diff --git a/neuralnetworks/1.2/vts/functional/Callbacks.cpp b/neuralnetworks/1.2/vts/functional/Callbacks.cpp
index cfaf91d..a607a08 100644
--- a/neuralnetworks/1.2/vts/functional/Callbacks.cpp
+++ b/neuralnetworks/1.2/vts/functional/Callbacks.cpp
@@ -14,160 +14,128 @@
  * limitations under the License.
  */
 
+#define LOG_TAG "Callbacks"
+
 #include "1.2/Callbacks.h"
+
 #include <android-base/logging.h>
 
-namespace android {
-namespace hardware {
-namespace neuralnetworks {
-namespace V1_2 {
-namespace implementation {
+#include <limits>
 
-CallbackBase::CallbackBase() : mNotified(false) {}
+namespace android::hardware::neuralnetworks::V1_2::implementation {
 
-CallbackBase::~CallbackBase() {
-    // Note that we cannot call CallbackBase::join_thread from here:
-    // CallbackBase is intended to be reference counted, and it is possible that
-    // the reference count drops to zero in the bound thread, causing the
-    // bound thread to call this destructor. If a thread tries to join
-    // itself, it throws an exception, producing a message like the
-    // following:
-    //
-    //     terminating with uncaught exception of type std::__1::system_error:
-    //     thread::join failed: Resource deadlock would occur
-}
+constexpr Timing kNoTiming = {.timeOnDevice = std::numeric_limits<uint64_t>::max(),
+                              .timeInDriver = std::numeric_limits<uint64_t>::max()};
 
-void CallbackBase::wait() {
-    std::unique_lock<std::mutex> lock(mMutex);
-    mCondition.wait(lock, [this] { return mNotified; });
-    join_thread_locked();
-}
-
-bool CallbackBase::on_finish(std::function<bool(void)> post_work) {
-    std::lock_guard<std::mutex> lock(mMutex);
-    if (mPostWork != nullptr) {
-        LOG(ERROR) << "CallbackBase::on_finish -- a post-work function has already been bound to "
-                      "this callback object";
-        return false;
-    }
-    if (post_work == nullptr) {
-        LOG(ERROR) << "CallbackBase::on_finish -- the new post-work function is invalid";
-        return false;
-    }
-    mPostWork = std::move(post_work);
-    return true;
-}
-
-bool CallbackBase::bind_thread(std::thread&& asyncThread) {
-    std::lock_guard<std::mutex> lock(mMutex);
-    if (mThread.joinable()) {
-        LOG(ERROR) << "CallbackBase::bind_thread -- a thread has already been bound to this "
-                      "callback object";
-        return false;
-    }
-    if (!asyncThread.joinable()) {
-        LOG(ERROR) << "CallbackBase::bind_thread -- the new thread is not joinable";
-        return false;
-    }
-    mThread = std::move(asyncThread);
-    return true;
-}
-
-void CallbackBase::join_thread() {
-    std::lock_guard<std::mutex> lock(mMutex);
-    join_thread_locked();
-}
-
-void CallbackBase::notify() {
-    {
-        std::lock_guard<std::mutex> lock(mMutex);
-        mNotified = true;
-        if (mPostWork != nullptr) {
-            bool success = mPostWork();
-            if (!success) {
-                LOG(ERROR) << "CallbackBase::notify -- post work failed";
-            }
-        }
-    }
-    mCondition.notify_all();
-}
-
-void CallbackBase::join_thread_locked() {
-    if (mThread.joinable()) {
-        mThread.join();
-    }
-}
-
-PreparedModelCallback::PreparedModelCallback()
-    : mErrorStatus(ErrorStatus::GENERAL_FAILURE), mPreparedModel(nullptr) {}
-
-PreparedModelCallback::~PreparedModelCallback() {}
+// PreparedModelCallback methods begin here
 
 Return<void> PreparedModelCallback::notify(ErrorStatus errorStatus,
                                            const sp<V1_0::IPreparedModel>& preparedModel) {
-    mErrorStatus = errorStatus;
-    mPreparedModel = preparedModel;
-    CallbackBase::notify();
+    {
+        std::lock_guard<std::mutex> hold(mMutex);
+
+        // quick-return if object has already been notified
+        if (mNotified) {
+            return Void();
+        }
+
+        // store results and mark as notified
+        mErrorStatus = errorStatus;
+        mPreparedModel = preparedModel;
+        mNotified = true;
+    }
+
+    mCondition.notify_all();
     return Void();
 }
 
 Return<void> PreparedModelCallback::notify_1_2(ErrorStatus errorStatus,
                                                const sp<V1_2::IPreparedModel>& preparedModel) {
-    mErrorStatus = errorStatus;
-    mPreparedModel = preparedModel;
-    CallbackBase::notify();
-    return Void();
+    return notify(errorStatus, preparedModel);
 }
 
-ErrorStatus PreparedModelCallback::getStatus() {
+void PreparedModelCallback::wait() const {
+    std::unique_lock<std::mutex> lock(mMutex);
+    mCondition.wait(lock, [this] { return mNotified; });
+}
+
+ErrorStatus PreparedModelCallback::getStatus() const {
     wait();
     return mErrorStatus;
 }
 
-sp<V1_0::IPreparedModel> PreparedModelCallback::getPreparedModel() {
+sp<V1_0::IPreparedModel> PreparedModelCallback::getPreparedModel() const {
     wait();
     return mPreparedModel;
 }
 
-ExecutionCallback::ExecutionCallback() : mErrorStatus(ErrorStatus::GENERAL_FAILURE) {}
-
-ExecutionCallback::~ExecutionCallback() {}
+// ExecutionCallback methods begin here
 
 Return<void> ExecutionCallback::notify(ErrorStatus errorStatus) {
-    mErrorStatus = errorStatus;
-    mOutputShapes = {};
-    mTiming = {.timeOnDevice = UINT64_MAX, .timeInDriver = UINT64_MAX};
-    CallbackBase::notify();
+    notifyInternal(errorStatus, {}, kNoTiming);
     return Void();
 }
 
 Return<void> ExecutionCallback::notify_1_2(ErrorStatus errorStatus,
                                            const hidl_vec<OutputShape>& outputShapes,
                                            const Timing& timing) {
-    mErrorStatus = errorStatus;
-    mOutputShapes = outputShapes;
-    mTiming = timing;
-    CallbackBase::notify();
+    if (errorStatus == ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) {
+        // outputShapes must not be empty if OUTPUT_INSUFFICIENT_SIZE.
+        if (outputShapes.size() == 0) {
+            LOG(ERROR) << "Notified with empty output shape vector when OUTPUT_INSUFFICIENT_SIZE";
+            notifyInternal(ErrorStatus::GENERAL_FAILURE, {}, kNoTiming);
+            return Void();
+        }
+    } else if (errorStatus != ErrorStatus::NONE) {
+        // outputShapes must be empty if errorStatus is neither NONE nor OUTPUT_INSUFFICIENT_SIZE.
+        if (outputShapes.size() != 0) {
+            LOG(ERROR) << "Notified with non-empty output shape vector when error status is "
+                          "neither NONE nor OUTPUT_INSUFFICIENT_SIZE";
+            notifyInternal(ErrorStatus::GENERAL_FAILURE, {}, kNoTiming);
+            return Void();
+        }
+    }
+    notifyInternal(errorStatus, outputShapes, timing);
     return Void();
 }
 
-ErrorStatus ExecutionCallback::getStatus() {
+void ExecutionCallback::wait() const {
+    std::unique_lock<std::mutex> lock(mMutex);
+    mCondition.wait(lock, [this] { return mNotified; });
+}
+
+ErrorStatus ExecutionCallback::getStatus() const {
     wait();
     return mErrorStatus;
 }
 
-const std::vector<OutputShape>& ExecutionCallback::getOutputShapes() {
+const std::vector<OutputShape>& ExecutionCallback::getOutputShapes() const {
     wait();
     return mOutputShapes;
 }
 
-Timing ExecutionCallback::getTiming() {
+Timing ExecutionCallback::getTiming() const {
     wait();
     return mTiming;
 }
 
-}  // namespace implementation
-}  // namespace V1_2
-}  // namespace neuralnetworks
-}  // namespace hardware
-}  // namespace android
+void ExecutionCallback::notifyInternal(ErrorStatus errorStatus,
+                                       const hidl_vec<OutputShape>& outputShapes,
+                                       const Timing& timing) {
+    {
+        std::lock_guard<std::mutex> hold(mMutex);
+
+        // quick-return if object has already been notified
+        if (mNotified) {
+            return;
+        }
+
+        mErrorStatus = errorStatus;
+        mOutputShapes = outputShapes;
+        mTiming = timing;
+        mNotified = true;
+    }
+    mCondition.notify_all();
+}
+
+}  // namespace android::hardware::neuralnetworks::V1_2::implementation
diff --git a/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp b/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
index adbf224..ac92a5b 100644
--- a/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
+++ b/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
@@ -26,6 +26,7 @@
 #include <cstdio>
 #include <cstdlib>
 #include <random>
+#include <thread>
 
 #include "1.2/Callbacks.h"
 #include "GeneratedTestHarness.h"
diff --git a/neuralnetworks/1.2/vts/functional/include/1.2/Callbacks.h b/neuralnetworks/1.2/vts/functional/include/1.2/Callbacks.h
index 212a887..2992c0c 100644
--- a/neuralnetworks/1.2/vts/functional/include/1.2/Callbacks.h
+++ b/neuralnetworks/1.2/vts/functional/include/1.2/Callbacks.h
@@ -17,299 +17,218 @@
 #ifndef ANDROID_HARDWARE_NEURALNETWORKS_V1_2_CALLBACKS_H
 #define ANDROID_HARDWARE_NEURALNETWORKS_V1_2_CALLBACKS_H
 
+#include <android-base/thread_annotations.h>
+#include <android/hardware/neuralnetworks/1.0/IExecutionCallback.h>
+#include <android/hardware/neuralnetworks/1.0/IPreparedModelCallback.h>
 #include <android/hardware/neuralnetworks/1.2/IExecutionCallback.h>
 #include <android/hardware/neuralnetworks/1.2/IPreparedModelCallback.h>
 #include <hidl/Status.h>
-#include <chrono>
 #include <condition_variable>
-#include <functional>
 #include <mutex>
-#include <thread>
 
-namespace android {
-namespace hardware {
-namespace neuralnetworks {
-namespace V1_2 {
-namespace implementation {
+/*
+ * The Callback classes are used internally by the NeuralNetworks runtime to
+ * synchronize between different threads. An asynchronous task is launched
+ * paired with a callback object. When a client thread requires the output being
+ * generated by the asynchronous task, the client thread can wait for the result
+ * and be blocked until it has completed. Any wait may safely be called
+ * concurrently, even on the same callback object. When the asynchronous task
+ * has finished its workload, it must immediately call "notify*". If the
+ * asynchronous task has failed to launch, the function that tried to launch the
+ * asynchronous task must immediately call "notify*". This "notify*" call
+ * awakens any client threads waiting on the callback object.
+ *
+ * These classes exist to enable synchronization across HIDL. When
+ * synchronization is only required in the same process, consider using
+ * std::future, std::mutex, std::condition_variable, or std::experimental::latch
+ * instead.
+ */
+
+namespace android::hardware::neuralnetworks::V1_2::implementation {
 
 using V1_0::ErrorStatus;
 
 /**
- * The CallbackBase class is used internally by the NeuralNetworks runtime to
- * synchronize between different threads. An asynchronous task is launched
- * paired with a callback object. When a client thread requires the output being
- * generated by the asynchronous task, the client thread can wait for the result
- * and be blocked until it has completed or a timeout condition has been
- * reached. Any wait* may safely be called concurrently, even on the same
- * callback object. When the asynchronous task has finished its workload, it
- * must immediately call "notify". If the asynchronous task has failed to launch,
- * the function that tried to launch the asynchronous task must immediately call
- * "notify". This "notify" call awakens any client threads waiting on the
- * callback object.
- *
- * The CallbackBase class implements some of the base synchronization common to
- * both PrepareModelCallback and ExecutionCallback. For consistency, any HIDL
- * callback class must inherit from CallbackBase as well as the HIDL callback
- * interface it implements.
- *
- * This class exists to enable synchronization across HIDL. When synchronization
- * is only required in the same process, consider using std::future, std::mutex,
- * std::condition_variable, or std::experimental::latch instead.
- */
-class CallbackBase {
-  public:
-    CallbackBase();
-    ~CallbackBase();
-
-    /**
-     * CallbackBase::wait blocks until notify has been called on the callback
-     * object.
-     */
-    void wait();
-
-    /**
-     * CallbackBase::wait_for blocks until notify has been called on the
-     * callback object or the time duration from the time the wait_for function
-     * was called has expired, whichever comes first.
-     *
-     * @return Status std::cv_status::no_timeout if the callback was notified
-     *                before the time duration expired, std::cv_status::timeout
-     *                otherwise.
-     */
-    template <class Rep, class Period>
-    std::cv_status wait_for(const std::chrono::duration<Rep, Period>& timeout_duration);
-
-    /**
-     * CallbackBase::on_finish binds a function to the callback object. This
-     * bound function will be executed when CallbackBase::notify is called,
-     * before any calls to wait* return. (Note that CallbackBase::wait_for can
-     * return std::cv_status::timeout before CallbackBase::notify is called for
-     * the first time, and hence before the bound function is executed.)
-     *
-     * The bound function must not synchronize with or otherwise access the
-     * callback object it is bound to, as this could cause a deadlock.
-     *
-     * CallbackBase::on_finish can be called at most once on a given callback
-     * object, and the call to CallbackBase::on_finish must finish before
-     * CallbackBase::notify is called.
-     *
-     * @param post_work Function to be invoked the first time
-     *                  CallbackBase::notify is called. Must have a target --
-     *                  i.e., must not compare equal to nullptr. post_work
-     *                  returns true if it successfully completes, false if it
-     *                  fails.
-     * @return bool True if the function was successfully bound, false if
-     *              unsuccessful.
-     *
-     * TODO: Why does the return value of the callback matter?
-     */
-    bool on_finish(std::function<bool(void)> post_work);
-
-    /**
-     * CallbackBase::bind_thread binds a thread to the event for later use by
-     * CallbackBase::join_thread.
-     *
-     * The thread must be passed using std::move.
-     *
-     * Once a thread is bound with CallbackBase::bind_thread, the client code
-     * should ensure that one of the following occurs before the event is
-     * destroyed:
-     * - CallbackBase::join_thread has been called.
-     * - CallbackBase::wait has been called.
-     * - CallbackBase::wait_for has been called and returned other than
-     *   std::cv_status::no_timeout.
-     *
-     * The bound thread shall not call any CallbackBase method with the
-     * exception of CallbackBase::notify, which it must call when the thread has
-     * finished its computation.
-     *
-     * CallbackBase::bind_thread can be called at most once on a given callback
-     * object.
-     *
-     * @param asyncThread Thread to be bound to the callback object. The thread
-     *                    object must represent a thread of execution -- i.e.,
-     *                    asyncThread.joinable() must be true.
-     * @return bool True if successful, false if thread was not properly bound.
-     */
-    bool bind_thread(std::thread&& asyncThread);
-
-    /**
-     * CallbackBase::join_thread ensures that the thread (if any) bound to this
-     * event with CallbackBase::bind_thread has fully finished and cleaned its
-     * resources. It is legal to call this function multiple times, concurrently
-     * or sequentially.
-     */
-    void join_thread();
-
-  protected:
-    /**
-     * CallbackBase::notify enables all prior and future wait* calls on the
-     * callback object to proceed. The call to CallbackBase::notify happens
-     * before any wait* calls on this callback object return (except in the case
-     * of wait_for timing out). The asynchronous call the callback object is
-     * paired with must ensure that any update to state that should be visible
-     * to the caller of wait* happens before the call to CallbackBase::notify.
-     *
-     * CallbackBase::notify must be called exactly once on a given callback
-     * object.
-     */
-    void notify();
-
-  private:
-    // Same as CallbackBase::join_thread but assumes we already hold a lock on
-    // mMutex.
-    void join_thread_locked();
-
-    bool mNotified;
-    std::mutex mMutex;
-    std::condition_variable mCondition;
-    std::function<bool(void)> mPostWork;
-    std::thread mThread;
-};
-
-/**
  * The PreparedModelCallback class is used to receive the error status of
  * preparing a model as well as the prepared model from a task executing
- * asynchronously with respect to the runtime. If a calling thread calls wait*
+ * asynchronously with respect to the runtime. If a calling thread calls wait
  * or get* on a PreparedModelCallback object and the corresponding asynchronous
  * task has not finished preparing the model, the calling thread will block
- * until the asynchronous task has either called notify or notify_1_2. For more
- * information on the synchronization behavior, refer to the CallbackBase class.
+ * until the asynchronous task has either called notify or notify_1_2.
  *
- * This class inherits the basic blocking and signaling calls from
- * CallbackBase, and implements the HIDL notify and notify_1_2 calls from
- * IPreparedModelCallback. This callback object is passed as an argument to
- * IDevice::prepareModel.
+ * If the callback object is notified more than once, only the results of the
+ * first call to notify* are used, and the results from subsequent calls are
+ * discarded.
+ *
+ * This callback object is passed as an argument to IDevice::prepareModel*.
  */
-class PreparedModelCallback : public CallbackBase, public IPreparedModelCallback {
+class PreparedModelCallback : public IPreparedModelCallback {
   public:
-    PreparedModelCallback();
-    ~PreparedModelCallback() override;
-
     /**
-     * IPreparedModelCallback::notify and IPreparedModelCallback::notify_1_2
-     * mark the callback object with the return status of the asynchronous
-     * model preparation along with the prepared model, and call
-     * CallbackBase::notify, enabling all prior and future wait* calls on the
-     * PreparedModelCallback object to proceed. For more information on the
-     * synchronization behavior, refer to the CallbackBase class.
+     * IPreparedModelCallback::notify marks the callback object with the return
+     * status of the asynchronous model preparation along with the prepared
+     * model, and allows all prior and future wait calls on the
+     * PreparedModelCallback object to proceed.
      *
-     * Either IPreparedModelCallback::notify or IPreparedModelCallback::notify_1_2
-     * must be called exactly once on a given PreparedModelCallback object.
+     * Either IPreparedModelCallback::notify or
+     * IPreparedModelCallback::notify_1_2 must be called on a given
+     * PreparedModelCallback object.
+     *
+     * If the callback object is notified more than once, only the results of
+     * the first call to notify* are used, and the results from subsequent calls
+     * are discarded.
      *
      * @param status Error status returned from asynchronously preparing the
-     *               model; will be:
-     *               - NONE if the asynchronous preparation was successful
-     *               - DEVICE_UNAVAILABLE if driver is offline or busy
-     *               - GENERAL_FAILURE if there is an unspecified error
-     *               - INVALID_ARGUMENT if the input model is invalid
+     *     model; will be:
+     *     - NONE if the asynchronous preparation was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if the input model is invalid
      * @param preparedModel Returned model that has been prepared for execution,
-     *                      nullptr if the model was unable to be prepared.
+     *     nullptr if the model was unable to be prepared.
      */
     Return<void> notify(ErrorStatus status, const sp<V1_0::IPreparedModel>& preparedModel) override;
+
+    /**
+     * IPreparedModelCallback::notify_1_2 marks the callback object with the
+     * return status of the asynchronous model preparation along with the
+     * prepared model, and allows all prior and future wait calls on the
+     * PreparedModelCallback object to proceed.
+     *
+     * Either IPreparedModelCallback::notify or
+     * IPreparedModelCallback::notify_1_2 must be called on a given
+     * PreparedModelCallback object.
+     *
+     * If the callback object is notified more than once, only the results of
+     * the first call to notify* are used, and the results from subsequent calls
+     * are discarded.
+     *
+     * @param status Error status returned from asynchronously preparing the
+     *     model; will be:
+     *     - NONE if the asynchronous preparation was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if the input model is invalid
+     * @param preparedModel Returned model that has been prepared for execution,
+     *     nullptr if the model was unable to be prepared.
+     */
     Return<void> notify_1_2(ErrorStatus status,
                             const sp<V1_2::IPreparedModel>& preparedModel) override;
 
     /**
+     * PreparedModelCallback::wait blocks until notify* has been called on the
+     * callback object.
+     */
+    void wait() const;
+
+    /**
      * Retrieves the error status returned from the asynchronous task launched
-     * by IDevice::prepareModel. If IDevice::prepareModel has not finished
+     * by IDevice::prepareModel*. If IDevice::prepareModel* has not finished
      * asynchronously preparing the model, this call will block until the
      * asynchronous task notifies the object.
      *
      * @return status Error status returned from asynchronously preparing the
-     *                model; will be:
-     *                - NONE if the asynchronous preparation was successful
-     *                - DEVICE_UNAVAILABLE if driver is offline or busy
-     *                - GENERAL_FAILURE if there is an unspecified error
-     *                - INVALID_ARGUMENT if the input model is invalid
+     *     model; will be:
+     *     - NONE if the asynchronous preparation was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if the input model is invalid
      */
-    ErrorStatus getStatus();
+    ErrorStatus getStatus() const;
 
     /**
      * Retrieves the model that has been prepared for execution from the
-     * asynchronous task launched by IDevice::prepareModel. If
-     * IDevice::prepareModel has not finished asynchronously preparing the
+     * asynchronous task launched by IDevice::prepareModel*. If
+     * IDevice::prepareModel* has not finished asynchronously preparing the
      * model, this call will block until the asynchronous task notifies the
      * object.
      *
      * @return preparedModel Returned model that has been prepared for
-     *                       execution, nullptr if the model was unable to be
-     *                       prepared.
+     *     execution, nullptr if the model was unable to be prepared.
      */
-    sp<V1_0::IPreparedModel> getPreparedModel();
+    sp<V1_0::IPreparedModel> getPreparedModel() const;
 
   private:
-    ErrorStatus mErrorStatus;
+    mutable std::mutex mMutex;
+    mutable std::condition_variable mCondition;
+    bool mNotified GUARDED_BY(mMutex) = false;
+    ErrorStatus mErrorStatus = ErrorStatus::GENERAL_FAILURE;
     sp<V1_0::IPreparedModel> mPreparedModel;
 };
 
 /**
- * The ExecutionCallback class is used to receive the error status of the
- * execution from a task executing asynchronously with respect to the runtime.
- * If a calling thread calls wait* or get* on a PreparedModelCallback object and
- * the corresponding asynchronous task has not finished the execution, the
- * calling thread will block until the asynchronous task has either called notify
- * or notify_1_2. For more information on the synchronization behavior, refer to
- * the CallbackBase class.
+ * The ExecutionCallback class is used to receive the results of the execution
+ * from a task executing asynchronously with respect to the runtime. If a
+ * calling thread calls wait or get* on a ExecutionCallback object and the
+ * corresponding asynchronous task has not finished the execution, the calling
+ * thread will block until the asynchronous task has either called notify or
+ * notify_1_2.
  *
- * This class inherits the basic blocking and signaling calls from
- * CallbackBase, and implements the HIDL notify and notify_1_2 calls from
- * IExecutionCallback. This callback object is passed as an argument to
- * IPreparedModel::execute.
+ * If the callback object is notified more than once, only the results of the
+ * first call to notify* are used, and the results from subsequent calls are
+ * discarded.
+ *
+ * This callback object is passed as an argument to IPreparedModel::execute*.
  */
-class ExecutionCallback : public CallbackBase, public IExecutionCallback {
+class ExecutionCallback : public IExecutionCallback {
   public:
-    ExecutionCallback();
-    ~ExecutionCallback() override;
-
     /**
-     * IExecutionCallback::notify and IExecutionCallback::notify_1_2 mark the
-     * callback object with the return status of the asynchronous execution that
-     * held this callback and enable all prior and future wait* calls on the
-     * ExecutionCallback object to proceed. For more information on the
-     * synchronization behavior, refer to the CallbackBase class.
+     * IExecutionCallback::notify marks the callback object with the return
+     * status of the asynchronous execution that held this callback and enables
+     * all prior and future wait calls on the ExecutionCallback object to
+     * proceed.
      *
      * Either IExecutionCallback::notify or IExecutionCallback::notify_1_2 must
-     * be called exactly once on a given ExecutionCallback object.
+     * be called on a given ExecutionCallback object.
+     *
+     * If the callback object is notified more than once, only the results of
+     * the first call to notify* are used, and the results from subsequent calls
+     * are discarded.
      *
      * @param status Error status returned from launching the asynchronous task
-     *               (if the launch fails) or from the asynchronous task itself
-     *               (if the launch succeeds). Must be:
-     *               - NONE if the asynchronous execution was successful
-     *               - DEVICE_UNAVAILABLE if driver is offline or busy
-     *               - GENERAL_FAILURE if there is an unspecified error
-     *               - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is
-     *                 not large enough to store the resultant values
-     *               - INVALID_ARGUMENT if the input request is invalid
+     *     (if the launch fails) or from the asynchronous task itself (if the
+     *     launch succeeds). Must be:
+     *     - NONE if the asynchronous execution was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is not large
+     *         enough to store the resultant values
+     *     - INVALID_ARGUMENT if the input request is invalid
      */
     Return<void> notify(ErrorStatus status) override;
 
     /**
-     * Similar to IExecutionCallback::notify, but for V1_2::IPreparedModel to
-     * also notify output shapes along with error status.
+     * IExecutionCallback::notify_1_2 marks the callback object with the results
+     * (error status, dynamic output shapes, and timing information) of the
+     * asynchronous execution that held this callback and enables all prior and
+     * future wait calls on the ExecutionCallback object to proceed.
+     *
+     * Either IExecutionCallback::notify or IExecutionCallback::notify_1_2 must
+     * be called on a given ExecutionCallback object.
+     *
+     * If the callback object is notified more than once, only the results of
+     * the first call to notify* are used, and the results from subsequent calls
+     * are discarded.
      *
      * @param status Error status returned from launching the asynchronous task
-     *               (if the launch fails) or from the asynchronous task itself
-     *               (if the launch succeeds). Must be:
-     *               - NONE if the asynchronous execution was successful
-     *               - DEVICE_UNAVAILABLE if driver is offline or busy
-     *               - GENERAL_FAILURE if the asynchronous task resulted in an
-     *                 unspecified error
-     *               - OUTPUT_INSUFFICIENT_SIZE if at least one output
-     *                 operand buffer is not large enough to store the
-     *                 corresponding output
-     *               - INVALID_ARGUMENT if one of the input arguments to
-     *                 prepareModel is invalid
+     *     (if the launch fails) or from the asynchronous task itself (if the
+     *     launch succeeds). Must be:
+     *     - NONE if the asynchronous execution was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if the asynchronous task resulted in an unspecified
+     *         error
+     *     - OUTPUT_INSUFFICIENT_SIZE if at least one output operand buffer is
+     *         not large enough to store the corresponding output
+     *     - INVALID_ARGUMENT if one of the input arguments to prepareModel is
+     *         invalid
      * @param outputShapes A list of shape information of model output operands.
-     *                     The index into "outputShapes" corresponds to the index
-     *                     of the output operand in the Request outputs vector.
-     *                     outputShapes must be empty unless the status is either
-     *                     NONE or OUTPUT_INSUFFICIENT_SIZE.
-     * @return Timing Duration of execution. Unless MeasureTiming::YES was passed when
-     *                launching the execution and status is NONE, all times must
-     *                be reported as UINT64_MAX. A driver may choose to report
-     *                any time as UINT64_MAX, indicating that particular measurement is
-     *                not available.
+     *     The index into "outputShapes" corresponds to the index of the output
+     *     operand in the Request outputs vector. outputShapes must be empty
+     *     unless the status is either NONE or OUTPUT_INSUFFICIENT_SIZE.
+     * @param Timing Duration of execution. Unless MeasureTiming::YES was passed
+     *     when launching the execution and status is NONE, all times must be
+     *     reported as UINT64_MAX. A driver may choose to report any time as
+     *     UINT64_MAX, indicating that particular measurement is not available.
      */
     Return<void> notify_1_2(ErrorStatus status, const hidl_vec<OutputShape>& outputShapes,
                             const Timing& timing) override;
@@ -321,81 +240,87 @@
     }
 
     /**
+     * ExecutionCallback::wait blocks until notify* has been called on the
+     * callback object.
+     */
+    void wait() const;
+
+    /**
      * Retrieves the error status returned from the asynchronous task launched
      * by either IPreparedModel::execute or IPreparedModel::execute_1_2. If
      * IPreparedModel::execute or IPreparedModel::execute_1_2 has not finished
-     * asynchronously executing, this call will block until the asynchronous task
-     * notifies the object.
+     * asynchronously executing, this call will block until the asynchronous
+     * task notifies the object.
      *
      * @return status Error status returned from launching the asynchronous task
-     *                (if the launch fails) or from the asynchronous task itself
-     *                (if the launch succeeds). Must be:
-     *                - NONE if the asynchronous execution was successful
-     *                - DEVICE_UNAVAILABLE if driver is offline or busy
-     *                - GENERAL_FAILURE if the asynchronous task resulted in an
-     *                  unspecified error
-     *                - OUTPUT_INSUFFICIENT_SIZE if at least one output
-     *                  operand buffer is not large enough to store the
-     *                  corresponding output
-     *                - INVALID_ARGUMENT if one of the input arguments to
-     *                  prepareModel is invalid
+     *     (if the launch fails) or from the asynchronous task itself (if the
+     *     launch succeeds). Must be:
+     *     - NONE if the asynchronous execution was successful
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if the asynchronous task resulted in an unspecified
+     *         error
+     *     - OUTPUT_INSUFFICIENT_SIZE if at least one output operand buffer is
+     *         not large enough to store the corresponding output
+     *     - INVALID_ARGUMENT if one of the input arguments to prepareModel is
+     *         invalid
      */
-    ErrorStatus getStatus();
+    ErrorStatus getStatus() const;
 
     /**
      * Retrieves the output shapes returned from the asynchronous task launched
-     * by IPreparedModel::execute_1_2. If IPreparedModel::execute_1_2 has not finished
-     * asynchronously executing, this call will block until the asynchronous task
-     * notifies the object.
+     * by IPreparedModel::execute_1_2. If IPreparedModel::execute_1_2 has not
+     * finished asynchronously executing, this call will block until the
+     * asynchronous task notifies the object.
      *
-     * If the asynchronous task was launched by IPreparedModel::execute, an empty vector
-     * will be returned.
+     * If the asynchronous task was launched by IPreparedModel::execute, an
+     * empty vector will be returned.
      *
-     * @return outputShapes A list of shape information of model output operands.
-     *                      The index into "outputShapes" corresponds to the index
-     *                      of the output operand in the Request outputs vector.
-     *                      outputShapes must be empty unless the status is either
-     *                      NONE or OUTPUT_INSUFFICIENT_SIZE.
+     * @return outputShapes A list of shape information of model output
+     *     operands. The index into "outputShapes" corresponds to the index of
+     *     the output operand in the Request outputs vector. outputShapes must
+     *     be empty unless the status is either NONE or
+     *     OUTPUT_INSUFFICIENT_SIZE. outputShaps may be empty if the status is
+     *     NONE and all model output operands are fully-specified at execution
+     *     time. outputShapes must have the same number of elements as the
+     *     number of model output operands if the status is
+     *     OUTPUT_INSUFFICIENT_SIZE, or if the status is NONE and the model has
+     *     at least one output operand that is not fully-specified.
      */
-    const std::vector<OutputShape>& getOutputShapes();
+    const std::vector<OutputShape>& getOutputShapes() const;
 
     /**
-     * Retrieves the duration of execution ofthe asynchronous task launched
-     * by IPreparedModel::execute_1_2. If IPreparedModel::execute_1_2 has not finished
-     * asynchronously executing, this call will block until the asynchronous task
-     * notifies the object.
+     * Retrieves the duration of execution of the asynchronous task launched by
+     * IPreparedModel::execute_1_2. If IPreparedModel::execute_1_2 has not
+     * finished asynchronously executing, this call will block until the
+     * asynchronous task notifies the object.
      *
-     * If the asynchronous task was launched by IPreparedModel::execute, every time
-     * must be UINT64_MAX.
+     * If the asynchronous task was launched by IPreparedModel::execute, every
+     * time must be UINT64_MAX.
      *
-     * @return timing Duration of the execution. Every time must be UINT64_MAX unless
-     *                the status is NONE.
+     * @return timing Duration of the execution. Every time must be UINT64_MAX
+     *     unless the status is NONE.
      */
-    Timing getTiming();
+    Timing getTiming() const;
 
   private:
+    /*
+     * ExecutionCallback::notifyInternal stores the results of the execution
+     * (status, output shapes, and timing information) in the ExecutionCallback
+     * object before any call to wait or get* return. It then enables all prior
+     * and future wait calls on the ExecutionCallback object to proceed.
+     */
+    void notifyInternal(ErrorStatus errorStatus, const hidl_vec<OutputShape>& outputShapes,
+                        const Timing& timing);
+
+    // members
+    mutable std::mutex mMutex;
+    mutable std::condition_variable mCondition;
+    bool mNotified GUARDED_BY(mMutex) = false;
     ErrorStatus mErrorStatus = ErrorStatus::GENERAL_FAILURE;
     std::vector<OutputShape> mOutputShapes = {};
     Timing mTiming = {};
 };
 
-// template function implementation(s) below this point
-
-template <class Rep, class Period>
-std::cv_status CallbackBase::wait_for(const std::chrono::duration<Rep, Period>& timeout_duration) {
-    std::unique_lock<std::mutex> lock(mMutex);
-    std::cv_status status =
-            mCondition.wait_for(lock, timeout_duration, [this] { return mNotified; });
-    if (status != std::cv_status::timeout) {
-        join_thread_locked();
-    }
-    return status;
-}
-
-}  // namespace implementation
-}  // namespace V1_2
-}  // namespace neuralnetworks
-}  // namespace hardware
-}  // namespace android
+}  // namespace android::hardware::neuralnetworks::V1_2::implementation
 
 #endif  // ANDROID_HARDWARE_NEURALNETWORKS_V1_2_CALLBACKS_H