diff --git a/Sources/TranscriberCore/Engines/CohereMLXBackend.swift b/Sources/TranscriberCore/Engines/CohereMLXBackend.swift
index f66eed4..f52735f 100644
--- a/Sources/TranscriberCore/Engines/CohereMLXBackend.swift
+++ b/Sources/TranscriberCore/Engines/CohereMLXBackend.swift
@@ -72,8 +72,14 @@ public protocol CohereMLXAudioDurationReading: Sendable {
 /// Terminal errors surfaced by `CohereMLXBackend.transcribe(_:)`. Marked
 /// terminal so `TranscriptionWorker.isTransient` (which classifies by type)
 /// does not retry — deterministic local inference would just loop again.
+///
+/// `reason` is a structural summary only (tri-gram stats, unique-word fraction)
+/// and never contains transcript content. `TranscriptionWorker` persists
+/// `String(describing: error)` into `transcript.md`/`metadata.json` on
+/// failure, so anything embedded in this case lands on disk — keep
+/// transcript text out of the payload.
 public enum CohereMLXBackendError: Error, Sendable, Equatable {
-    case degenerateOutput(reason: String, sample: String)
+    case degenerateOutput(reason: String)
 }
 
 /// Heuristic guard against decode-loop output (model emits the same phrase
@@ -96,7 +102,12 @@ enum DegenerateOutputDetector {
             let key = "\(words[i]) \(words[i + 1]) \(words[i + 2])"
             counts[key, default: 0] += 1
         }
+        // Require BOTH an absolute minimum repeat count AND a high fraction.
+        // The fraction alone would flag a 30-word transcript with three
+        // repeats of "thank you very" (3/28 ≈ 11%) as degenerate; an
+        // observed real-world loop has 12+ repeats, well above the floor.
         if let (top, n) = counts.max(by: { $0.value < $1.value }),
+           n >= 5,
            Double(n) / Double(total) > 0.08 {
             return "tri-gram \"\(top)\" repeats \(n)/\(total) times"
         }
@@ -168,9 +179,12 @@ public final class CohereMLXBackend: TranscriptionEngine, @unchecked Sendable {
         )
         let output = try await adapter.transcribe(localRequest)
         if let reason = DegenerateOutputDetector.evaluate(output.text) {
-            let sample = String(output.text.prefix(120))
-            Log.engine.error("Cohere MLX produced degenerate output: \(reason, privacy: .public)")
-            throw CohereMLXBackendError.degenerateOutput(reason: reason, sample: sample)
+            // Sample stays in the unified log at `.private` for live debugging
+            // and is intentionally NOT carried on the thrown error — the
+            // worker persists `String(describing:)` to disk and meeting
+            // content must not land in the failure artifact.
+            Log.engine.error("Cohere MLX produced degenerate output: \(reason, privacy: .public); sample: \(output.text.prefix(120), privacy: .private)")
+            throw CohereMLXBackendError.degenerateOutput(reason: reason)
         }
         let utterances: [EngineResponse.Utterance]
         if output.segments.isEmpty {
diff --git a/Tests/TranscriberCoreTests/Engines/CohereRustBackendTests.swift b/Tests/TranscriberCoreTests/Engines/CohereRustBackendTests.swift
index 169e37c..aae4812 100644
--- a/Tests/TranscriberCoreTests/Engines/CohereRustBackendTests.swift
+++ b/Tests/TranscriberCoreTests/Engines/CohereRustBackendTests.swift
@@ -407,16 +407,38 @@ final class CohereMLXBackendTests: XCTestCase {
             XCTFail("Backend must throw on degenerate output, not silently return")
         } catch let error as CohereMLXBackendError {
             switch error {
-            case .degenerateOutput(let reason, let sample):
+            case .degenerateOutput(let reason):
                 XCTAssertTrue(reason.contains("tri-gram") || reason.contains("unique-word"),
                               "Reason should identify the failure mode, got: \(reason)")
-                XCTAssertFalse(sample.isEmpty, "Sample should include a snippet of the failing transcript")
-                XCTAssertLessThanOrEqual(sample.count, 120, "Sample is capped at 120 chars")
+                let serialized = String(describing: error)
+                XCTAssertFalse(
+                    serialized.contains("I think that's what I'm hearing"),
+                    "String(describing:) must not embed transcript content; TranscriptionWorker persists this verbatim to disk"
+                )
             }
         } catch {
             XCTFail("Expected CohereMLXBackendError.degenerateOutput, got \(error)")
         }
     }
+
+    func testDetectorDoesNotFireOnShortTranscriptWithFewRepeats() {
+        // 30+ words, "thank you very" tri-gram appears 3 times. Before the
+        // `n >= 5` floor was added this fired (3/28 ≈ 11% > 8%) and the
+        // backend threw, costing the user a usable transcript.
+        let words: [String] = [
+            "thank", "you", "very", "much", "for",
+            "joining", "the", "call", "today",
+            "thank", "you", "very", "much", "again",
+            "before", "we", "wrap",
+            "thank", "you", "very", "kindly",
+            "and", "have", "a", "wonderful", "rest", "of", "your", "afternoon", "everyone"
+        ]
+        XCTAssertGreaterThanOrEqual(words.count, 30)
+        XCTAssertNil(
+            DegenerateOutputDetector.evaluate(words.joined(separator: " ")),
+            "Three repeats of a common tri-gram in a short transcript is not degenerate"
+        )
+    }
 }
 
 private final class RecordingLocalAdapter: CohereMLXTranscribing, @unchecked Sendable {