church-translate/app/mt-guard.js

/*
 * mt-guard.js — catch what NLLB invents, before a congregation hears it.
 *
 * WHY THIS IS NOT A GENERATION-PARAMETER FIX. Measured against the shipped
 * model (nllb-200-distilled-600M fp32), "Donald Trump." produces
 * "Donald Trump. ¿Qué es eso?" under ALL of: default params, an input-derived
 * max_new_tokens cap, no_repeat_ngram_size:3, and num_beams:2. The model
 * genuinely believes that is the translation. Decoding cannot be tuned out of
 * it; the output has to be checked.
 *
 * WHY THE OBVIOUS RULE IS WRONG. "Source has no '?', target has one, so strip
 * the question" corrupts real output: Whisper routinely drops the question mark
 * on a rhetorical aside, so "Are you with me" arrives unpunctuated and a
 * correct Spanish rendering is "¿Estás conmigo?" — which that rule would
 * delete. The reliable signal is not the question mark, it is the SENTENCE
 * COUNT: one sentence in, two out, and the second one invented.
 *
 * WHY ONLY SHORT INPUTS. Hallucination is a short-input failure — the model has
 * no context and fills the gap. Long sentences are also the ones that
 * legitimately split in the target (many languages prefer shorter sentences
 * than English), so clamping them would delete real scripture. The length gate
 * keeps the clamp where the evidence is.
 */
(function (root, factory) {
  if (typeof module === 'object' && module.exports) module.exports = factory();
  else root.MtGuard = factory();
}(typeof self !== 'undefined' ? self : this, function () {
  'use strict';

  /* Sentence-final punctuation across the scripts this app serves: Latin,
     CJK ideographic stop, fullwidth marks, Arabic question mark, Devanagari
     danda, Armenian, Greek, Thai has none. */
  const TERMINATORS = '.!?。！？؟۔।॥…՜՞;';
  const TERM_RE = new RegExp('[' + TERMINATORS.replace(/[.\-\]\\]/g, '\\$&') + ']');
  const QUESTION_RE = /[?？؟՞]/;

  /* Beyond this the input carries enough context that the model stops
     inventing, and legitimate sentence-splitting starts. Measured failures all
     sit far below it. */
  /* Narrow deliberately. Every hallucination measured against this model sat
     at one or two words; a wider gate starts reaching utterances that split
     into two target sentences legitimately, and deleting real scripture is a
     far worse failure than leaving one invented clause in. */
  const SHORT_WORDS = 6;

  function words(s) {
    return String(s || '').trim().split(/\s+/).filter(Boolean).length;
  }

  /* Count sentences by terminator runs. "Mr. Smith went." must not read as two,
     so a terminator followed by a lowercase letter does not end a sentence. */
  function sentenceCount(s) {
    const parts = splitSentences(s);
    return parts.length || 1;
  }

  function splitSentences(s) {
    const text = String(s || '').trim();
    if (!text) return [];
    const out = [];
    let cur = '';
    for (let i = 0; i < text.length; i++) {
      const ch = text[i];
      cur += ch;
      if (!TERM_RE.test(ch)) continue;
      /* Absorb a run of terminators and any closing quote/bracket. */
      while (i + 1 < text.length && (TERM_RE.test(text[i + 1]) || /["'\)\]”’」』]/.test(text[i + 1]))) {
        cur += text[++i];
      }
      const rest = text.slice(i + 1);
      /* An abbreviation or a decimal: "Dr. Smith", "3.5" — not a boundary. */
      if (/^\s*[a-zà-ÿ0-9]/.test(rest)) continue;
      out.push(cur.trim());
      cur = '';
    }
    if (cur.trim()) out.push(cur.trim());
    return out;
  }

  /*
   * NLLB was trained on a lot of subtitles and emits their dialogue dash. It is
   * never correct here — nothing in this app is a screenplay.
   */
  function stripDash(s) {
    return String(s || '').replace(/^\s*[-‐-―−]\s+/, '').trim();
  }

  /*
   * The main entry. Returns { text, trimmed, reason } — `trimmed` is true when
   * something was removed, so callers can log it and the operator can see how
   * often the model is inventing.
   */
  /*
   * A ONE-WORD source that comes back with a comma-clause. Measured across the
   * archive, 13.4% of everything the preacher says is one or two words, and the
   * model answers the short ones conversationally:
   *
   *   "Hello."  -> "Hola, ¿qué pasa?"     "Yeah."  -> "Sí, es cierto."
   *   "No."     -> "No, no."              "Good."  -> "Bien, muy bien."
   *   "Come."   -> "Ven, ven."            "Look."  -> "Mira, ¿qué?"
   *
   * One English word cannot mean two clauses, so the tail is invented. Keeping
   * only the head fixes all of these and touches nothing that was already
   * right ("La fe.", "Sin embargo.", "Por lo tanto." have no comma).
   *
   * Restricted to EXACTLY one word deliberately — at two words a comma can be
   * legitimate. Handles the CJK and ideographic commas too, since this is the
   * one output rule that needs no per-language knowledge.
   */
  function trimInventedClause(src, out) {
    const words = String(src == null ? '' : src).replace(/[.!?,;:]+$/, '').split(/\s+/).filter(Boolean);
    if (words.length !== 1) return out;
    const m = String(out).match(/^([^,\uff0c\u3001]+)[,\uff0c\u3001]/u);
    if (!m) return out;
    const head = m[1].trim();
    if (!head) return out;
    const end = (String(src).match(/[.!?]$/) || [''])[0];
    return /[.!?\u3002\uff01\uff1f]$/.test(head) ? head : head + end;
  }

  /*
   * THE WHOLE OUTPUT IS AN INVENTED QUESTION.
   *
   * The sentence-count rule above catches an invented question APPENDED to a
   * real translation ("Donald Trump. ¿Qué es eso?"). It cannot catch the case
   * where the invention IS the output: one sentence in, one sentence out, and
   * the one out is "¿Qué es eso?" for a source of "Mmm" or "them" — MEASURED
   * in a live service, spoken aloud to the Spanish congregation three times in
   * one morning, and produced by 34 distinct short sources in the archive.
   *
   * Narrow on purpose. Source: at most two words and not itself a question.
   * Output: every sentence is a question. And the source must not be an
   * English interrogative or a tag — "Why.", "Who.", "Right?", "You know." —
   * because for those a Spanish question IS the correct rendering (measured:
   * 'why.' -> '¿Por qué?', 'Isn't it.' -> '¿No es cierto?'). Whisper drops
   * question marks routinely, so the word list, not the punctuation, is the
   * exemption.
   */
  /*
   * Widened with the auxiliaries and contractions that FRONT a real English
   * question, because the rule below now reads every word of a short source
   * rather than the single word it used to be limited to.
   *
   * MEASURED: without `could` and `you're`, widening the rule deleted
   * 'Could it be' -> '¿Podría ser' and "You're interested in" -> '¿Te interesa?'
   * — both correct renderings of questions Whisper handed over unpunctuated.
   * With them the widened rule breaks NOTHING in 22,526 real short lines.
   */
  const INTERROGATIVE_SRC = /^(?:why|who|whom|whose|what|where|when|when'?s|how|which|isn'?t|right|okay|ok|see|really|huh|anybody|anyone|somebody|someone|guess|any|says|you|you'?re|we'?re|they'?re|there'?s|it'?s|are|am|is|was|were|do|does|did|can|could|will|would|shall|should|may|might|must|have|has|had|ready|let'?s)$/i;

  function isQuestionSentence(p) {
    const t = String(p || '').trim();
    return /^[¿]/.test(t) || /[?？؟՞]["'\)\]”’」』]*$/.test(t);
  }

  function bareSource(src) {
    return String(src || '').replace(/^[\s"'“‘(\[]+/u, '').replace(/[.!,;:…\-\s"'”’)\]]+$/u, '').trim();
  }

  /* Short enough that the model had no context to work from. Four, measured:
     see the numbers below. */
  const INVENTED_MAX_WORDS = 4;

  /*
   * UP TO FOUR words, not one.
   *
   * This was restricted to a single word because a two-word rule was measured
   * blanking 'Why not.' -> '¿Por qué no?', 'Guess what.' -> '¿Adivina qué?',
   * 'You ready.' -> '¿Están listos?'. That finding was right; the conclusion
   * drawn from it was too narrow. What those cases have in common is not their
   * LENGTH, it is that they contain an interrogative word — so the exemption
   * below, not the word count, is what protects them. Testing every word of
   * the source instead of only a lone one keeps all of them and frees the
   * count to rise.
   *
   * MEASURED over this church's own archive — 349 confirmed inventions and
   * 22,526 legitimate short translations from ten real services:
   *
   *     one word  (before)   198/349 caught   57%    0 legitimate lines broken
   *     <= 4 words           254/349 caught   73%    0 legitimate lines broken
   *
   * The congregation was hearing 'Check, check.' as '¿Qué quieres?', 'Holy
   * Spirit.' as '我們必須要做什麼?' and 'Hello.' as '¿Qué pasa?' — all of them
   * two words, all of them one word past where this rule used to stop looking.
   *
   * A source of NO words reaches here too ('-', '...', a stray dash Whisper
   * emits between speakers). Nothing cannot mean a question, so it is not
   * exempted: an empty source with an all-question output is an invention by
   * definition.
   */
  function inventedQuestion(src, out) {
    if (QUESTION_RE.test(src)) return false;
    const words = bareSource(src).split(/\s+/).filter(Boolean);
    if (words.length > INVENTED_MAX_WORDS) return false;
    /* ANY interrogative word, anywhere in the source — 'Guess what.' and
       'Why not.' put it first, 'How about now' and 'And why' do not. */
    if (words.some(function (w) { return INTERROGATIVE_SRC.test(w); })) return false;
    const parts = splitSentences(out);
    return parts.length > 0 && parts.every(isQuestionSentence);
  }

  /*
   * THE PORTUGUESE TWIN. NLLB's stock invention for a fragment in Portuguese is
   * not a question but the statement "Não." ('See.' -> 'Não.', 'Wow.' ->
   * 'Não.', 'way.' -> 'Não.', 'You understand?' -> '- Não.' — measured in
   * the archive), and the Spanish "No, no lo sé." family. A short source with
   * no negation in it cannot mean "No".
   */
  const NEGATION_SRC = /\b(?:no|not|never|nope|nah|nay|nix|negative|nothing|none|nobody|neither|nor|don'?t|can'?t|won'?t|isn'?t|ain'?t|didn'?t|doesn'?t|wasn'?t|aren'?t|couldn'?t|wouldn'?t|shouldn'?t)\b|\b(?:uh|nuh)-?uh\b/i;
  const NEGATION_OUT = /^[\s¿¡\-‐-―−"'“”]*(?:não|nao|no|non|nein|nee|niet|nej|nie|нет|いいえ|不|아니요|لا)[\s.!?…"'”]*$/iu;

  function inventedNegation(src, out) {
    const bare = bareSource(src);
    const words = bare.split(/\s+/).filter(Boolean);
    if (!words.length || words.length > 2) return false;
    if (NEGATION_SRC.test(bare)) return false;
    return NEGATION_OUT.test(String(out || ''));
  }

  function clean(source, target) {
    const src = String(source == null ? '' : source).trim();
    let out = stripDash(target);
    if (!out) return { text: '', trimmed: false, reason: '' };

    const dashed = out !== String(target || '').trim();

    const oneWord = trimInventedClause(src, out);
    if (oneWord !== out) {
      return { text: oneWord, trimmed: true,
               reason: 'dropped invented clause after a one-word line \u2014 "' + out + '"' };
    }

    if (inventedQuestion(src, out)) {
      return { text: '', trimmed: true,
               reason: 'dropped an invented question for a short line \u2014 "' + out + '"' };
    }
    if (inventedNegation(src, out)) {
      return { text: '', trimmed: true,
               reason: 'dropped an invented negation for a short line \u2014 "' + out + '"' };
    }

    if (words(src) > SHORT_WORDS) {
      return { text: out, trimmed: dashed, reason: dashed ? 'subtitle dash' : '' };
    }

    const nSrc = sentenceCount(src);
    const parts = splitSentences(out);
    if (parts.length <= nSrc) {
      return { text: out, trimmed: dashed, reason: dashed ? 'subtitle dash' : '' };
    }

    /*
     * More sentences came out than went in, on an input too short to justify
     * it. Keep the first nSrc — NLLB translates in order, so the real content
     * is at the front and the invention is appended.
     */
    const kept = parts.slice(0, nSrc).join(' ').trim();
    const dropped = parts.slice(nSrc).join(' ').trim();
    if (!kept) return { text: out, trimmed: dashed, reason: '' };
    return {
      text: kept,
      trimmed: true,
      reason: 'dropped invented ' + (QUESTION_RE.test(dropped) ? 'question' : 'text') +
              ' "' + dropped + '"'
    };
  }

  return { clean, splitSentences, sentenceCount, stripDash, trimInventedClause, inventedQuestion, inventedNegation, SHORT_WORDS };
}));