diff --git a/docs/agents/performance.md b/docs/agents/performance.md
index 1688e8b..1341b92 100644
--- a/docs/agents/performance.md
+++ b/docs/agents/performance.md
@@ -21,8 +21,8 @@
 
 ## Timing Attack Protection
 
-- `tokensMatch()` enforces a **100ms floor** on token comparison to mask cache hit vs miss timing
-- Each auth request holds a servlet thread for at least 100ms during token validation
+- `tokensMatch()` applies **randomized jitter (80–120ms)** on token comparison to mask cache hit vs miss timing
+- Each auth request holds a servlet thread for the jitter duration during token validation
 - **Thread pool sizing**: with a 200-thread pool, max theoretical auth throughput is ~2,000 req/s
 - Size `server.tomcat.threads.max` accordingly for auth-heavy workloads
 
diff --git a/docs/agents/security.md b/docs/agents/security.md
index ae57634..09329d8 100644
--- a/docs/agents/security.md
+++ b/docs/agents/security.md
@@ -25,6 +25,12 @@
 - Tokens **BCrypt-hashed** before storage (never plaintext)
 - Token comparison results **cached** (Caffeine, 1hr) to reduce BCrypt overhead
 - Token prefix (8-char) enables **O(1) DB lookups** vs O(n) BCrypt scans
+- **DUMMY_HASH** must be a valid 60-char BCrypt hash (not a short placeholder) so `BCryptPasswordEncoder.matches()` performs a full computation on the "user not found" path, preventing user enumeration via timing
+
+### Sub-token Validation
+
+- Fallback linear scan in `validateSubToken()` is bounded to `maxClients` entries to limit worst-case BCrypt comparisons
+- Primary path (bearer with client ID) uses direct O(1) lookup via `getByUserIdAndClient()`
 
 ### Authentication Flow
 
diff --git a/ogiri-core/src/main/kotlin/com/quantipixels/ogiri/security/tokens/OgiriTokenService.kt b/ogiri-core/src/main/kotlin/com/quantipixels/ogiri/security/tokens/OgiriTokenService.kt
index 85db53e..90b1a04 100644
--- a/ogiri-core/src/main/kotlin/com/quantipixels/ogiri/security/tokens/OgiriTokenService.kt
+++ b/ogiri-core/src/main/kotlin/com/quantipixels/ogiri/security/tokens/OgiriTokenService.kt
@@ -33,6 +33,7 @@ import jakarta.servlet.http.HttpServletResponse
 import java.time.Instant
 import java.time.temporal.ChronoUnit
 import java.util.Base64
+import java.util.concurrent.ThreadLocalRandom
 import java.util.concurrent.TimeUnit
 import org.slf4j.LoggerFactory
 import org.springframework.beans.factory.ObjectProvider
@@ -144,11 +145,11 @@ open class OgiriTokenService<T : OgiriToken>(
 
     val result = tokenEqualityCache.get(key) { passwordEncoder.matches(token, tokenHash) }
 
-    // Ensure minimum 100ms to mask cache hit vs miss timing
+    // Randomized jitter to mask cache hit vs miss timing without a fixed floor
+    val jitter = ThreadLocalRandom.current().nextLong(80_000_000L, 120_000_000L)
     val elapsed = System.nanoTime() - startTime
-    val minDelayNanos = 100_000_000L // 100ms
-    if (elapsed < minDelayNanos) {
-      Thread.sleep((minDelayNanos - elapsed) / 1_000_000)
+    if (elapsed < jitter) {
+      Thread.sleep((jitter - elapsed) / 1_000_000)
     }
 
     return result
@@ -723,8 +724,11 @@ open class OgiriTokenService<T : OgiriToken>(
   companion object {
     private val logger = LoggerFactory.getLogger(OgiriTokenService::class.java)
 
-    // Pre-computed BCrypt hash for timing normalization
-    private const val DUMMY_HASH = "\$2a\$10\$dummyhashvalueforconstanttimecheck"
+    // Pre-computed valid BCrypt hash for constant-time comparison when user is not found.
+    // Must be a properly formatted 60-char BCrypt hash so that BCryptPasswordEncoder.matches()
+    // performs a full BCrypt computation rather than short-circuiting on pattern validation.
+    private const val DUMMY_HASH =
+        "\$2a\$10\$II9GOMND5IObe5IA9/7EXuflfP5U77aqrWyXCEXEFpXolIDmBbLmS"
   }
 
   @Transactional
@@ -847,6 +851,7 @@ open class OgiriTokenService<T : OgiriToken>(
         repository
             .findByUserIdAndTokenSubtypeOrderByUpdatedAtDesc(user.getOgiriUserId(), subTokenName)
             .filter { OgiriTokenType.of(it.tokenType) == OgiriTokenType.SUB }
+            .take(maxClients.toInt())
     return all.any { tokenMatches(it, tokenField) && registration.validate(tokenField) }
   }