diff --git a/docs/api/async-logger-has-failed.md b/docs/api/async-logger-has-failed.md
index 9a04f7d..a827acd 100644
--- a/docs/api/async-logger-has-failed.md
+++ b/docs/api/async-logger-has-failed.md
@@ -36,7 +36,7 @@ Detailed rules explaining key parameters and behaviors
 - `run()` clears previous failure state at startup.
 - `run()` also clears the stored `last_error()` string at startup before drain work begins.
 - If the worker loop raises an error, the logger records that failure and exposes it through this flag.
-- Once set by a failed run, the flag stays `true` until a later `run()` start resets it.
+- Once set by a failed run, the flag stays `true` until a later `run()` invocation actually starts and resets it.
 - This helper is intentionally compact and should usually be paired with `last_error()` for details.
 - Failure state is about runtime drain execution, not whether records were dropped due to overflow policy.
 
@@ -73,7 +73,7 @@ e.g.:
 
 - If `has_failed()` is `true`, callers should inspect `last_error()` or `state()` for more context.
 
-- `close()` or `shutdown()` do not clear this flag by themselves; only a later `run()` start resets it.
+- `close()` or `shutdown()` do not clear this flag by themselves; only a later `run()` that has already started resets it.
 
 ### Notes
 
diff --git a/docs/api/async-logger-last-error.md b/docs/api/async-logger-last-error.md
index a0fc26f..0f54585 100644
--- a/docs/api/async-logger-last-error.md
+++ b/docs/api/async-logger-last-error.md
@@ -36,7 +36,7 @@ Detailed rules explaining key parameters and behaviors
 - `run()` resets the stored error string when the worker starts.
 - If the worker loop fails, the error text is captured from the raised exception.
 - An empty string normally means no failure has been recorded.
-- Once a failure string is recorded, it stays in place until a later `run()` start clears it.
+- Once a failure string is recorded, it stays in place until a later `run()` invocation actually starts and clears it.
 - This helper reports worker execution errors, not ordinary overflow or backpressure conditions.
 - A new successful `run()` attempt clears any previously stored error text before drain work begins again.
 
@@ -73,7 +73,7 @@ e.g.:
 
 - If callers need broader context than just the error text, they should use `state()`.
 
-- `close()` or `shutdown()` do not clear a previously recorded error string by themselves.
+- `close()` or `shutdown()` do not clear a previously recorded error string by themselves; the reset happens when a later `run()` has already started.
 
 ### Notes
 
diff --git a/docs/api/async-logger-run.md b/docs/api/async-logger-run.md
index 3f8cc89..f30b3c5 100644
--- a/docs/api/async-logger-run.md
+++ b/docs/api/async-logger-run.md
@@ -34,7 +34,7 @@ pub async fn[S : @bitlogger.Sink] AsyncLogger::run(self : AsyncLogger[S]) -> Uni
 Detailed rules explaining key parameters and behaviors
 
 - `run()` sets `is_running` to `true` before worker execution begins.
-- Every invocation clears previous failure state first by setting `has_failed=false` and `last_error()` to an empty string.
+- Every invocation clears previous failure state first by setting `has_failed=false` and `last_error()` to an empty string once that `run()` call has actually started executing.
 - The method then keeps draining records until `queue.get()` stops with `AsyncLoggerClosed` or a worker error escapes.
 - On a normal queue-close exit, `run()` clears `is_running` and returns normally.
 - On failure, the logger records `has_failed=true`, stores the error text in `last_error`, clears `is_running`, and then raises the error back out of `run()`.
diff --git a/src-async/BitLoggerAsync_test.mbt b/src-async/BitLoggerAsync_test.mbt
index 3eefa52..e9fee5c 100644
--- a/src-async/BitLoggerAsync_test.mbt
+++ b/src-async/BitLoggerAsync_test.mbt
@@ -355,6 +355,57 @@ async test "async logger records worker failures and wait_idle stops early" {
   inspect(logger.pending_count(), content="0")
 }
 
+async test "later started run resets async failure state before draining remaining backlog" {
+  let writes : Ref[Int] = Ref(0)
+  let flushes : Ref[Int] = Ref(0)
+  let logger = async_logger(
+    @bitlogger.callback_sink(fn(_) {
+      writes.val += 1
+    }),
+    config=AsyncLoggerConfig::new(
+      max_pending=4,
+      overflow=AsyncOverflowPolicy::Blocking,
+      flush=AsyncFlushPolicy::Batch,
+    ),
+    min_level=@bitlogger.Level::Info,
+    target="async.retry",
+    flush=fn(_) -> Int raise {
+      flushes.val += 1
+      if flushes.val == 1 {
+        raise TestFlushError("flush exploded once")
+      }
+      1
+    },
+  )
+
+  @async.with_task_group(group => {
+    group.spawn_bg(allow_failure=true, () => logger.run())
+    logger.info("one")
+    logger.info("two")
+    logger.wait_idle()
+    inspect(logger.has_failed(), content="true")
+    inspect(logger.last_error().contains("TestFlushError"), content="true")
+    inspect(logger.is_running(), content="false")
+    inspect(logger.pending_count(), content="1")
+
+    group.spawn_bg(() => logger.run())
+    while !logger.is_running() {
+      @async.pause()
+    }
+    inspect(logger.has_failed(), content="false")
+    inspect(logger.last_error(), content="")
+    logger.shutdown()
+  })
+
+  inspect(writes.val, content="2")
+  inspect(flushes.val, content="2")
+  inspect(logger.has_failed(), content="false")
+  inspect(logger.last_error(), content="")
+  inspect(logger.is_running(), content="false")
+  inspect(logger.is_closed(), content="true")
+  inspect(logger.pending_count(), content="0")
+}
+
 async test "library async logger keeps a smaller async facade" {
   let written_targets : Ref[Array[String]] = Ref([])
   let written_messages : Ref[Array[String]] = Ref([])