diff --git a/src/analyze.ts b/src/analyze.ts
index d07d378..fb91957 100644
--- a/src/analyze.ts
+++ b/src/analyze.ts
@@ -430,14 +430,15 @@ export function analyze(text: string, options: AnalyzeOptions = {}): AnalysisRes
   // are symbols, not letters (category So): circled "Ⓐⓓⓜⓘⓝ", parenthesized,
   // squared. A run of >=2 such glyphs whose NFKC form is an ASCII word is the
   // same disguised-word attack as fullwidth/math styling (which tokenize and are
-  // handled per-token). Runs of styled DIGITS (①②③ -> 123) fold to non-letters
-  // and are ignored, keeping legitimate enclosed numbering out.
+  // handled per-token).
   {
     let runStart = -1;
     let runCount = 0;
     const flush = (end: number) => {
       if (runStart >= 0 && runCount >= 2) {
         const fold = text.slice(runStart, end).normalize('NFKC').normalize('NFC');
+        // Belt-and-braces: every character in the run already passed these two
+        // tests individually, so the concatenation cannot fail them.
         if (ASCII_PRINTABLE_RE.test(fold) && ASCII_LETTER_RE.test(fold)) {
           signals.confusable_word = true;
           words.push({
@@ -458,6 +459,8 @@ export function analyze(text: string, options: AnalyzeOptions = {}): AnalysisRes
       const width = cp > 0xffff ? 2 : 1;
       const ch = String.fromCodePoint(cp);
       const nf = ch.normalize('NFKC');
+      // Styled DIGITS (①②③ -> 123) fail ASCII_LETTER_RE and so never join a
+      // run, keeping legitimate enclosed numbering out.
       const styledLetter =
         !TOKEN_CHAR_RE.test(ch) &&
         nf !== ch &&
diff --git a/test/analyze.test.ts b/test/analyze.test.ts
index 032b8be..795cbea 100644
--- a/test/analyze.test.ts
+++ b/test/analyze.test.ts
@@ -77,6 +77,16 @@ describe('analyze — spoofed content', () => {
     expect(hot!.signals).toContain('confusable_word');
   });
 
+  it('flags Unicode non-characters as illegal code points', () => {
+    // Two disjoint ranges reach the same verdict: the BMP block U+FDD0–U+FDEF
+    // and the U+xxFFFE/U+xxFFFF pair that closes every plane.
+    for (const text of ['hi ﷐ there', 'hi ￾ there', 'hi \u{1fffe} there']) {
+      const r = analyze(text);
+      expect(r.signals.illegal).toBe(true);
+      expect(r.normalized).toBe('hi  there');
+    }
+  });
+
   it('reports codepoint-accurate word offsets', () => {
     const r = analyze(CUSTOMER_SAMPLE);
     for (const w of r.words) {
@@ -109,6 +119,20 @@ describe('analyze — legitimate content', () => {
     expect(r.normalized).toBe(text);
   });
 
+  it('does not call a styled word confusable unless the whole word folds to ASCII', () => {
+    // "𝗉𝗋é" has two styled letters that individually fold to ASCII, but the
+    // token folds to "pré" — still non-ASCII, so it is not a disguised word.
+    const r = analyze('𝗉𝗋é');
+    expect(r.signals.confusable_word).toBe(false);
+  });
+
+  it('ignores runs of enclosed digits, which fold to non-letters', () => {
+    // ①②③ → "123": a styled run, but not a disguised WORD. Legitimate
+    // enclosed numbering must survive untouched.
+    const r = analyze('see item ①②③ below');
+    expect(r.signals.confusable_word).toBe(false);
+  });
+
   it('treats expected scripts as legitimate for whole words', () => {
     // Without expectedScripts this could look suspicious in Latin-dominant
     // text; with them, whole Cyrillic words are the sender's normal traffic.