diff --git a/src/__tests__/helpers/ascii-boundary.ts b/src/__tests__/helpers/ascii-boundary.ts
new file mode 100644
index 000000000..48636a304
--- /dev/null
+++ b/src/__tests__/helpers/ascii-boundary.ts
@@ -0,0 +1,107 @@
+/**
+ * Find `\b` word boundaries that sit next to an accented letter in a regular
+ * expression's source.
+ *
+ * JavaScript defines `\b` over ASCII `\w`, so `santé\b` never matches before a
+ * space and `\bépisode` never matches after one; an alternative that starts or
+ * ends that way is dead. The check looks at what the source puts on each side
+ * of the boundary: a literal letter, a character class, or a group whose
+ * alternatives start (or end) with one. It is a text heuristic over the
+ * pattern source, not a regex engine: a boundary next to a backreference or a
+ * lookaround it does not unwrap is not judged.
+ */
+function isAccentedLetter(char: string | undefined): boolean {
+  return char !== undefined && char.charCodeAt(0) > 127 && /\p{L}/u.test(char);
+}
+
+function classEnd(source: string, open: number): number {
+  let i = open + 1;
+  while (i < source.length && !(source[i] === "]" && source[i - 1] !== "\\")) {
+    i += 1;
+  }
+  return i;
+}
+
+function classHasAccent(source: string, open: number): boolean {
+  return [...source.slice(open, classEnd(source, open))].some(isAccentedLetter);
+}
+
+/** Top-level alternatives of the group opening at `open`, and its close. */
+function alternatives(source: string, open: number): string[] {
+  const prefix = /^\(\?(?:<[=!]|[:=!])/.exec(source.slice(open))?.[0] ?? "(";
+  const alts: string[] = [];
+  let depth = 0;
+  let start = open + prefix.length;
+  for (let i = open; i < source.length; i += 1) {
+    const c = source[i];
+    if (c === "\\") {
+      i += 1;
+    } else if (c === "[") {
+      i = classEnd(source, i);
+    } else if (c === "(") {
+      depth += 1;
+    } else if (c === ")") {
+      depth -= 1;
+      if (depth === 0) {
+        alts.push(source.slice(start, i));
+        return alts;
+      }
+    } else if (c === "|" && depth === 1) {
+      alts.push(source.slice(start, i));
+      start = i + 1;
+    }
+  }
+  return alts;
+}
+
+function opensWithAccent(text: string): boolean {
+  if (text.startsWith("[")) return classHasAccent(text, 0);
+  if (text.startsWith("(")) return alternatives(text, 0).some(opensWithAccent);
+  return isAccentedLetter(text[0]);
+}
+
+function closesWithAccent(text: string): boolean {
+  const trimmed = text.replace(/(?:[?*+]|\{\d+(?:,\d*)?\})+\??$/, "");
+  const last = trimmed[trimmed.length - 1];
+  if (last === "]") {
+    let open = trimmed.length - 2;
+    while (
+      open >= 0 &&
+      !(trimmed[open] === "[" && trimmed[open - 1] !== "\\")
+    ) {
+      open -= 1;
+    }
+    return classHasAccent(trimmed, open);
+  }
+  if (last === ")") {
+    let depth = 0;
+    for (let i = trimmed.length - 1; i >= 0; i -= 1) {
+      if (trimmed[i - 1] === "\\") continue;
+      if (trimmed[i] === ")") depth += 1;
+      else if (trimmed[i] === "(" && --depth === 0) {
+        return alternatives(trimmed, i).some(closesWithAccent);
+      }
+    }
+    return false;
+  }
+  return isAccentedLetter(last);
+}
+
+/** Offsets of every `\b` in `source` that touches an accented letter. */
+export function accentedBoundaries(source: string): number[] {
+  const hits: number[] = [];
+  for (
+    let i = source.indexOf("\\b");
+    i >= 0;
+    i = source.indexOf("\\b", i + 2)
+  ) {
+    if (source[i - 1] === "\\") continue;
+    if (
+      opensWithAccent(source.slice(i + 2)) ||
+      closesWithAccent(source.slice(0, i))
+    ) {
+      hits.push(i);
+    }
+  }
+  return hits;
+}
diff --git a/src/lib/ai/coach/population-norms/__tests__/liveness.test.ts b/src/lib/ai/coach/population-norms/__tests__/liveness.test.ts
new file mode 100644
index 000000000..145e4e072
--- /dev/null
+++ b/src/lib/ai/coach/population-norms/__tests__/liveness.test.ts
@@ -0,0 +1,121 @@
+/**
+ * Every population-norm pattern of every locale has to match at least one
+ * realistic reference sentence below. A pattern that never fires exempts
+ * nothing and fails silently; French "en bonne santé" and Polish "zdrowi
+ * dorośli" were dead that way, the one behind an ASCII `\b` after "é", the
+ * other spelled for "dorosli" without the "ś".
+ */
+import { describe, expect, it } from "vitest";
+
+import { accentedBoundaries } from "@/__tests__/helpers/ascii-boundary";
+import { locales, type Locale } from "@/lib/i18n/config";
+
+import { POPULATION_NORM_PATTERNS } from "..";
+
+const SENTENCES: Record<Locale, readonly string[]> = {
+  en: [
+    "The normal range is 60 to 100.",
+    "Healthy adults sleep 7 to 9 hours.",
+    "The general population averages 120.",
+    "Guidelines recommend 150 minutes a week.",
+    "Adults generally need 7 hours of sleep.",
+    "The reference range is 4 to 6.",
+    "The recommended amount is 2 litres.",
+  ],
+  de: [
+    "Der Normalbereich liegt bei 60 bis 100.",
+    "Die normale Herzfrequenz liegt bei 60.",
+    "Gesunde Erwachsene schlafen 8 Stunden.",
+    "Ärzte empfehlen 150 Minuten pro Woche.",
+    "Erwachsene brauchen in der Regel 8 Stunden.",
+    "Normale Werte liegen bei 4 bis 6.",
+    "Der Puls liegt normalerweise zwischen 60 und 100.",
+  ],
+  fr: [
+    "La plage normale va de 60 à 100.",
+    "Les adultes en bonne santé dorment 8 heures.",
+    "La plupart des adultes dorment 8 heures.",
+    "Les médecins recommandent 150 minutes.",
+    "Les adultes ont besoin de 8 heures.",
+    "Elle se situe généralement entre 60 et 100.",
+  ],
+  es: [
+    "El rango normal va de 60 a 100.",
+    "Los adultos sanos duermen 8 horas.",
+    "La mayoría de los adultos duerme 8 horas.",
+    "Los médicos recomiendan 150 minutos.",
+    "Los adultos generalmente necesitan 8 horas.",
+    "Se sitúa normalmente entre 60 y 100.",
+  ],
+  it: [
+    "L'intervallo normale va da 60 a 100.",
+    "Gli adulti sani dormono 8 ore.",
+    "La maggior parte degli adulti dorme 8 ore.",
+    "I medici raccomandano 150 minuti.",
+    "Gli adulti di solito hanno bisogno di 8 ore.",
+    "Si colloca di solito tra 60 e 100.",
+  ],
+  pl: [
+    "Prawidłowy zakres wynosi 60 do 100.",
+    "Zdrowi dorośli śpią 8 godzin.",
+    "Większość dorosłych śpi 8 godzin.",
+    "Lekarze zalecają 150 minut.",
+    "Dorośli zwykle potrzebują 8 godzin.",
+    "Zakres referencyjny to 4 do 6.",
+    "Wynosi zwykle od 60 do 100.",
+  ],
+  nl: [
+    "Het normale bereik is 60 tot 100.",
+    "Een normale hartslag ligt rond 60.",
+    "De meeste volwassenen slapen 8 uur.",
+    "De algemene bevolking slaapt 7 uur.",
+    "Artsen adviseren 150 minuten per week.",
+    "Volwassenen hebben doorgaans nodig: 7 tot 9 uur slaap.",
+    "Het referentiebereik is 4 tot 6.",
+    "Je hartslag ligt meestal tussen 60 en 100.",
+  ],
+  ko: [
+    "정상 범위는 60~100이에요.",
+    "참고 범위는 4~6이에요.",
+    "건강한 성인은 8시간 자요.",
+    "일반 인구는 7시간 자요.",
+    "전문가들은 150분을 권장해요.",
+    "성인은 보통 8시간이 필요해요.",
+    "보통 60에서 100 사이예요.",
+  ],
+  "pt-BR": [
+    "A faixa normal vai de 60 a 100.",
+    "Os valores de referência vão de 4 a 6.",
+    "Adultos saudáveis dormem 8 horas.",
+    "A maioria dos adultos dorme 8 horas.",
+    "Os médicos recomendam 150 minutos por semana.",
+    "Os adultos precisam de 8 horas de sono.",
+    "A pressão arterial normal fica abaixo de 120/80.",
+  ],
+};
+
+describe("population-norm banks: no pattern is dead", () => {
+  it.each(locales)(
+    "every %s pattern matches one of its sentences",
+    (locale) => {
+      const dead = POPULATION_NORM_PATTERNS[locale]
+        .map((pattern, index): [number, RegExp] => [index, pattern])
+        .filter(
+          ([, pattern]) => !SENTENCES[locale].some((s) => pattern.test(s)),
+        )
+        .map(([index, pattern]) => `[${index}] /${pattern.source}/`);
+      expect(dead, `${locale} patterns no sentence matches`).toEqual([]);
+    },
+  );
+
+  it.each(locales)(
+    "no %s pattern puts an ASCII word boundary next to an accented letter",
+    (locale) => {
+      const offenders = POPULATION_NORM_PATTERNS[locale]
+        .map((pattern, index): [string, RegExp] => [`[${index}]`, pattern])
+        .filter(([, pattern]) => accentedBoundaries(pattern.source).length > 0)
+        .map(([name]) => name);
+      expect(offenders, `${locale} population-norm patterns`).toEqual([]);
+    },
+  );
+});
diff --git a/src/lib/ai/coach/population-norms/de.ts b/src/lib/ai/coach/population-norms/de.ts
index 3a4780b96..d90de39b0 100644
--- a/src/lib/ai/coach/population-norms/de.ts
+++ b/src/lib/ai/coach/population-norms/de.ts
@@ -4,7 +4,7 @@ export const patterns: readonly RegExp[] = [
   /\bdie\s+normale\b/i,
   /\bnormal(?:er|e)?\s+(?:ruhe)?(?:herzfrequenz|puls|bereich|wert)\b/i,
   /\b(?:gesunde|die\s+meisten|der\s+durchschnittliche?)\s+(?:erwachsene|menschen|bevölkerung)\b/i,
-  /\b(?:leitlinien|ärzte|expert\w*)\s+(?:sagen|empfehlen|raten|betrachten)\b/i,
+  /(?<![\w\u00C0-\u024F])(?:leitlinien|ärzte|expert\w*)\s+(?:sagen|empfehlen|raten|betrachten)\b/i,
   /\b(?:erwachsene|menschen)\s+(?:brauchen|benötigen)\s+(?:in\s+der\s+regel|generell|typischerweise|normalerweise)\b/i,
   /\b(?:allgemeine|normale|typische)\s+(?:referenz)?(?:bereich|spanne|werte)\b/i,
   /\bliegt\s+(?:normalerweise|üblicherweise|in\s+der\s+regel)\s+(?:bei|zwischen)\b/i,
diff --git a/src/lib/ai/coach/population-norms/fr.ts b/src/lib/ai/coach/population-norms/fr.ts
index b193b6048..a17e767b9 100644
--- a/src/lib/ai/coach/population-norms/fr.ts
+++ b/src/lib/ai/coach/population-norms/fr.ts
@@ -1,7 +1,7 @@
 /** Population-norm framing, fr. */
 export const patterns: readonly RegExp[] = [
   /\bla\s+(?:plage|fourchette)\s+normale\b/i,
-  /\b(?:les\s+)?(?:adultes|personnes)\s+en\s+bonne\s+santé\b/i,
+  /\b(?:les\s+)?(?:adultes|personnes)\s+en\s+bonne\s+santé(?![\w\u00C0-\u024F])/i,
   /\b(?:la\s+plupart\s+des|la\s+moyenne\s+des)\s+(?:adultes|gens|personnes)\b/i,
   /\b(?:les\s+)?(?:recommandations|médecins|experts?)\s+(?:disent|recommandent|conseillent|suggèrent)\b/i,
   /\b(?:les\s+)?adultes\s+(?:ont\s+généralement|ont\s+besoin|nécessitent)\b/i,
diff --git a/src/lib/ai/coach/population-norms/pl.ts b/src/lib/ai/coach/population-norms/pl.ts
index 0ca306e87..b16b62006 100644
--- a/src/lib/ai/coach/population-norms/pl.ts
+++ b/src/lib/ai/coach/population-norms/pl.ts
@@ -1,10 +1,10 @@
 /** Population-norm framing, pl. */
 export const patterns: readonly RegExp[] = [
-  /\b(?:normalny|prawidłowy)\s+(?:zakres|przedział)\b/i,
-  /\bzdrow(?:i|ych)\s+doros(?:li|łych)\b/i,
-  /\b(?:większość|średnia)\s+(?:doros(?:łych|li)|ludzi|osób)\b/i,
-  /\b(?:wytyczne|lekarze|eksperci)\s+(?:mówią|zalecają|radzą|sugerują)\b/i,
-  /\bdoros(?:li|łych)\s+(?:zazwyczaj|zwykle|na\s+ogół)\s+(?:potrzebują|wymagają)\b/i,
+  /\b(?:normalny|prawidłowy)\s+(?:zakres|przedział)(?![\w\u00C0-\u024F])/i,
+  /\bzdrow(?:i|ych)\s+doro(?:[sś]li|słych)\b/i,
+  /(?<![\w\u00C0-\u024F])(?:większość|średnia)\s+(?:doro(?:słych|[sś]li)|ludzi|osób)\b/i,
+  /\b(?:wytyczne|lekarze|eksperci)\s+(?:mówią|zalecają|radzą|sugerują)(?![\w\u00C0-\u024F])/i,
+  /\bdoro(?:[sś]li|słych)\s+(?:zazwyczaj|zwykle|na\s+ogół)\s+(?:potrzebują|wymagają)(?![\w\u00C0-\u024F])/i,
   /\b(?:zakres|przedział)\s+(?:referencyjny|normy|prawidłowy|zalecany)\b/i,
   /\b(?:wynosi|mieści\s+się)\s+(?:zazwyczaj|zwykle|na\s+ogół)\s+(?:od|między|około)\b/i,
 ];
diff --git a/src/lib/ai/safety/__tests__/locale-bank-liveness.test.ts b/src/lib/ai/safety/__tests__/locale-bank-liveness.test.ts
new file mode 100644
index 000000000..9aefc11ae
--- /dev/null
+++ b/src/lib/ai/safety/__tests__/locale-bank-liveness.test.ts
@@ -0,0 +1,402 @@
+/**
+ * Every pattern of every locale's outbound-screen bank has to match at least
+ * one realistic sentence below.
+ *
+ * A pattern that can never fire is a contract the screen only appears to
+ * enforce, and nothing else notices: the bank stays green because the other
+ * patterns carry the tests. JavaScript's `\b` is an ASCII boundary, so next to
+ * an accented letter it never matches: French "à cause de", "causé" and
+ * "continuez à", the French and Italian "probabilité" / "probabilità" and the
+ * Dutch "één procent" all walked past their patterns, and a `\w*` stem cannot
+ * run through "ł", so Polish "doprowadziło do" did too. A new pattern needs a
+ * sentence here; a sentence that stops matching after an edit points at the
+ * pattern it broke.
+ */
+import { describe, expect, it } from "vitest";
+
+import { accentedBoundaries } from "@/__tests__/helpers/ascii-boundary";
+import { locales, type Locale } from "@/lib/i18n/config";
+
+import { bank as en } from "../outbound-screen/locales/en";
+import { bank as de } from "../outbound-screen/locales/de";
+import { bank as fr } from "../outbound-screen/locales/fr";
+import { bank as es } from "../outbound-screen/locales/es";
+import { bank as it_ } from "../outbound-screen/locales/it";
+import { bank as pl } from "../outbound-screen/locales/pl";
+import { bank as nl } from "../outbound-screen/locales/nl";
+import { bank as ko } from "../outbound-screen/locales/ko";
+import { bank as ptBR } from "../outbound-screen/locales/pt-br";
+import type { LocaleBank } from "../outbound-screen/shared";
+
+const BANKS: Record<Locale, LocaleBank> = {
+  en,
+  de,
+  fr,
+  es,
+  it: it_,
+  pl,
+  nl,
+  ko,
+  "pt-BR": ptBR,
+};
+
+const SENTENCES: Record<Locale, readonly string[]> = {
+  en: [
+    "Increase your metformin to 2000 mg daily.",
+    "You might want to consider increasing your insulin to 20 units.",
+    "Take 2000 mg daily.",
+    "Skip your evening dose.",
+    "Stop taking metformin.",
+    "I'd recommend skipping your evening dose.",
+    "Take an extra ramipril.",
+    "Stop your blood thinner before the run.",
+    "Leave your metformin out tomorrow.",
+    "I'd recommend stopping your blood thinner.",
+    "Step up to 2.4 mg next week.",
+    "Lower it to 5 mg.",
+    "You should try 5 mg.",
+    "Skip your dose for two weeks to see.",
+    "That is a 12% risk.",
+    "Your risk is about 12%.",
+    "Your risk is roughly twelve percent.",
+    "Twelve percent is your risk.",
+    "Your 10-year cardiovascular risk is about 12%.",
+    "About 12% is your 10-year cardiovascular risk.",
+    "Your ASCVD score comes out at 14%.",
+    "A 14% figure is what ASCVD gives.",
+    "SCORE2 would put you in a higher bracket.",
+    "Your SCORE2 lands in the high-risk band.",
+    "Your 10-year cardiovascular risk is elevated.",
+    "Your weight fell because you slept more.",
+    "It was caused by stress.",
+    "That is due to salt.",
+    "It led to a dip.",
+    "It resulted in a dip.",
+    "It is the result of travel.",
+    "Sleep is the culprit.",
+    "It was driven by stress.",
+    "Thanks to the walks, it fell.",
+    "Owing to the heat, it rose.",
+    "Salt is responsible for it.",
+    "Keep taking your 7.5 mg as prescribed.",
+  ],
+  de: [
+    "Erhöhe dein Metformin auf 2000 mg.",
+    "Nimm 2 Tabletten am Abend.",
+    "Pausiere dein Insulin.",
+    "Setz das Metformin ab.",
+    "Lass die Abenddosis weg.",
+    "Hör auf, Metformin zu nehmen.",
+    "Du solltest das Metformin absetzen.",
+    "Nimm abends eine Tablette mehr.",
+    "Hör mit dem Insulin auf.",
+    "Verzichte auf das Insulin.",
+    "Erhöhe auf 7,5 mg.",
+    "Reduziere auf 5 mg.",
+    "Die nächste Stufe sind 10 mg.",
+    "Es ist sinnvoll, auf 7,5 mg zu erhöhen.",
+    "Ich empfehle 5 mg.",
+    "Halbiere die Dosis für zwei Wochen.",
+    "Dein Risiko liegt bei etwa 14%.",
+    "Das sind 14% Risiko.",
+    "Dein Risiko liegt bei vierzehn Prozent.",
+    "Vierzehn Prozent Risiko bleiben.",
+    "Dein 10-Jahres-Risiko liegt bei 14%.",
+    "14% beträgt dein 10-Jahres-Risiko.",
+    "Der SCORE2 ergibt 14%.",
+    "Mit 14% laut SCORE2.",
+    "SCORE2 stuft dich ein als hoch.",
+    "SCORE2 zeigt einen hohen Risikobereich.",
+    "Dein 10-Jahres-Risiko ist erhöht.",
+    "Das Gewicht sank, weil du mehr geschlafen hast.",
+    "Es sank wegen des Schlafs.",
+    "Das verursacht Stress.",
+    "Es stieg aufgrund der Hitze.",
+    "Das führte zu einem Anstieg.",
+    "Er wurde durch Stress verursacht.",
+    "Schuld ist das Salz.",
+    "Der Auslöser war Salz.",
+    "Salz ist verantwortlich für den Anstieg.",
+    "Nimm dein Metformin weiter wie verordnet.",
+  ],
+  fr: [
+    "Augmentez votre metformine à 2000 mg.",
+    "Prenez 2 comprimés le soir.",
+    "Sautez votre dose du soir.",
+    "Arrêtez de prendre la metformine.",
+    "Je recommande de suspendre votre insuline.",
+    "Prenez un comprimé en plus.",
+    "Augmentez à 2,4 mg.",
+    "Réduisez à 5 mg.",
+    "Je recommande 5 mg.",
+    "La prochaine dose sera de 10 mg.",
+    "Votre risque est d'environ 12%.",
+    "Il y a 12% de risque.",
+    "Votre risque cardiovasculaire à 10 ans.",
+    "Votre risque est de douze pour cent.",
+    "Douze pour cent de risque.",
+    "SCORE2 donne 12%.",
+    "12% selon SCORE2.",
+    "SCORE2 vous place dans la tranche haute.",
+    "Votre poids a baissé parce que vous avez mieux dormi.",
+    "Il a monté à cause de la chaleur.",
+    "Le sel cause cela.",
+    "Il a monté en raison de la chaleur.",
+    "Cela entraîne une hausse.",
+    "Le sommeil a conduit à une baisse.",
+    "C'est dû à la chaleur.",
+    "Le sel est responsable de la hausse.",
+    "Il a baissé grâce à la marche.",
+    "Le stress provoque une hausse.",
+    "Continuez à prendre votre metformine.",
+  ],
+  es: [
+    "Aumente su metformina a 2000 mg.",
+    "Tome 2 comprimidos por la noche.",
+    "Omita la dosis de la noche.",
+    "Deje de tomar la metformina.",
+    "Tome una pastilla de más.",
+    "Deje el anticoagulante.",
+    "Aumente a 2,4 mg.",
+    "Reduzca a 5 mg.",
+    "Recomiendo 5 mg.",
+    "La próxima dosis será de 10 mg.",
+    "Su riesgo es del 12%.",
+    "Hay un 12% de riesgo.",
+    "Su riesgo cardiovascular a 10 años.",
+    "Su riesgo es de doce por ciento.",
+    "Doce por ciento de riesgo.",
+    "SCORE2 da 12%.",
+    "12% según SCORE2.",
+    "SCORE2 lo coloca en la franja alta.",
+    "Su peso bajó porque durmió más.",
+    "Subió debido a la sal.",
+    "Subió a causa de la sal.",
+    "La sal causa eso.",
+    "La sal provoca una subida.",
+    "Eso llevó a una bajada.",
+    "La sal es responsable de la subida.",
+    "Bajó gracias a la caminata.",
+    "Fue impulsado por el estrés.",
+    "Siga tomando su metformina.",
+  ],
+  it: [
+    "Aumenti la metformina a 2000 mg.",
+    "Prenda 2 compresse la sera.",
+    "Salti la dose serale.",
+    "Smetta di prendere la metformina.",
+    "Prenda una compressa in più.",
+    "Aumenti a 2,4 mg.",
+    "Riduca a 5 mg.",
+    "Raccomando 5 mg.",
+    "La prossima dose sarà di 10 mg.",
+    "Il suo rischio è del 12%.",
+    "C'è il 12% di rischio.",
+    "Il suo rischio cardiovascolare a 10 anni.",
+    "Il rischio è del dodici per cento.",
+    "Dodici per cento di rischio.",
+    "SCORE2 dà 12%.",
+    "12% secondo SCORE2.",
+    "SCORE2 la colloca nella fascia alta.",
+    "Il suo peso è calato perché ha dormito di più.",
+    "È salito a causa di troppo sale.",
+    "Il sale causa questo.",
+    "Il sale provoca un aumento.",
+    "Il sonno ha portato a un calo.",
+    "È dovuto a stress.",
+    "Il sale è responsabile di un aumento.",
+    "È calato grazie a una camminata.",
+    "È stato guidato da stress.",
+    "Continui a prendere la metformina.",
+  ],
+  pl: [
+    "Zwiększ metforminę do 2000 mg.",
+    "Proszę przyjąć 2 tabletki wieczorem.",
+    "Pomiń wieczorną dawkę.",
+    "Przestań brać metforminę.",
+    "Weź dodatkową tabletkę.",
+    "Zwiększ do 2,4 mg.",
+    "Zmniejsz do 5 mg.",
+    "Zalecam 5 mg.",
+    "Następna dawka to 10 mg.",
+    "Ryzyko wynosi 12%.",
+    "To 12% ryzyka.",
+    "Ryzyko zawału w ciągu 10 lat.",
+    "Ryzyko wynosi dwanaście procent.",
+    "Dwanaście procent ryzyka.",
+    "SCORE2 daje 12%.",
+    "12% według SCORE2.",
+    "SCORE2 umieszcza cię w grupie wysokiego ryzyka.",
+    "Waga spadła, ponieważ spał Pan więcej.",
+    "Wzrosło z powodu soli.",
+    "Sól powoduje wzrost.",
+    "Stres spowodował spadek.",
+    "Przyczyną jest sól.",
+    "To doprowadziło do spadku.",
+    "Wzrosło wskutek stresu.",
+    "Spadło dzięki spacerom.",
+    "Sól jest odpowiedzialna za wzrost.",
+    "Kontynuuj przyjmowanie leku.",
+  ],
+  nl: [
+    "Verhoog je metformine naar 2000 mg per dag.",
+    "Bouw je dosis af naar 5 mg.",
+    "Zet je insuline op 20 eenheden.",
+    "Het is beter om je metformine naar 2000 mg te verhogen.",
+    "Neem 's avonds 2 tabletten.",
+    "Halveer je insulinedosis.",
+    "Sla de avonddosis vandaag over.",
+    "Stop met het innemen van metformine.",
+    "Stop je bloedverdunner voor het hardlopen.",
+    "Laat de metformine morgen weg.",
+    "Je zou de metformine een week kunnen stoppen.",
+    "Neem een extra ramipril.",
+    "Verhoog naar 2,4 mg.",
+    "Verlaag naar 5 mg.",
+    "Ik adviseer 5 mg.",
+    "De volgende stap is 10 mg.",
+    "Halveer je dosis twee weken om te zien wat er gebeurt.",
+    "Je risico ligt op ongeveer 12%.",
+    "Dat is 12% risico.",
+    "Je risico is twaalf procent.",
+    "Twaalf procent risico.",
+    "Eén procent kans, meer niet.",
+    "Je 10-jaarsrisico is 12%.",
+    "12% is je 10-jaarsrisico.",
+    "SCORE2 geeft 12%.",
+    "12% volgens SCORE2.",
+    "SCORE2 plaatst je in de hoge groep.",
+    "SCORE2 zet je in de hoogrisicogroep.",
+    "Je 10-jaarsrisico is verhoogd.",
+    "Je gewicht daalde omdat je meer sliep.",
+    "Het daalde doordat je meer sliep.",
+    "Het steeg vanwege de hitte.",
+    "Zout veroorzaakt dat.",
+    "Het steeg als gevolg van stress.",
+    "Dat leidde tot een daling.",
+    "Dat komt door het zout.",
+    "Het daalde dankzij het wandelen.",
+    "Het is de schuld van het zout.",
+    "De oorzaak is zout.",
+    "Zout is verantwoordelijk voor de stijging.",
+    "Blijf je metformine nemen.",
+  ],
+  ko: [
+    "메트포르민을 2000mg으로 증량하세요.",
+    "저녁 인슐린 용량을 건너뛰세요.",
+    "메트포르민 복용을 중단하세요.",
+    "인슐린을 추가로 맞으세요.",
+    "2.4mg으로 늘려 보세요.",
+    "5mg으로 줄여 보세요.",
+    "권장 용량은 5mg이에요.",
+    "5mg을 복용하세요.",
+    "다음 단계는 10mg이에요.",
+    "약을 절반으로 줄여 2주간 해 보세요.",
+    "위험도는 약 12%예요.",
+    "12% 위험이 있어요.",
+    "10년 위험도는 12%예요.",
+    "12%가 10년 위험도예요.",
+    "SCORE2로 12%예요.",
+    "12%는 SCORE2 기준이에요.",
+    "SCORE2로 보면 고위험군에 해당해요.",
+    "SCORE2 기준 고위험군이에요.",
+    "10년 위험도는 높아요.",
+    "수면 때문에 체중이 줄었어요.",
+    "스트레스로 인해 올랐어요.",
+    "소금 탓에 올랐어요.",
+    "소금이 원인이에요.",
+    "소금이 상승을 유발해요.",
+    "소금이 상승을 초래해요.",
+    "소금이 상승을 야기해요.",
+    "산책 덕분에 내려갔어요.",
+    "스트레스가 영향을 줘서 올랐어요.",
+    "처방대로 복용하세요.",
+  ],
+  "pt-BR": [
+    "Aumente sua metformina para 2000 mg por dia.",
+    "Tome 2 comprimidos à noite.",
+    "Pule a dose da noite.",
+    "Pare de tomar a metformina.",
+    "Pare o anticoagulante antes da corrida.",
+    "Você poderia suspender a metformina por uma semana.",
+    "Tome um comprimido a mais.",
+    "Aumente a dose para 2,4 mg.",
+    "Reduza a dose para 5 mg.",
+    "Recomendo 5 mg.",
+    "A próxima dose será de 10 mg.",
+    "Dobre a dose por duas semanas para testar.",
+    "Seu risco é de cerca de 12%.",
+    "Há 12% de risco.",
+    "Seu risco é de doze por cento.",
+    "Doze por cento de risco.",
+    "Seu risco cardiovascular em 10 anos é de 12%.",
+    "12% é o seu risco cardiovascular em 10 anos.",
+    "O SCORE2 dá 12%.",
+    "12% pelo SCORE2.",
+    "O SCORE2 coloca você na faixa alta.",
+    "O SCORE2 indica alto risco.",
+    "Seu risco cardiovascular em 10 anos está elevado.",
+    "Seu peso caiu porque você dormiu mais.",
+    "Subiu por causa do sal.",
+    "Subiu devido ao sal.",
+    "O sal causa isso.",
+    "O sal provoca uma alta.",
+    "Isso levou a uma queda.",
+    "Isso resultou em uma queda.",
+    "Subiu em razão do calor.",
+    "Subiu em decorrência do estresse.",
+    "Caiu graças às caminhadas.",
+    "O sal é responsável pela alta.",
+    "Isso é culpa do sal.",
+    "Continue tomando sua metformina conforme prescrito.",
+  ],
+};
+
+function patternsOf(bank: LocaleBank): Array<[string, RegExp]> {
+  const groups: Record<string, readonly RegExp[]> = {
+    doseChangeClass: bank.doseChangeClass(bank.vocabulary(null)),
+    dose: bank.dose,
+    risk: bank.risk,
+    causal: bank.causal,
+    continuation: [bank.continuation],
+    changeStem: [bank.changeStem],
+  };
+  return Object.entries(groups).flatMap(([group, patterns]) =>
+    patterns.map((pattern, index): [string, RegExp] => [
+      `${group}[${index}]`,
+      pattern,
+    ]),
+  );
+}
+
+describe("outbound-screen banks: no pattern is dead", () => {
+  it("walks a real number of patterns", () => {
+    const total = locales.reduce(
+      (sum, locale) => sum + patternsOf(BANKS[locale]).length,
+      0,
+    );
+    expect(total).toBeGreaterThan(300);
+  });
+
+  it.each(locales)(
+    "every %s pattern matches one of its sentences",
+    (locale) => {
+      const dead = patternsOf(BANKS[locale])
+        .filter(
+          ([, pattern]) => !SENTENCES[locale].some((s) => pattern.test(s)),
+        )
+        .map(([name, pattern]) => `${name} /${pattern.source.slice(0, 120)}/`);
+      expect(dead, `${locale} patterns no sentence matches`).toEqual([]);
+    },
+  );
+
+  it.each(locales)(
+    "no %s pattern puts an ASCII word boundary next to an accented letter",
+    (locale) => {
+      const offenders = patternsOf(BANKS[locale])
+        .filter(([, pattern]) => accentedBoundaries(pattern.source).length > 0)
+        .map(([name]) => name);
+      expect(offenders, `${locale} outbound-screen patterns`).toEqual([]);
+    },
+  );
+});
diff --git a/src/lib/ai/safety/outbound-screen/locales/fr.ts b/src/lib/ai/safety/outbound-screen/locales/fr.ts
index f02bb4124..fd63cc39d 100644
--- a/src/lib/ai/safety/outbound-screen/locales/fr.ts
+++ b/src/lib/ai/safety/outbound-screen/locales/fr.ts
@@ -5,6 +5,7 @@ import {
   objectSlot,
   TARGET,
   ENGINE,
+  WORD_END,
   ELIDED,
   medicationExtras,
   type LocaleBank,
@@ -94,11 +95,17 @@ export const bank: LocaleBank = {
     /\b\d{1,3}\s*%\s+(?:de\s+)?(?:risque|probabilit[ée]|chance)/i,
     /\brisque\s+(?:cardiovasculaire|cardiaque|d'avc|de\s+mortalit[ée])\s+[àa]\s+(?:10|dix)\s+ans\b/i,
     // digit percent as a risk figure, either order (the noun need not be adjacent)
-    new RegExp(`\\b${RISK_NOUN_FR}\\b[^.?!]{0,40}\\d{1,3}\\s*%`, "i"),
-    new RegExp(`\\d{1,3}\\s*%[^.?!]{0,40}\\b${RISK_NOUN_FR}\\b`, "i"),
+    new RegExp(`\\b${RISK_NOUN_FR}${WORD_END}[^.?!]{0,40}\\d{1,3}\\s*%`, "i"),
+    new RegExp(`\\d{1,3}\\s*%[^.?!]{0,40}\\b${RISK_NOUN_FR}${WORD_END}`, "i"),
     // spelled-out percent as a risk figure, either order
-    new RegExp(`\\b${RISK_NOUN_FR}\\b[^.?!]{0,40}\\b${PCT_WORD_FR}\\b`, "i"),
-    new RegExp(`\\b${PCT_WORD_FR}\\b[^.?!]{0,40}\\b${RISK_NOUN_FR}\\b`, "i"),
+    new RegExp(
+      `\\b${RISK_NOUN_FR}${WORD_END}[^.?!]{0,40}\\b${PCT_WORD_FR}\\b`,
+      "i",
+    ),
+    new RegExp(
+      `\\b${PCT_WORD_FR}\\b[^.?!]{0,40}\\b${RISK_NOUN_FR}${WORD_END}`,
+      "i",
+    ),
     // named engine + qualifying number, either order
     new RegExp(`\\b${ENGINE}\\b[^.?!]{0,50}${QUAL_NUM_FR}`, "i"),
     new RegExp(`${QUAL_NUM_FR}[^.?!]{0,50}\\b${ENGINE}\\b`, "i"),
@@ -107,8 +114,8 @@ export const bank: LocaleBank = {
   ],
   causal: [
     /\bparce\s+qu\w*/i,
-    /\b[àa]\s+cause\s+de\b/i,
-    /\bcaus(?:e|es|ent|é|ée)\b/i,
+    /(?<![\w\u00C0-\u024F])[àa]\s+cause\s+de\b/i,
+    /\bcaus(?:e|es|ent|é|ée)(?![\w\u00C0-\u024F])/i,
     /\ben\s+raison\s+de\b/i,
     /\bentra[îi]n\w*/i,
     /\ba\s+conduit\s+[àa]/i,
@@ -118,7 +125,7 @@ export const bank: LocaleBank = {
     /\bprovoqu\w*/i,
   ],
   continuation:
-    /\b(?:continuez?\s+(?:à|de|le)|gardez|comme\s+prescrit|tel\s+que\s+prescrit)\b/i,
+    /\b(?:continuez?\s+(?:à|de|le)|gardez|comme\s+prescrit|tel\s+que\s+prescrit)(?![\w\u00C0-\u024F])/i,
   changeStem:
     /\b(?:augment|réduis|réduir|diminu|baiss|abaiss|doubl|arrêt|saut)\w*/i,
 };
diff --git a/src/lib/ai/safety/outbound-screen/locales/it.ts b/src/lib/ai/safety/outbound-screen/locales/it.ts
index 0a866e8ae..6ca4e3234 100644
--- a/src/lib/ai/safety/outbound-screen/locales/it.ts
+++ b/src/lib/ai/safety/outbound-screen/locales/it.ts
@@ -5,6 +5,7 @@ import {
   objectSlot,
   TARGET,
   ENGINE,
+  WORD_END,
   ELIDED,
   medicationExtras,
   type LocaleBank,
@@ -85,10 +86,16 @@ export const bank: LocaleBank = {
     /\brischio\s+(?:del?|dell'|[èe]\s+del?|di\s+circa)\s+(?:circa\s+|~)?\d{1,3}\s*%/i,
     /\b\d{1,3}\s*%\s+(?:di\s+)?(?:rischio|probabilit[àa])/i,
     /\brischio\s+(?:cardiovascolare|cardiaco|di\s+ictus|di\s+mortalit[àa])\s+a\s+(?:10|dieci)\s+anni\b/i,
-    new RegExp(`\\b${RISK_NOUN_IT}\\b[^.?!]{0,40}\\d{1,3}\\s*%`, "i"),
-    new RegExp(`\\d{1,3}\\s*%[^.?!]{0,40}\\b${RISK_NOUN_IT}\\b`, "i"),
-    new RegExp(`\\b${RISK_NOUN_IT}\\b[^.?!]{0,40}\\b${PCT_WORD_IT}\\b`, "i"),
-    new RegExp(`\\b${PCT_WORD_IT}\\b[^.?!]{0,40}\\b${RISK_NOUN_IT}\\b`, "i"),
+    new RegExp(`\\b${RISK_NOUN_IT}${WORD_END}[^.?!]{0,40}\\d{1,3}\\s*%`, "i"),
+    new RegExp(`\\d{1,3}\\s*%[^.?!]{0,40}\\b${RISK_NOUN_IT}${WORD_END}`, "i"),
+    new RegExp(
+      `\\b${RISK_NOUN_IT}${WORD_END}[^.?!]{0,40}\\b${PCT_WORD_IT}\\b`,
+      "i",
+    ),
+    new RegExp(
+      `\\b${PCT_WORD_IT}\\b[^.?!]{0,40}\\b${RISK_NOUN_IT}${WORD_END}`,
+      "i",
+    ),
     new RegExp(`\\b${ENGINE}\\b[^.?!]{0,50}${QUAL_NUM_IT}`, "i"),
     new RegExp(`${QUAL_NUM_IT}[^.?!]{0,50}\\b${ENGINE}\\b`, "i"),
     new RegExp(`\\b${ENGINE}\\b[^.?!]{0,60}${RESULT_VERB_IT}`, "i"),
diff --git a/src/lib/ai/safety/outbound-screen/locales/nl.ts b/src/lib/ai/safety/outbound-screen/locales/nl.ts
index 4cbac6f3a..8420ad6ec 100644
--- a/src/lib/ai/safety/outbound-screen/locales/nl.ts
+++ b/src/lib/ai/safety/outbound-screen/locales/nl.ts
@@ -7,6 +7,7 @@ import {
   TARGET,
   LETTER,
   ENGINE,
+  WORD_START,
   medicationExtras,
   type LocaleBank,
 } from "../shared";
@@ -140,8 +141,14 @@ export const bank: LocaleBank = {
       "i",
     ),
     new RegExp(`\\b\\d{1,3}\\s*%\\s+${RISK_NOUN_NL}\\b`, "i"),
-    new RegExp(`\\b${RISK_NOUN_NL}\\b[^.?!]{0,40}\\b${PCT_WORD_NL}\\b`, "i"),
-    new RegExp(`\\b${PCT_WORD_NL}\\b[^.?!]{0,40}\\b${RISK_NOUN_NL}\\b`, "i"),
+    new RegExp(
+      `\\b${RISK_NOUN_NL}\\b[^.?!]{0,40}${WORD_START}${PCT_WORD_NL}\\b`,
+      "i",
+    ),
+    new RegExp(
+      `${WORD_START}${PCT_WORD_NL}\\b[^.?!]{0,40}\\b${RISK_NOUN_NL}\\b`,
+      "i",
+    ),
     new RegExp(`\\b${HORIZON_NL}\\b[^.?!]{0,40}${QUAL_NUM_NL}`, "i"),
     new RegExp(`${QUAL_NUM_NL}[^.?!]{0,40}\\b${HORIZON_NL}\\b`, "i"),
     new RegExp(`\\b${ENGINE}\\b[^.?!]{0,50}${QUAL_NUM_NL}`, "i"),
diff --git a/src/lib/ai/safety/outbound-screen/locales/pl.ts b/src/lib/ai/safety/outbound-screen/locales/pl.ts
index 5a0f9b2dc..50163363c 100644
--- a/src/lib/ai/safety/outbound-screen/locales/pl.ts
+++ b/src/lib/ai/safety/outbound-screen/locales/pl.ts
@@ -98,10 +98,10 @@ export const bank: LocaleBank = {
     /\bpowoduj\w*/i,
     /\bspowodowa\w*/i,
     /\bprzyczyn\w*/i,
-    /\bdoprowadzi\w*\s+do\b/i,
+    /\bdoprowadzi[\w\u00C0-\u024F]*\s+do\b/i,
     /\bwskutek\b/i,
     /\bdzi[ęe]ki\b/i,
-    /\bodpowiedzialn\w*\s+za\b/i,
+    /\bodpowiedzialn[\w\u00C0-\u024F]*\s+za\b/i,
   ],
   continuation:
     /\b(?:kontynuuj|przyjmuj\s+dalej|zgodnie\s+z\s+zaleceniem|pozosta[ńn]\s+przy)\b/i,
diff --git a/src/lib/ai/safety/outbound-screen/shared.ts b/src/lib/ai/safety/outbound-screen/shared.ts
index 2089a552e..51bd3c533 100644
--- a/src/lib/ai/safety/outbound-screen/shared.ts
+++ b/src/lib/ai/safety/outbound-screen/shared.ts
@@ -5,32 +5,32 @@
  * is one module in `./locales/`.
  */
 
-export /**
+/**
  * Count units — a dose expressed as a number of dosage forms ("2 tablets",
  * "zwei Hübe" written as "2 Hübe", "1 comprimé"). The same forms the
  * medication wizard offers (`src/lib/medications/dose-units.ts`), in every
  * locale the banks cover. Without them "take 3 tablets a day" carried no unit
  * the screen recognised and passed every pattern.
  */
-const COUNT_UNIT =
+export const COUNT_UNIT =
   "(?:tablets?|pills?|capsules?|drops?|puffs?|sprays?|tabletten?|kapseln?|tropfen|hübe|hub|sprühstöße|comprimés?|gélules?|gouttes?|bouffées?|comprimidos?|pastillas?|c[áa]psulas?|gotas?|compress[ae]|pastigli[ae]|gocce|goccia|tabletk[aęiy]|tabletek|kapsułk[aęiy]|kapsułek|krople|kropli|druppels?|pillen|pufjes?|pastilhas?|dr[áa]geas?|jatos?|borrifadas?)";
 
-export /**
+/**
  * Dose units for a target dose a model must never name. Mass and insulin
  * units plus the count forms; `ml` is left out on purpose because
  * "increase your water intake to 2000 ml" is ordinary hydration advice.
  */
-const DRUG_UNIT = `(?:mg|mcg|µg|iu|ui|i\\.u\\.|units?|unidades?|unità|unités?|jednostek|jednostki|eenheden|eenheid|ie|i\\.e\\.|einheiten|${COUNT_UNIT})`;
+export const DRUG_UNIT = `(?:mg|mcg|µg|iu|ui|i\\.u\\.|units?|unidades?|unità|unités?|jednostek|jednostki|eenheden|eenheid|ie|i\\.e\\.|einheiten|${COUNT_UNIT})`;
 
-export /**
+/**
  * Dose units, shared across every locale. The GLP-1 + general oral-dose
  * vocabulary; `ie` / `i.e.` / `einheiten` / `unità` / `jednostk` cover the
  * insulin-unit spellings the locales use (`eenheden` is the Dutch one, `ui`
  * the Portuguese abbreviation of "unidades internacionais").
  */
-const DOSE_UNIT = `(?:ml|${DRUG_UNIT})`;
+export const DOSE_UNIT = `(?:ml|${DRUG_UNIT})`;
 
-export /**
+/**
  * What may follow a dose unit. A concentration is a lab value, not a dose:
  * "keep your glucose under 140 mg/dL" and "eGFR above 60 ml/min" name no
  * medication, and a bare `\b` let both trip the "try … N mg" pattern. A
@@ -38,9 +38,9 @@ export /**
  * matches. The lookahead also ends a unit that closes on a non-word letter
  * ("unità", "i.e.") where `\b` cannot.
  */
-const UNIT_END = "(?!\\w|\\s*\\/\\s*(?:dl|l|min|mmol)\\b)";
+export const UNIT_END = "(?!\\w|\\s*\\/\\s*(?:dl|l|min|mmol)\\b)";
 
-export /**
+/**
  * One to four words of object between a change verb and its target — "your
  * metformin", "the evening insulin dose", "votre metformine". The original
  * English pattern accepted only "it" or "your dose" in that slot, so naming the
@@ -51,27 +51,37 @@ export /**
  * dietary noun: "lower your sodium to 1500 mg" and "cut caffeine to 200 mg"
  * are nutrition guidance, not a dose change.
  */
-function objectSlot(stopWords: string, maxWords = 4): string {
+export function objectSlot(stopWords: string, maxWords = 4): string {
   return `(?:(?!(?:${stopWords})(?=\\s))[^\\s\\d.?!,;:]+\\s+){0,${maxWords}}?`;
 }
 
-export /**
+/**
  * A count written as a word: "take two tablets", "nimm zwei Tabletten",
  * "prenez deux comprimés", "weź dwie tabletki". The union of every covered
  * locale, because a provider mixes languages. Articles ("a", "un", "eine")
  * are left out on purpose: "take a tablet with food" is an article, not a
  * count, and the "eine Tablette mehr" shape is its own pattern below.
  */
-const SPELLED_COUNT =
+export const SPELLED_COUNT =
   "(?:one|two|three|four|five|six|half\\s+an?|zwei|drei|vier|fünf|sechs|eine?n?\\s+halbe|halbe|deux|trois|quatre|cinq|demi|dos|tres|cuatro|cinco|media|due|tre|quattro|cinque|mezza|dwie|dwa|trzy|cztery|pięć|jedn[ąa]|pół|één|twee|drie|vier|vijf|zes|halve|dois|duas|três|quatro|meia|meio)";
 
-export /** A target dose: a number, a drug unit, and an honest unit end. */
-const TARGET = `(?:[\\d.,]+\\s*${DRUG_UNIT}${UNIT_END}|(?<![\\w])${SPELLED_COUNT}\\s+${COUNT_UNIT}${UNIT_END})`;
+/** A target dose: a number, a drug unit, and an honest unit end. */
+export const TARGET = `(?:[\\d.,]+\\s*${DRUG_UNIT}${UNIT_END}|(?<![\\w])${SPELLED_COUNT}\\s+${COUNT_UNIT}${UNIT_END})`;
 
-export /** The letters a drug name is spelled with in the Latin-script locales. */
-const LETTER = "a-zà-ÿąćęłńóśźż";
+/**
+ * Word edges for a word that starts or ends in an accented letter. JavaScript's
+ * `\b` is defined over ASCII `\w`, so `santé\b` never matches before a space
+ * and `\bépisode` never matches after one. These treat the Latin letters with
+ * diacritics (U+00C0 to U+024F) as word letters, which is what `\b` was meant
+ * to mean.
+ */
+export const WORD_START = "(?<![\\w\\u00C0-\\u024F])";
+export const WORD_END = "(?![\\w\\u00C0-\\u024F])";
 
-export /**
+/** The letters a drug name is spelled with in the Latin-script locales. */
+export const LETTER = "a-zà-ÿąćęłńóśźż";
+
+/**
  * Common generic-name stems (INN): ramipril, losartan, bisoprolol,
  * atorvastatin, amlodipine, metformin, semaglutide, apixaban, omeprazole,
  * furosemide. A surface that passes no schedule (a nudge, the document chat)
@@ -79,7 +89,7 @@ export /**
  * so "April" is not a drug; the stems are chosen so no everyday word ends in
  * one ("comparing" is why heparin is spelled out rather than stemmed).
  */
-const INN_STEM = `(?<![${LETTER}])[${LETTER}]{2,}(?:pril|sartan|olol|statin|dipin|dypin|formin|glutid|gliflozin|gliptin|xaban|ksaban|prazol|semid)[${LETTER}]{0,4}(?![${LETTER}])|(?<![${LETTER}])(?:enoxa|dalte)?heparin[${LETTER}]{0,4}`;
+export const INN_STEM = `(?<![${LETTER}])[${LETTER}]{2,}(?:pril|sartan|olol|statin|dipin|dypin|formin|glutid|gliflozin|gliptin|xaban|ksaban|prazol|semid)[${LETTER}]{0,4}(?![${LETTER}])|(?<![${LETTER}])(?:enoxa|dalte)?heparin[${LETTER}]{0,4}`;
 
 export const ENGINE = "(?:framingham|ascvd|score2|qrisk)";