@@ -26,10 +26,12 @@ All notable changes to CallShield will be documented in this file.
   Dutch, Portuguese, Italian, French and a few more, and so are "Is this
   Sarah?" and "sorry, wrong number". On its own it blocks only in aggressive
   mode, since a real kid with a new phone writes the same words. Someone
-  else's parent ("this is Maya's mom, here's my new number") doesn't count.
+  else's parent ("this is Maya's mom, here's my new number") doesn't count,
+  and neither does a business's "Wrong number? Reply STOP" footer or an RCS
+  or chat message, which never reaches the inbox that says who's a stranger.
 - The SMS regression test runs 5,000 real smishing reports from the IMC 2025
   dataset (CC BY 4.0) and reports recall for each language. Content rules
-  catch 13% of them, up from 10% before reply bait, and 35% of the wrong
+  catch 12% of them, up from 10% before reply bait, and 26% of the wrong
   number and "hi mum" scams, up from 4.5%.
 - The report Worker runs the current code. It had been serving a build from
   March, so reports carried no reporter identity and the trending lists and
 
@@ -12,7 +12,7 @@
 <p align="center">
   <a href="https://github.com/SysAdminDoc/CallShield/releases/latest"><img src="https://img.shields.io/github/v/release/SysAdminDoc/CallShield?style=flat-square&color=a6e3a1" alt="Release"></a>
   <img src="https://img.shields.io/badge/Spam%20Numbers-51%2C363-f38ba8?style=flat-square" alt="51,363 Numbers">
-  <img src="https://img.shields.io/badge/JVM%20unit%20tests-1740-94e2d5?style=flat-square" alt="1740 JVM unit tests">
+  <img src="https://img.shields.io/badge/JVM%20unit%20tests-1742-94e2d5?style=flat-square" alt="1742 JVM unit tests">
   <img src="https://img.shields.io/badge/Android-10%2B-89b4fa?style=flat-square" alt="Android 10+">
   <img src="https://img.shields.io/badge/License-MIT-cba6f7?style=flat-square" alt="MIT License">
   <img src="https://img.shields.io/badge/API%20Keys-None-fab387?style=flat-square" alt="No required API keys">
@@ -216,7 +216,7 @@ The maintainer trains it by hand from the CallShield database (50K positive and
 
 SMS content regressions use a separate CC0 corpus that CallShield wrote itself. It covers eight languages, the three sender forms, each kind of link, legitimate messages and hard negatives, and the test reports precision, recall and false-positive rate by language and message type without shipping personal data. Every spam message it catches is checked again in disguise (zero-width characters between letters, Cyrillic look-alike letters, fullwidth text, a hidden character in its link) and each copy has to be caught too: `./gradlew :app:testDebugUnitTest --tests com.sysadmindoc.callshield.data.SmsEvaluationCorpusTest`.
 
-The same test runs 5,000 real smishing reports sampled from the IMC 2025 dataset (Agarwal, Papasavva, Suarez-Tangil and Vasek, "Fishing for Smishing", CC BY 4.0) in 63 languages. It prints recall for each language, with precision and false alarms taken from the clean messages above, and each large language has a floor it can't drop below. The number is humbling. Content rules alone catch 13% of those reports, and 35% of the wrong number and "hi mum" scams. Part of the gap is the dataset: it replaced every link with a placeholder, so the link rules see a link but never its host. `python scripts/sample_imc25_corpus.py` redraws the sample from a pinned copy of the dataset, and `IMC25-NOTICE.txt` next to it in `app/src/test/resources/sms-corpus/` credits it.
+The same test runs 5,000 real smishing reports sampled from the IMC 2025 dataset (Agarwal, Papasavva, Suarez-Tangil and Vasek, "Fishing for Smishing", CC BY 4.0) in 63 languages. It prints recall for each language, with precision and false alarms taken from the clean messages above, and each large language has a floor it can't drop below. The number is humbling. Content rules alone catch 12% of those reports, and 26% of the wrong number and "hi mum" scams. Part of the gap is the dataset: it replaced every link with a placeholder, so the link rules see a link but never its host. `python scripts/sample_imc25_corpus.py` redraws the sample from a pinned copy of the dataset, and `IMC25-NOTICE.txt` next to it in `app/src/test/resources/sms-corpus/` credits it.
 
 ## Features
 
@@ -515,12 +515,12 @@ RELEASE_KEY_PASSWORD=...
 ## Testing
 
 ```bash
-./gradlew testDebugUnitTest   # 1740 tests
+./gradlew testDebugUnitTest   # 1742 tests
 ./gradlew connectedDebugAndroidTest -Pandroid.testInstrumentationRunnerArguments.class=com.sysadmindoc.callshield.platform.TargetSdkBehaviorSmokeTest
 ./gradlew verifyPipelineTests # Cloudflare Worker (node) + data-pipeline and translation checks (python)
 ```
 
-The suite is **1740 total JVM unit tests**, run with Robolectric wherever a screen, a service or the database is involved.
+The suite is **1742 total JVM unit tests**, run with Robolectric wherever a screen, a service or the database is involved.
 
 Two GitHub workflows run without building the app. **Validation** runs the Worker and
 pipeline suites on every push except report-only ones (`run-pipeline-tests.ps1 -CorrectnessOnly`),
@@ -557,7 +557,7 @@ language in [issue #7](https://github.com/SysAdminDoc/CallShield/issues/7).
 | Community API | Cloudflare Workers |
 | URL Safety | Local spam-domain data, with optional PhishTank and OpenPhish |
 | Verification | Local Gradle, lint, and release-artifact checks |
-| Tests | 1740 JVM unit tests (JUnit) |
+| Tests | 1742 JVM unit tests (JUnit) |
 | Strings | 1731 string resources and 37 plural groups (translation-ready) |
 | Accessibility | 100+ content descriptions, 48dp touch targets |
 | Min SDK | 29 (Android 10) |
 
@@ -20,6 +20,9 @@ internal object SmsReplyBait {
     /** How far back from a parent word to look for "Maya's" or "my". */
     private const val POSSESSIVE_WINDOW = 24
 
+    /** "Sorry, wrong number" opens a stranger's text; further in, it's usually a business's opt-out footer. */
+    private const val WRONG_NUMBER_OPENING = 40
+
     private const val START = "(?<![\\p{L}\\p{N}])"
     private const val END = "(?![\\p{L}\\p{N}])"
     private const val SPACE = "[\\s\\p{Zs}]"
@@ -86,6 +89,10 @@ internal object SmsReplyBait {
                 "|did(?:n$APOSTROPHE?t| not) you saved? my number|long time no (?:see|talk|chat|speak))$END",
         )
 
+    // A business's first text ends "Wrong number? Reply STOP to opt out".
+    private val optOut =
+        Regex("(?iu)$START(?:(?:reply|text|send)$SPACE+[\"'“]?stop|opt$SPACE?-?$SPACE?out|unsubscribe)$END")
+
     // "Hi, is this Sarah?" The name has to be capitalized: "is this the
     // number for the flyer?" is someone with a reason to write.
     private val isThisName =
@@ -110,7 +117,12 @@ internal object SmsReplyBait {
         val signsOff = parentSignsOff.containsMatchIn(text)
         if ((parents.isNotEmpty() || signsOff) && newPhone.containsMatchIn(text)) return true
         if ((signsOff || parents.any { it !in ambiguousParents }) && chatApp.containsMatchIn(text)) return true
-        return wrongNumber.containsMatchIn(text) || (visibleLength <= MAX_OPENER_LENGTH && isThisName.containsMatchIn(text))
+        return isWrongNumberOpener(text) || (visibleLength <= MAX_OPENER_LENGTH && isThisName.containsMatchIn(text))
+    }
+
+    private fun isWrongNumberOpener(text: String): Boolean {
+        val match = wrongNumber.find(text) ?: return false
+        return match.range.first <= WRONG_NUMBER_OPENING && !optOut.containsMatchIn(text)
     }
 
     /** The parent words [text] speaks to, skipping ones that name someone's parent. */
 
@@ -738,12 +738,14 @@ class SpamRepository(
         body: String,
         realtimeCall: Boolean = true,
         prefsSnapshot: Preferences? = null,
+        fromSmsInbox: Boolean = true,
     ): SpamCheckResult =
         spamRepositoryImpl.isSpamSms(
             number = number,
             body = body,
             realtimeCall = realtimeCall,
             prefsSnapshot = prefsSnapshot,
+            fromSmsInbox = fromSmsInbox,
         )
 
     // ── Pipeline trace (diagnostic) ─────────────────────────────────────
 
@@ -49,6 +49,18 @@ class SpamRepositoryAdapter(
             prefsSnapshot = prefsSnapshot,
         )
 
+    override suspend fun checkSpamMessageOutsideInbox(
+        number: String,
+        body: String,
+        prefsSnapshot: Preferences?,
+    ): SpamCheckResult =
+        repository.isSpamSms(
+            number = number,
+            body = body,
+            prefsSnapshot = prefsSnapshot,
+            fromSmsInbox = false,
+        )
+
     override suspend fun syncDatabase(force: Boolean): SyncResult = repository.syncFromGitHub(force = force)
 
     override suspend fun blockNumber(
 
@@ -275,6 +275,7 @@ class SpamRepositoryImpl(
         body: String,
         realtimeCall: Boolean = true,
         prefsSnapshot: Preferences? = null,
+        fromSmsInbox: Boolean = true,
     ): SpamCheckResult {
         val prefs = prefsSnapshot ?: settingsRepository.readPrefsSnapshot()
         val senderProvenance = senderProvenanceResolver.resolve(number, senderRegionIso)
@@ -283,9 +284,12 @@ class SpamRepositoryImpl(
             canonicalPhone.isNotBlank() &&
                 checkerDependencies.smsContextChecker.isTrustedSender(context, canonicalPhone)
         // Reply bait only means something in a stranger's first text. A rescan
-        // finds the old message itself in the inbox, so it applies only live.
+        // finds the old message itself in the inbox, so it applies only live,
+        // and an RCS or chat message never reaches the inbox, so every one of
+        // them would read as a first text.
         val smsFirstContact =
             realtimeCall &&
+                fromSmsInbox &&
                 !smsContextTrusted &&
                 canonicalPhone.isNotBlank() &&
                 checkerDependencies.smsContextChecker.isFirstMessageFrom(context, canonicalPhone)
 
@@ -19,4 +19,15 @@ interface SpamCheckRepository {
         realtimeCall: Boolean,
         prefsSnapshot: Preferences?,
     ): SpamCheckResult
+
+    /**
+     * A live message the SMS inbox never holds (RCS, a chat app's notification).
+     * Whether it's someone's first text can't be read from the inbox, so it's
+     * never treated as one.
+     */
+    suspend fun checkSpamMessageOutsideInbox(
+        number: String,
+        body: String,
+        prefsSnapshot: Preferences?,
+    ): SpamCheckResult = checkSpamSms(number, body, realtimeCall = true, prefsSnapshot = prefsSnapshot)
 }
@@ -12,11 +12,16 @@ class CheckSpamSmsUseCase(
         body: String,
         realtimeCall: Boolean = true,
         prefsSnapshot: Preferences? = null,
+        fromSmsInbox: Boolean = true,
     ): SpamCheckResult =
-        repository.checkSpamSms(
-            number = number,
-            body = body,
-            realtimeCall = realtimeCall,
-            prefsSnapshot = prefsSnapshot,
-        )
+        if (fromSmsInbox) {
+            repository.checkSpamSms(
+                number = number,
+                body = body,
+                realtimeCall = realtimeCall,
+                prefsSnapshot = prefsSnapshot,
+            )
+        } else {
+            repository.checkSpamMessageOutsideInbox(number, body, prefsSnapshot)
+        }
 }
@@ -251,7 +251,7 @@ class RcsNotificationListener : NotificationListenerService() {
         val result =
             senderNumber
                 .takeIf { spamBlockingEnabled && filterAsciiDigits(it).length >= 7 }
-                ?.let { number -> checkSpamSms(number, effectiveBody.orEmpty(), prefsSnapshot = prefs) }
+                ?.let { number -> checkSpamSms(number, effectiveBody.orEmpty(), prefsSnapshot = prefs, fromSmsInbox = false) }
         val contentVerdict =
             effectiveBody?.takeIf { spamBlockingEnabled && result == null }?.let { bodyText ->
                 contentVerdict(
 
@@ -8,6 +8,8 @@ import android.database.MatrixCursor
 import android.net.Uri
 import android.provider.Telephony
 import androidx.test.core.app.ApplicationProvider
+import com.sysadmindoc.callshield.data.repository.SpamRepositoryAdapter
+import com.sysadmindoc.callshield.domain.usecase.CheckSpamSmsUseCase
 import kotlinx.coroutines.runBlocking
 import org.junit.After
 import org.junit.Assert.assertEquals
@@ -92,6 +94,16 @@ class ReplyBaitFirstContactTest {
             assertFalse(fixture.repository.isSpamSms(number, body, realtimeCall = false).isSpam)
         }
 
+    @Test
+    fun `an RCS or chat message never counts as a first text`() =
+        runBlocking {
+            // They never reach the SMS inbox, so every one would read as a stranger's first.
+            val result = CheckSpamSmsUseCase(SpamRepositoryAdapter(fixture.repository))(number, body, fromSmsInbox = false)
+
+            assertFalse(result.signals.toString(), "reply_bait" in result.signals)
+            assertFalse(result.isSpam)
+        }
+
     @Test
     fun `a short code is not a stranger's phone`() {
         val context = ApplicationProvider.getApplicationContext<Application>()