diff --git a/include/rcp/redundancy.hpp b/include/rcp/redundancy.hpp
index 4bc261c..4d83b17 100644
--- a/include/rcp/redundancy.hpp
+++ b/include/rcp/redundancy.hpp
@@ -72,9 +72,16 @@ class RedundantRequestFn {
 
         if (!cfg_.auto_promote) return ec;
 
-        // Promote standby on retriable failure.
+        // Promote standby on retriable failure. Pass along the exact
+        // RequestFn* this call observed active (captured under the lock
+        // above, before the call): if two send() calls race on the same
+        // failing active pointer, only the one whose promote_from() runs
+        // first actually flips active_; the other's observed pointer no
+        // longer matches active_ by the time it acquires the lock, so it is
+        // a no-op instead of an unconditional toggle that would flip
+        // active_ right back (see promote_from()).
         if (ec == ErrClosed || ec == ErrTimeout) {
-            promote();
+            promote_from(active);
             for (int i = 0; i < cfg_.max_retries; ++i) {
                 RequestFn* retry_active;
                 {
@@ -109,6 +116,22 @@ class RedundantRequestFn {
     }
 
 private:
+    // promote_from is send()'s internal, CAS-style counterpart to the public
+    // promote() above: it only flips active_ away from the specific pointer
+    // the caller observed failing. If active_ has already moved on (e.g. a
+    // concurrent send() on the same observed-failing pointer promoted first),
+    // this is a no-op rather than re-toggling — without this guard, two
+    // send() calls that both observe the primary failing concurrently would
+    // together apply the toggle twice (primary->standby, then straight back
+    // standby->primary), silently leaving active_ on the confirmed-bad
+    // primary for every later caller (REQ-RED-006).
+    void promote_from(RequestFn* observed_active) {
+        std::lock_guard<std::mutex> lk(mu_);
+        if (active_ == observed_active) {
+            active_ = (active_ == &primary_) ? &standby_ : &primary_;
+        }
+    }
+
     RequestFn primary_;
     RequestFn standby_;
     RequestFn* active_;
diff --git a/tests/test_redundancy.cpp b/tests/test_redundancy.cpp
index d7cffc0..4ed08ae 100644
--- a/tests/test_redundancy.cpp
+++ b/tests/test_redundancy.cpp
@@ -20,8 +20,12 @@
 #include "rcp/mock.hpp"
 #include "rcp/redundancy.hpp"
 
+#include <condition_variable>
 #include <functional>
 #include <memory>
+#include <mutex>
+#include <thread>
+#include <vector>
 
 using namespace rcp;
 using namespace std::chrono_literals;
@@ -150,3 +154,107 @@ TEST_CASE("redundancy: RedundantRequestFn is itself usable as an rcp::RequestFn
     REQUIRE_FALSE(ec);
     REQUIRE(resp.id == req.id);
 }
+
+// ── Concurrency regression: promote() must not be a blind toggle ───────────
+//
+// send() reads active_ under a short lock, then calls the RequestFn pointer
+// *outside* the lock. Two concurrent send() calls can therefore both observe
+// active_ == &primary_, both have the primary fail, and both attempt to
+// promote. An unconditional toggle (active_ = active_==&primary_ ? &standby_
+// : &primary_) would apply twice in that case -- the first promote flips
+// primary->standby, the second (serialized behind the same mutex, but blind
+// to what the caller actually observed) flips it straight back
+// standby->primary -- silently reverting to the confirmed-bad primary and
+// defeating failover for every later caller. This is a real regression risk
+// for REQ-RED-006 ("subsequent send() calls shall continue to be served by
+// the standby without reverting to the primary on their own").
+//
+// The two send() calls are driven through an explicit two-phase handshake
+// (entered_cv / release_cv), following this project's established pattern
+// for deterministic concurrency tests (see e.g. shmem's
+// "admits up to queue_capacity concurrent callers" case): both threads are
+// held inside their (still-primary) RequestFn call -- i.e. both have already
+// captured active_ == &primary_ under send()'s lock -- until both have
+// entered, and are then released together so their promote attempts
+// genuinely race on the *same* observed-primary pointer. This makes the race
+// deterministic instead of depending on OS scheduling luck.
+TEST_CASE("redundancy: concurrent send() failures on the same observed primary promote "
+          "exactly once and never revert to primary",
+          "[redundancy][thread][REQ-RED-006]") {
+    constexpr int kThreads = 8;
+
+    std::mutex              mu;
+    std::condition_variable entered_cv;
+    std::condition_variable release_cv;
+    int                     entered_count = 0;
+    bool                    may_release   = false;
+
+    // Always fails (like fail_fn), but first blocks every caller until all
+    // kThreads callers are simultaneously inside the primary call -- forcing
+    // every send() to have observed active_ == &primary_ before any of them
+    // can reach promote_from().
+    RequestFn blocking_fail_fn = [&](const Context&, const acf::AcfMessageInfo&,
+                                      const std::vector<uint8_t>&, acf::AcfMessageInfo&,
+                                      std::vector<uint8_t>&) {
+        {
+            std::lock_guard<std::mutex> lk(mu);
+            ++entered_count;
+        }
+        entered_cv.notify_all();
+
+        std::unique_lock<std::mutex> lk(mu);
+        release_cv.wait(lk, [&] { return may_release; });
+        return ErrClosed;
+    };
+
+    // The standby is a trivial, stateless always-succeeds fn rather than a
+    // mock::Server (which is not itself documented/guaranteed thread-safe
+    // for concurrent dispatch()) -- this test's own retries after promotion
+    // are deliberately concurrent, and must not introduce a second, unrelated
+    // race of their own on top of the one under test.
+    RequestFn always_ok_fn = [](const Context&, const acf::AcfMessageInfo&,
+                                 const std::vector<uint8_t>&, acf::AcfMessageInfo&,
+                                 std::vector<uint8_t>&) { return std::error_code{}; };
+
+    redundancy::RedundantRequestFn rr(blocking_fail_fn, always_ok_fn);
+    REQUIRE(rr.is_primary_active());
+
+    auto req = gpio_read_request();
+
+    std::vector<std::thread> threads;
+    threads.reserve(kThreads);
+    for (int i = 0; i < kThreads; ++i) {
+        threads.emplace_back([&] {
+            acf::AcfMessageInfo  out;
+            std::vector<uint8_t> out_payload;
+            rr.send(Context{}, req, {}, out, out_payload);
+        });
+    }
+
+    // Wait until every thread's send() is blocked inside the primary call --
+    // each has already read active_ == &primary_ under the lock, before any
+    // of them can call promote_from().
+    {
+        std::unique_lock<std::mutex> lk(mu);
+        entered_cv.wait(lk, [&] { return entered_count == kThreads; });
+    }
+
+    // Release them all together: every thread's primary call now returns
+    // ErrClosed and races to promote_from(observed == &primary_) at
+    // (approximately) the same time, serialized only by RedundantRequestFn's
+    // internal mutex.
+    {
+        std::lock_guard<std::mutex> lk(mu);
+        may_release = true;
+    }
+    release_cv.notify_all();
+
+    for (auto& th : threads) th.join();
+
+    // Exactly one net promotion must have occurred: active_ must be the
+    // standby, never reverted back to the primary that every caller observed
+    // failing (REQ-RED-006). With the old blind-toggle promote(), an even
+    // number of concurrent promote attempts on the same observed pointer
+    // cancel back out to &primary_, and this REQUIRE fails.
+    REQUIRE_FALSE(rr.is_primary_active());
+}