without a year: ISO 8601 offers no year-less calendar date — the truncated --MM-DD form was removed in the 2004 revision, and reduced accuracy drops components from the right, never the left. So where the year is implied by context, CE uses the strict <Day> <Month> sequence with the month spelled out entirely and no ordinal suffix. Spelling out the month is what makes the value safe: a numeric year-less date cannot be read without knowing whether the writer meant day-first or month-first.

• Correct 2026-08-08
• Correct 8 August (only where the year is implied by context)
• Incorrect 8 August 2026 (a date with a year must be ISO 8601)
• Incorrect August 8th, 2026 (contains a comma and the ordinal suffix th)
• Incorrect 8 Aug (uses an abbreviation)

• Times: must use the 24-hour clock format hh:mm or hh:mm:ss, optionally appended with a UTC offset (for example 14:30 or 17:23:37-04:00). A date and a time given as one value join with the ISO 8601 T separator (2026-08-08T17:23:37-04:00).

6. Measurements and SI units

Source: BIPM SI conventions, adapted for natural prose.

Specification:

• Standard metric / SI units and widely accepted derivatives (such as Celsius) are permitted.
• A single space must always separate the numeric value and the textual unit or symbol (10 m, not 10m).
• Celsius and symbols: temperature can be expressed as degrees Celsius, degrees C, or using the standard symbol °C. If using the symbol, it must strictly be the Unicode degree sign (U+00B0) followed by a capital C.
• Decimal marker: ISO 80000-1:2022 and CGPM 2003 permit either a point or a comma. CE picks the point. Write 3.14, not 3,14.
• Digit grouping: a thin space every three digits, never a comma or a point (1 000 000, not 1,000,000). ISO reserves comma and point for the decimal marker; using either to group digits reintroduces the regional ambiguity the marker choice exists to close.

• Correct 20 degrees C
• Correct 20 °C
• Correct 3.14
• Correct 1 000 000
• Incorrect 20°C (missing space)
• Incorrect 20 º C (uses the masculine ordinal indicator instead of the degree sign)
• Incorrect 3,14 (comma as decimal marker)
• Incorrect 1,000,000 (comma as grouping separator)

7. Deterministic title case

Source: adapted from APA style (length rules) combined with CE dictionary part-of-speech (POS) logic.

Specification:

Because traditional title case requires human interpretation of context, CE requires a computationally unambiguous rule for headers and titles:

1. Positional override: always capitalize the first and last word of a title, regardless of POS.
2. Length override: always capitalize words of four letters or more.
3. Dictionary rule: for words of three letters or fewer, capitalize them unless they are defined in the CE attractor subgraph dictionary exclusively as an article (a, an, the), a coordinating conjunction (and, but, for, or, nor), or a preposition (at, by, in, of, on, to, up).
4. If a short word has multiple POS classifications in the CE dictionary (for example “up” as an adverb and a preposition), default to capitalized.

8. Pronouns

Source: singular they, epicenity, and the historical thou / you number collapse.

Specification:

• The unmarked third-person pronoun is they / them / their / theirs (singular they). Verb agreement is plural, as in ordinary English.
• Gendered pronouns (he, she, and their inflections) remain legal when the referent is a known person whose gender is part of the text. She sat on the bank. is CE.
• If a pronoun would be ambiguous among antecedents, repeat the noun or an alias declared in the graph. Do not ask the reader to guess.

CE enforces they as the unmarked form for one reason. English already sacrificed a singular/plural distinction for the exact same reason. Historically, thou was singular and you was plural; you absorbed both numbers. Epicene they is that collapse applied to gender. Epicenity in English is machine-forward: the agents that will emit most CE text are not gendered.

• Correct They sat on the bank. (gender unspecified or irrelevant)
• Correct She sat on the bank. (known person; gender is part of the text)
• Incorrect A user may change his password. (generic referent; write their)

9. References

Source: Halliday and Hasan, Cohesion in English (1976), for the endophora / exophora / anaphora / cataphora division and for the treatment of substitution and ellipsis as reference. Hawkins (1978) and Martin (1992) for homophora and associative reference. Clark (1975) for bridging. Ross (1969, 1970) for sluicing and gapping. Reinhart (1983) for bound variables. Webber (1988) for discourse deixis. Levinson, Pragmatics (1983), for deixis. CGEL for the surface categories.

Provenance note. Every trigger and target name below is standard published terminology. The two-axis arrangement, the six trigger families, and the numbering 1–30 are CE’s own organization of that material, not a taxonomy quoted from any single source.

Specification:

A referring expression points at something. The published literature names this phenomenon many times over because it answers two independent questions with one vocabulary: what the expression looks like, and where the thing it points at sits. Anaphora, cataphora, endophora, exophora, and homophora answer the second question. Deixis answers the first. Held apart, the two lists are short and closed. Merged, they generate a name for every cell of the cross-product and the subject looks unbounded. CE holds them apart.

A CE document therefore records reference on two axes. Axis 1 is the trigger: the token or construction that does the pointing. Axis 2 is the target: where resolution lands. Any trigger may take more than one kind of target, and the pairing is what a parser resolves.

Axis 1: triggers

The trigger inventory is closed. Every item is a function word, a determined noun phrase, or a construction defined by a gap. This is the list a parser scans for.

Pro-forms

Determined noun phrases

Quantificational and comparative

Ellipsis. Reference by absence. The trigger is a gap, not a token.

Non-nominal

Inference-mediated. These resolve through world knowledge rather than through a search of the text, so they have no bounded resolution procedure.

Axis 2: targets

The same trigger takes different targets in different sentences. That is situational-exophoric when the speaker points and anaphoric when it follows a clause on the page. One form, two resolutions, no need for a separate name per pairing.

Non-referential forms

Several triggers have uses in which they refer to nothing. A resolver that chases these produces a spurious antecedent, so CE marks them as non-referential rather than leaving them to the resolution pass.

• Expletive it — It is raining.
• Extraposition it — It is obvious that he lied.
• Cleft it — It was John who called.
• Existential there — There is a problem.
• Idiomatic it — take it easy; make it
• Generic you and they — You never know. They say so.
• Predicative noun phrases — John is the president. This ascribes a property; it does not refer.
• Bound variables — Every chef burned their soup.
• Frozen definites — in the end; the fact that; The Hague

Scope and policy

The paragraph is the atomic unit of a CE document, so the resolution search space for a text-internal target is a single paragraph. Triggers 1 to 27 have a bounded candidate set and a mechanical resolution procedure inside that space. Triggers 28 to 30 do not, because the relation that licenses them is world knowledge: room to ceiling is a good bridge and room to carburetor is not, and no rule over the text separates the two. That boundary is real rather than an accident of terminology, and CE sits on the near side of it.

• Endophoric reference (1 to 27) resolves within the paragraph. A cross-paragraph antecedent is not CE.
• Homophoric reference resolves through @graph. A uniquely identifiable definite is legal when it carries an IRI, which is what makes the sun computable rather than assumed.
• Exophoric reference to the physical situation is not resolvable from the document and is not CE. Participant reference (I, you) stays legal, since the participants are document metadata.
• Ellipsis (19 to 23) is rejected. The gap carries no token, so it cannot appear on @graph, and CE-0009 already requires every word to be a graph node.
• Inference-mediated reference (28 to 30) is rejected with a canonical rewrite rather than resolved.
• Non-referential forms are legal and are marked as non-referring, not resolved.
• Ambiguity is rejected. Where a trigger has more than one candidate target of matching agreement, repeat the noun or use an alias declared in the graph, as in §8.

• Correct The chef burned the soup. They had walked away from the stove. (Trigger 1, anaphoric, one candidate, same paragraph.)
• Correct The sun heats the ocean surface. (Trigger 10, homophoric, pinned on @graph.)
• Incorrect I entered the room. The ceiling was low. (Trigger 28. Write: The room had a low ceiling.)
• Incorrect She can swim and he can too. (Trigger 20. Write: She can swim and he can swim too.)
• Incorrect The chef told the waiter that he was late. (Trigger 1 with two candidates. Repeat the noun.)
• Incorrect Look at that. (Trigger 4, situational-exophoric, no target in the document.)

None of the above is machine-checked yet. These rules move into the linter rules list and take the next free CE-000n as they become executable.

10. Figurative language

Source: literal vs figurative language; dead metaphor; Lakoff and Johnson, Metaphors We Live By (1980), for source-domain mapping; MetaNet as a computational catalogue of those mappings; Open English WordNet for figurative readings that have already split into their own synset. Construction families and dead-metaphor clusters are taken from Banned: The Definitive Guide (announcement; catalogue), a personal anti-cliché style guide written to stop both human and model prose from regressing to interchangeable defaults.

Provenance note. The Banned guide is a taste document with an accumulation rule: one instance of a pattern may be fine, the fifth is a crutch. CE does not work that way. A sentence is CE or it is not. CE takes the guide’s negative inventory — the constructions and metaphor clusters that substitute form for content — and restates it as a closed allowlist. Register preferences the source records (erotica diction, a ban on sentence-initial And / But, avoidance of the em dash) are not imported. Dashes stay as in the conventions table. CE is not an AI detector; a word that happens to be over-represented in model output is banned here only when it is underspecified or figurative, not because a classifier would flag it.

Specification:

CE pins one sense per open-class word. A live metaphor uses a word as if it belonged to another domain: the vehicle’s predicates are true of the tenor only after a mapping the reader must reconstruct. That mapping is a second parse. Until it is an inventoried multi-word expression with one IRI, it is not CE. The default rewrite is the literal equivalent.

Not every figurative-looking string is a live metaphor. English has already lexicalized many of them. OEWN lists crash as a market event and as a collision; those are two synsets, not a comparison in flight. A closed allowlist is what separates the two cases. Until that list exists, rewrite. The inventory procedure is next steps, item 8.

Three outcomes, and only three:

• Literalize. The intended meaning is written with the ordinary senses of the words. This is the default.
• Admit. The expression is on the allowlist: either a figurative OEWN synset pinned on @graph, or a frozen multi-word expression whose dictionary module maps the whole string to one IRI (OntoLex-Lemon). The parts are not parsed as a composition.
• Reject. The expression is a live metaphor, a stock simile, a Banned construction, or an unnamed placeholder. There is no CE form that keeps the decoration.

Metonymy is already rejected as inference-mediated reference (§9, trigger 29). Personification of an atmosphere or an absence (silence that stretches, a room that holds its breath) is the same failure: the subject cannot bear the predicate. Idioms without a module entry rewrite to the proposition they stand for (kick the bucket → die).

A meaning-bearing clash is still a clash. “Take arms against a sea of troubles” fails domain consistency on purpose. The linter may flag it; it must not silently rewrite it. The author either literalizes, or the sentence is not CE.

Construction families that are not CE

These are architectural habits, not word lists. Each one simulates depth, tension, or emotion without naming the event, the referent, or the consequence. The Banned guide’s Part 1 is the source catalogue; the families below are the CE restatement. The rewrite is always: name what happened.

Metaphor clusters that start as reject

These clusters are the Banned guide’s Part 2, treated as the seed of the reject set. A later inventory pass may promote a member that OEWN already stores as its own synset. Until then, they rewrite to a literal predication in the tenor’s domain.

Underspecified register words (delve, figurative tapestry / landscape, pivotal, nestled, stands as a testament) fail the same test: they assert importance or motion without a measurable predicate. Replace them with the specific verb or cut them. They are not banned for being “AI words”.

• Correct The chef was angry at the waiter. (Attitude and object named.)
• Correct The stock market fell 2 percent. (If crash is later allowlisted as the market-event synset, that form is also CE once pinned.)
• Incorrect Something shifts behind his eyes. (Unnamed event. Write the belief or the expression.)
• Incorrect The silence stretches between them. (Silence is not an agent. Write who does not speak, and for how long.)
• Incorrect She was pulled toward him. (Gravitational vehicle. Write that she approached him, or that she wanted to.)
• Incorrect The news hit like a punch to the gut. (Stock impact simile. Name the emotion.)
• Incorrect He left, highlighting his frustration. (Narrator gloss. Write: He left because he was frustrated. Or cut the because-clause and leave the leaving.)
• Incorrect It is not a setback. It is a lesson. (Negative parallelism. Write the lesson, or write the setback.)

None of the above is machine-checked yet. The allowlist and the reject catalogue become data — a versioned JSON-LD module next to the lexicon — and then a CE-000n check. Until that module exists, authors and converting agents rewrite to the literal equivalent by default.