/**
 * Phrase dictionaries.
 *
 * These lists are DATA, not user interface: they are the vocabulary Trazum
 * looks for inside the prompts it optimises. Real prompts mix languages, so
 * the dictionaries deliberately cover English and Spanish at the same time and
 * a single prompt can trigger entries from both.
 *
 * Adding a language here means adding entries to these lists — it is unrelated
 * to the report language, which lives in `src/i18n/`.
 *
 * Every entry earns its place by preserving meaning. Anything whose removal
 * could change what the prompt asks for belongs in the `aggressive` level, or
 * does not belong here at all.
 */
/**
 * Languages the trimming dictionaries below actually cover.
 *
 * Exported because the report has to be able to say so. Until this existed,
 * a French or German prompt came back with `No rule found anything to trim` —
 * which reads as "your prompt is already efficient" and meant "I do not speak
 * your language". Same class of mistake as the one `--reorder` had: the tool
 * knew something it was not telling the reader.
 *
 * `phrases.test.js` asserts this list matches the languages the dictionaries are
 * actually grouped under, so adding a dictionary without listing it — or listing
 * one without the entries — fails rather than misleads.
 */
export declare const PHRASE_LANGUAGES: readonly string[];
/** Long phrases and their shorter equivalent. */
export declare const VERBOSE_PHRASES: ReadonlyArray<readonly [string, string]>;
/** Courtesy: adds nothing for the model and costs tokens on every call. */
export declare const POLITENESS: readonly string[];
/** Filler and throat-clearing with no content. */
export declare const FILLER: readonly string[];
/** Intensifiers: they rarely change the task and almost always add tokens. */
/**
 * Words that add emphasis and nothing else, so dropping one cannot change what
 * the prompt asks for.
 *
 * **A quantifier is not an intensifier**, and the distinction is not visible in
 * a dictionary translated word by word. Spanish gets `muy` and deliberately not
 * `mucho`; the first draft of the other five languages did not keep that line
 * and shipped `muito`, `molto` and `heel`, each of which does both jobs:
 *
 *     Hai molto tempo per rispondere.   →   Hai tempo per rispondere.
 *
 * "You have much time" became "you have time". Caught by running the five
 * languages through the rules rather than by reading the list, and pinned by a
 * test that keeps those three words out.
 */
export declare const INTENSIFIERS: readonly string[];
/** Hedges that weaken the instruction without adding information. */
export declare const HEDGES: readonly string[];
/**
 * Self-verification instructions.
 *
 * On current models these trigger over-verification: the model already checks
 * its own work, and asking explicitly fires extra steps paid for in output
 * tokens. Aggressive level because some workflows do want explicit
 * verification.
 *
 * The full forms come first on purpose: removing only the core phrase would
 * leave a dangling fragment such as "Before answering." — worse than leaving
 * the sentence alone.
 *
 * That includes whatever introduces the instruction. "You MUST double-check
 * your answer before responding" reduced to "You must." for exactly this
 * reason: the phrase matched, the subject and modal in front of it did not,
 * and what survived was a sentence that says nothing. Anything that can open
 * one of these instructions belongs in the list, ahead of the bare form.
 */
export declare const SELF_CHECK: readonly string[];
/** Words shouted in capitals that cost fewer tokens in lowercase. */
export declare const SHOUTED_WORDS: readonly string[];
/** Emphasis prefixes at the start of a line. */
export declare const EMPHASIS_PREFIXES: readonly string[];
/**
 * Vocabulary suggesting the task needs a more capable model.
 * Multilingual for the same reason as the dictionaries above.
 */
export declare const COMPLEX_SIGNALS: readonly string[];
/** Vocabulary suggesting a cheaper model would do. */
export declare const SIMPLE_SIGNALS: readonly string[];
/**
 * Phrases that refer *backwards* to something earlier in the prompt, by language.
 *
 * These are what makes reordering unsafe. "Summarise the text above" is correct
 * where it sits and nonsense if moved in front of the text it points at, so a
 * block containing one of these must stay where it is — and so must everything
 * after it, because moving a later block past a pinned one changes their order
 * relative to each other.
 *
 * Deliberately generous. A false positive costs a saving that was available; a
 * false negative silently changes what the prompt asks for, which is the one
 * thing this project will not trade for tokens.
 *
 * **Grouped by language rather than kept as one flat list, because the flat list
 * hid a hole.** It held English and Spanish and was applied to every prompt, so
 * a French, German or Japanese author ran `--reorder` with no protection at all:
 * every refusal this module is built on silently did not apply to them, and
 * "Résumez le texte ci-dessus" was hoisted above the text and reported as a
 * saving. A list per language makes the coverage a thing you can look at, and
 * lets `UNCOVERED_SCRIPTS` below refuse what is still missing rather than
 * pretending.
 *
 * Every language is matched against every prompt. Language detection would be
 * one more thing to get wrong, and the cost of checking a French prompt for
 * German phrases is a missed saving, which is the direction this module errs in
 * anyway.
 */
export interface BackwardReferenceSet {
    /**
     * Whether a match must sit on a word boundary.
     *
     * True for anything written with spaces, so "aboveboard" does not pin a block.
     * False for Japanese and Chinese, which have no word boundaries at all — the
     * boundary test asks whether the neighbouring character is a letter, and in
     * 上記のテキスト it always is, so a boundary-matched CJK phrase would never
     * fire. A list that cannot match is worse than no list: it reads like cover.
     */
    wordBoundaries: boolean;
    phrases: readonly string[];
}
export declare const BACKWARD_REFERENCES_BY_LANGUAGE: Readonly<Record<string, BackwardReferenceSet>>;
/**
 * Every phrase, flattened. Kept for callers that only need the words.
 */
export declare const BACKWARD_REFERENCES: readonly string[];
/**
 * Scripts this module has no backward-reference phrases for.
 *
 * The point of naming them is that the alternative is what used to happen:
 * rearranging a Russian or Arabic prompt with nothing to stop it, and calling
 * the result a saving. `reorderForCache` refuses when it sees one of these,
 * which turns a silent hazard into a message somebody can act on — and into a
 * short list of pull requests, since adding a language is adding an array.
 *
 * Matched on the script, not the language: the question is not "which language
 * is this" but "is there any chance my phrase lists apply to it".
 */
export declare const UNCOVERED_SCRIPTS: ReadonlyArray<{
    name: string;
    pattern: RegExp;
}>;
/**
 * Phrases that mark a block as the *output contract* rather than as data.
 *
 * Needed because a fenced JSON block in a prompt is one of two completely
 * different things, and confusing them is the one way this analysis could do
 * harm. `Output format: {...}` describes what the model must return, and a
 * provider that accepts a response schema takes it as a request parameter
 * instead — those tokens leave the prompt on every call. `Input: {...}` inside a
 * few-shot example is *data the prompt needs*, and advising anyone to move it
 * would break their prompt.
 *
 * Nothing here guesses. A schema block with no such phrase before it is left
 * alone, because "probably an output format" is not good enough when being wrong
 * costs somebody a working prompt.
 *
 * Seven languages, the same set the rules cover. A prompt in a language absent
 * here raises nothing, which is a false negative and states itself as one — the
 * alternative is matching an English cue inside Japanese prose and calling the
 * result a saving.
 */
export declare const OUTPUT_CUES_BY_LANGUAGE: Readonly<Record<string, readonly string[]>>;
/** Every output cue, flattened. Order is not significant. */
export declare const OUTPUT_CUES: readonly string[];
//# sourceMappingURL=phrases.d.ts.map