diff --git a/Common/Eval.cpp b/Common/Eval.cpp
index 3704f27c2..d29d04584 100644
--- a/Common/Eval.cpp
+++ b/Common/Eval.cpp
@@ -122,4 +122,4 @@ void Eval<ElemType>::ResetState()
 template class Eval<double>; 
 template class Eval<float>;
 
-}}}
\ No newline at end of file
+}}}
diff --git a/MachineLearning/CNTKComputationNetworkLib/CompositeComputationNodes.h b/MachineLearning/CNTKComputationNetworkLib/CompositeComputationNodes.h
index 27d9cab86..4d3aa2191 100644
--- a/MachineLearning/CNTKComputationNetworkLib/CompositeComputationNodes.h
+++ b/MachineLearning/CNTKComputationNetworkLib/CompositeComputationNodes.h
@@ -204,15 +204,19 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         typedef ComputationNodeNonLooping<ElemType> Base; UsingComputationNodeMembers;
     public:
         //virtual ComputationNodeBase * NewThis(DEVICEID_TYPE deviceId, const wstring & name) = 0;
-        PreComputedNode(DEVICEID_TYPE deviceId, const wstring & name) : Base(deviceId, name)
-        {
-            // further initializations
-            m_hasComputed = false;
-        }
+        PreComputedNode(DEVICEID_TYPE deviceId, const wstring & name) :
+            Base(deviceId, name),
+            m_hasComputed(false)
+        { }
 
         // interface through which this node is operated on are these two functions
+
+        // check whether node has already undergone precomputation
         virtual bool HasComputed() const { return m_hasComputed; }
-        virtual void MarkComputed(const bool hasComputed)       // override this for further finalizing operation
+
+        // call this with 'false' at start and with 'true' at end
+        // This is used for resetting and updating from accumulators.
+        virtual void MarkComputed(const bool hasComputed)
         {
             m_hasComputed = hasComputed;
         }
@@ -223,7 +227,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         {
             Base::SaveToFile(fstream);
             fstream << m_hasComputed;
-            fstream << m_functionValues;
+            fstream << m_functionValues;    // TODO: why serialize if not yet computed?
         }
 
         virtual void LoadFromFile(File& fstream, size_t modelVersion) override
@@ -271,34 +275,46 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         bool m_hasComputed;
     };
 
-#define UsingPreComputedNodeMembers UsingComputationNodeMembersBoilerplate; using Base::m_hasComputed
+#define UsingPreComputedNodeMembers UsingComputationNodeMembers; using Base::m_hasComputed
 
     // -----------------------------------------------------------------------
-    // MeanNode (features)
+    // MeanInvStdDevNodeBase (features)  -- common base class for Mean and InvStdDev
     // -----------------------------------------------------------------------
 
     template<class ElemType>
-    class MeanNode : public PreComputedNode<ElemType>, public NumInputs<1>
+    class MeanInvStdDevNodeBase : public PreComputedNode<ElemType>, public NumInputs<1>
     {
         typedef PreComputedNode<ElemType> Base; UsingPreComputedNodeMembers;
-        static const std::wstring TypeName() { return L"Mean"; }
+        //static const std::wstring TypeName() { return L"MeanInvStdDev (base)"; }
     public:
-        MeanNode(DEVICEID_TYPE deviceId, const wstring & name) :
+        MeanInvStdDevNodeBase(DEVICEID_TYPE deviceId, const wstring & name) :
             PreComputedNode<ElemType>(deviceId, name),
-            m_numSamples(0)
+            m_numSamples(SIZE_MAX)
         { }
 
         virtual void LoadFromFile(File& fstream, size_t modelVersion) override
         {
             Base::LoadFromFile(fstream, modelVersion);
-            m_numSamples = 0;   // TODO: intended? Not loaded from file?
+            m_numSamples = SIZE_MAX;
         }
 
         virtual void /*PreComputedNode::*/MarkComputed(const bool hasComputed)
         {
             Base::MarkComputed(hasComputed);
-            if (m_hasComputed)
+            if (!m_hasComputed)     // initialize
+            {
+                if (IsAccumulating())
+                    LogicError("%ls %ls operation: MarkComputed(false) has been called while accumulating.", NodeName().c_str(), OperationName().c_str());
                 m_numSamples = 0;
+            }
+            else                    // finalize
+            {
+                if (!IsAccumulating())
+                    LogicError("%ls %ls operation: MarkComputed(true) has been called without MarkComputed(false) first.", NodeName().c_str(), OperationName().c_str());
+                if (m_numSamples == 0)
+                    LogicError("%ls %ls operation: No data accumulated during precomputation.", NodeName().c_str(), OperationName().c_str());
+                m_numSamples = SIZE_MAX;
+            }
         }
 
         virtual void ComputeInputPartial(const size_t /*inputIndex*/)
@@ -306,51 +322,68 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             LogicError("Mean operation should not be involved in the gradient calculation.");
         }
 
-        virtual void /*ComputationNodeNonLooping::*/EvaluateThisNodeNonLooping() override
-        {
-            if (!m_hasComputed)
-            {
-                Matrix<ElemType> &samples = Inputs(0)->FunctionValues();
-                Matrix<ElemType> &avg = FunctionValues();
-#if 1//NANCHECK
-                samples.HasNan("Mean-Samples");
-#endif
-
-                size_t numNewSamples = samples.GetNumCols();
-                Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + samples.GetNumCols()), samples, false,
-                                                         ConstOnes(numNewSamples, 1, samples.GetDeviceId()),
-                                                         false, (ElemType)m_numSamples / (m_numSamples + numNewSamples), avg);
-
-#if 1//NANCHECK
-                avg.HasNan("Mean-avg");
-                //ones.HasNan("Mean-ones");
-#endif
-
-                m_numSamples += numNewSamples;
-            }
-        }
-
-        virtual void /*ComputationNodeBase::*/Validate(bool isFinalValidationPass) override
-        {
-            Base::Validate(isFinalValidationPass);
-            if (!m_hasComputed)
-            {
-                FunctionValues().SetValue(0);    // reset accumulator
-                fprintf(stderr, "Mean: SetValue(0)\n");
-            }
-        }
-
         virtual void CopyTo(const ComputationNodePtr nodeP, const std::wstring& newName, const CopyNodeFlags flags) const
         {
             Base::CopyTo(nodeP, newName, flags);
             if (flags & CopyNodeFlags::copyNodeValue)
             {
-                auto node = dynamic_pointer_cast<MeanNode<ElemType>>(nodeP);
-                node->m_numSamples = m_numSamples;
+                if (m_numSamples != SIZE_MAX)
+                    LogicError("%ls %ls operation: CopyTo() called while accumulating.", NodeName().c_str(), OperationName().c_str());
+                auto node = dynamic_pointer_cast<MeanInvStdDevNodeBase<ElemType>>(nodeP);
+                node->m_numSamples = SIZE_MAX;
             }
         }
-    private:
-        size_t m_numSamples;    // TODO: move to base class?
+    protected:
+        size_t m_numSamples;    // (SIZE_MAX while outside accumulation state)
+        bool IsAccumulating() const { return m_numSamples != SIZE_MAX; }
+    };
+
+    // -----------------------------------------------------------------------
+    // MeanNode (features)
+    // -----------------------------------------------------------------------
+
+    template<class ElemType>
+    class MeanNode : public MeanInvStdDevNodeBase<ElemType>
+    {
+        typedef MeanInvStdDevNodeBase<ElemType> Base; ComputationNodeBoilerplate; UsingPreComputedNodeMembers;
+        static const std::wstring TypeName() { return L"Mean"; }
+    public:
+        MeanNode(DEVICEID_TYPE deviceId, const wstring & name) :
+            Base(deviceId, name)
+        { }
+
+        virtual void /*PreComputedNode::*/MarkComputed(const bool hasComputed)
+        {
+            Base::MarkComputed(hasComputed);
+            if (!m_hasComputed)     // initialize accumulation
+                FunctionValues().SetValue(0);
+            // no else branch because EvaluateThisNodeNonLooping() already leaves a valid mean in m_functionValues
+        }
+
+        virtual void /*ComputationNodeNonLooping::*/EvaluateThisNodeNonLooping() override
+        {
+            if (m_hasComputed)
+                return;     // not accumulating
+
+            if (!IsAccumulating())
+                LogicError("%ls %ls operation: MarkComputed(false) has not been called.", NodeName().c_str(), OperationName().c_str());
+
+            Matrix<ElemType> &samples = Inputs(0)->FunctionValues();
+            Matrix<ElemType> &avg = FunctionValues();
+
+#if 1//NANCHECK
+            samples.HasNan("Mean-Samples");
+#endif
+            size_t numNewSamples = samples.GetNumCols();
+            Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + samples.GetNumCols()), samples, false,
+                                                        ConstOnes(numNewSamples, 1, samples.GetDeviceId()),
+                                                        false, (ElemType)m_numSamples / (m_numSamples + numNewSamples), avg);
+#if 1//NANCHECK
+            avg.HasNan("Mean-avg");
+#endif
+
+            m_numSamples += numNewSamples;
+        }
     };
 
     template class MeanNode<float>;
@@ -362,32 +395,34 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     // -----------------------------------------------------------------------
 
     template<class ElemType>
-    class InvStdDevNode : public PreComputedNode<ElemType>, public NumInputs<1>
+    class InvStdDevNode : public MeanInvStdDevNodeBase<ElemType>
     {
-        typedef PreComputedNode<ElemType> Base; UsingPreComputedNodeMembers;
+        typedef MeanInvStdDevNodeBase<ElemType> Base; ComputationNodeBoilerplate; UsingPreComputedNodeMembers;
         static const std::wstring TypeName() { return L"InvStdDev"; }
     public:
         InvStdDevNode(DEVICEID_TYPE deviceId, const wstring & name) :
-            PreComputedNode<ElemType>(deviceId, name),
-            m_mean(deviceId), m_var(deviceId), m_temp(deviceId),
-            m_numSamples(0)
+            Base(deviceId, name),
+            m_mean(deviceId), m_var(deviceId), m_temp(deviceId)
         { }
 
-        virtual void LoadFromFile(File& fstream, size_t modelVersion) override
-        {
-            Base::LoadFromFile(fstream, modelVersion);
-            m_numSamples = 0;   // TODO: intended? not loading from file?
-        }
-
         virtual void /*PreComputedNode::*/MarkComputed(const bool hasComputed) override
         {
             Base::MarkComputed(hasComputed);
 
-            if (m_hasComputed && m_numSamples > 0)  //m_numSamples>0 means it's not called from model loading
+            if (!m_hasComputed) // initialize
+            {
+                // reset accumulators
+                size_t inputDim = Inputs(0)->GetNumRows();
+                m_mean.Resize(inputDim, 1);
+                m_var.Resize(inputDim, 1);
+                m_mean.SetValue(0);
+                m_var.SetValue(0);
+                FunctionValues().SetValue(0);   // also set this because not doing it may flag during debugging; avoids special-casing this
+            }
+            else                // finalize
             {
                 ElemType sqrtFloor = 1e-10f;
-
-                m_var.InplaceTruncateBottom(sqrtFloor); //prevent too small variance (and negative square roots)
+                m_var.InplaceTruncateBottom(sqrtFloor);     // prevent too small variance (and negative square roots due to numeric inaccuracy)
 #if 1//NANCHECK
                 m_var.HasNan("MarkComputed-InplaceTruncateBottom");
 #endif
@@ -402,63 +437,43 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                 m_var.HasNan("MarkComputed-ElementInverse()");
 #endif
                 FunctionValues().SetValue(m_var);
-
-                m_numSamples = 0;
             }
         }
 
-        virtual void ComputeInputPartial(const size_t /*inputIndex*/) override
-        {
-            LogicError("InvStdDev operation should not be involved in the gradient calculation.");
-        }
-
         virtual void /*ComputationNodeNonLooping::*/EvaluateThisNodeNonLooping() override
         {
-            if (!m_hasComputed)
-            {
-                Matrix<ElemType> &samples = Inputs(0)->FunctionValues();
+            if (m_hasComputed)
+                return;     // not accumulating
+
+            if (!IsAccumulating())
+                LogicError("%ls %ls operation: MarkComputed(false) has not been called.", NodeName().c_str(), OperationName().c_str());
+
+            Matrix<ElemType> &samples = Inputs(0)->FunctionValues();
 #if 1//NANCHECK
-                samples.HasNan("InvStdDev-Samples");
+            samples.HasNan("InvStdDev-Samples");
 #endif
-                m_temp.SetValue(m_mean);
-                size_t numNewSample = samples.GetNumCols();
-                Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + numNewSample), samples, false,
-                                                         ConstOnes(numNewSample, 1, samples.GetDeviceId()),
-                                                         false, (ElemType)m_numSamples / (m_numSamples + numNewSample), m_mean);
+            m_temp.SetValue(m_mean);
+            size_t numNewSample = samples.GetNumCols();
+            Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + numNewSample), samples, false,
+                                                        ConstOnes(numNewSample, 1, samples.GetDeviceId()),
+                                                        false, (ElemType)m_numSamples / (m_numSamples + numNewSample), m_mean);
 
-                m_temp -= m_mean;
-                m_temp.AssignElementPowerOf(m_temp, 2);
-                m_var += m_temp;
+            m_temp -= m_mean;
+            m_temp.AssignElementPowerOf(m_temp, 2);
+            m_var += m_temp;
 
-                m_temp.AssignDifferenceOf(samples, m_mean);
-                m_temp.AssignElementPowerOf(m_temp, 2);
+            m_temp.AssignDifferenceOf(samples, m_mean);
+            m_temp.AssignElementPowerOf(m_temp, 2);
 
-                Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + numNewSample), m_temp, false,
-                                                         ConstOnes(numNewSample, 1, samples.GetDeviceId()),
-                                                         false, (ElemType)m_numSamples / (m_numSamples + numNewSample), m_var);
+            Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / (m_numSamples + numNewSample), m_temp, false,
+                                                        ConstOnes(numNewSample, 1, samples.GetDeviceId()),
+                                                        false, (ElemType)m_numSamples / (m_numSamples + numNewSample), m_var);
 
 #if 1//NANCHECK
-                m_var.HasNan("InvStdDev-m_var");
+            m_var.HasNan("InvStdDev-m_var");
 #endif
 
-                m_numSamples += samples.GetNumCols();
-            }
-        }
-
-        virtual void /*ComputationNodeBase::*/Validate(bool isFinalValidationPass) override
-        {
-            Base::Validate(isFinalValidationPass);
-
-            if (!m_hasComputed)
-            {
-                size_t inputDim = Inputs(0)->GetNumRows();
-                m_mean.Resize(inputDim, 1);
-                m_var.Resize(inputDim, 1);
-                // reset accumulators
-                m_mean.SetValue(0);
-                m_var.SetValue(0);
-                fprintf(stderr, "InvStdDev: SetValue(0)\n");
-            }
+            m_numSamples += samples.GetNumCols();
         }
 
         virtual void MoveMatricesToDevice(const DEVICEID_TYPE deviceId) override
@@ -475,15 +490,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             if (flags & CopyNodeFlags::copyNodeValue)
             {
                 auto node = dynamic_pointer_cast<InvStdDevNode<ElemType>>(nodeP);
-                node->m_numSamples = m_numSamples;
-
                 node->m_mean = m_mean;
                 node->m_var = m_var;
-                node-> m_temp =  m_temp;
+                node->m_temp =  m_temp;
             }
         }
     private:
-        size_t m_numSamples;
         Matrix<ElemType> m_mean;
         Matrix<ElemType> m_var;
         Matrix<ElemType>  m_temp;
diff --git a/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp b/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp
index 92985017b..efb6eea51 100644
--- a/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp
+++ b/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp
@@ -430,7 +430,15 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         }
     }
 
-    // prepares the network for computation
+    bool ComputationNetwork::BuiltAndValidatedSubNetwork(const ComputationNodeBasePtr & rootNode)
+    {
+        return m_built.find(rootNode) != m_built.end();
+    }
+
+    // prepare to compute with the subnetwork that this rootNode depends on, including
+    //  - auto-detecting recurrent loops
+    //  - collect input and learnable nodes
+    //  - calling Validate() on all nodes lazily, which sizes all matrices (column dimensions get updated to MB size)
     // Done lazily, called for every minibatch's invocation of EvaluateNode(), but memoizing which nodes were done already.
     // BUGBUG? Lazy triggers on the root node. I.e. for two different root nodes (training, eval), it validates twice.
     void ComputationNetwork::BuildAndValidateSubNetwork(const ComputationNodeBasePtr rootNode)
@@ -558,53 +566,53 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     }
 #endif
 
-    template<class N> void ComputationNetwork::GetNodesRequiringX(std::list<ComputationNodeBasePtr> & nodesRequirePreComputation, const ComputationNodeBasePtr rootNode, bool checkComputed)
+    template<class N> void ComputationNetwork::GetNodesRequiringX(std::list<ComputationNodeBasePtr> & nodesRequiringX, const ComputationNodeBasePtr rootNode, bool checkComputed)
     {
-        if (rootNode == nullptr)        // find nodes from all available nodes
+        if (!rootNode)              // find nodes from all available nodes
         {
-            for (auto nodeIter = m_nameToNodeMap.begin(); nodeIter != m_nameToNodeMap.end(); nodeIter++)
+            for (const auto & nodep : m_nameToNodeMap)
             {
-                ComputationNodeBasePtr node = nodeIter->second;
-                if (node->RequiresPreCompute()) // TODO: why not check directly for the type with a dynamic_cast?
+                auto node = dynamic_pointer_cast<N>(nodep.second);
+                if (node)
                 {
-                    auto preComputedNode = static_pointer_cast<N>(node);
-                    if (!checkComputed || !preComputedNode->HasComputed())
-                        nodesRequirePreComputation.push_back(node);
+                    assert(node->RequiresPreCompute());
+                    if (!checkComputed || !node->HasComputed())
+                        nodesRequiringX.push_back(node);
                 }
             }
         }
         else                            // or for calculating a specific node
         {
-            const auto & nodes = GetEvalOrder(rootNode, false);
-            for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
+            for (const auto & nodei : GetEvalOrder(rootNode, false))
             {
-                ComputationNodeBasePtr node = *nodeIter;
-                if (node->RequiresPreCompute()) // TODO: why not check directly for the type with a dynamic_cast?
+                auto node = dynamic_pointer_cast<N>(nodei);
+                if (node)
                 {
-                    auto preComputedNode = static_pointer_cast<N>(node);
-                    if (!checkComputed || !preComputedNode->HasComputed())
-                        nodesRequirePreComputation.push_back(node);
+                    assert(node->RequiresPreCompute());
+                    if (!checkComputed || !node->HasComputed())
+                        nodesRequiringX.push_back(node);
                 }
             }
         }
+        nodesRequiringX.unique();
     }
 
     //return list of nodes that require precomputation and not precomputed yet.
     std::list<ComputationNodeBasePtr> ComputationNetwork::GetNodesRequiringPreComputation(const ComputationNodeBasePtr rootNode, bool checkComputed)
     {
-        std::list<ComputationNodeBasePtr> nodesRequirePreComputation;
-        GetNodesRequiringX<PreComputedNode<float>>(nodesRequirePreComputation, rootNode, checkComputed);
-        GetNodesRequiringX<PreComputedNode<double>>(nodesRequirePreComputation, rootNode, checkComputed);
-        return nodesRequirePreComputation;
+        std::list<ComputationNodeBasePtr> nodesRequiringX;
+        GetNodesRequiringX<PreComputedNode<float>>(nodesRequiringX, rootNode, checkComputed);
+        GetNodesRequiringX<PreComputedNode<double>>(nodesRequiringX, rootNode, checkComputed);
+        return nodesRequiringX;
     }
 
     //return list of nodes that require batch mode and not precomputed yet.
     std::list<ComputationNodeBasePtr> ComputationNetwork::GetNodesRequiringBatchMode(const ComputationNodeBasePtr rootNode, bool checkComputed)
     {
-        std::list<ComputationNodeBasePtr> nodesRequirePreComputation;
-        GetNodesRequiringX<BatchModeNode<float>>(nodesRequirePreComputation, rootNode, checkComputed);
-        GetNodesRequiringX<BatchModeNode<double>>(nodesRequirePreComputation, rootNode, checkComputed);
-        return nodesRequirePreComputation;
+        std::list<ComputationNodeBasePtr> nodesRequiringX;
+        GetNodesRequiringX<BatchModeNode<float>>(nodesRequiringX, rootNode, checkComputed);
+        GetNodesRequiringX<BatchModeNode<double>>(nodesRequiringX, rootNode, checkComputed);
+        return nodesRequiringX;
     }
 
     // The methods below determine evaluation order, which is tricky in presence of recurrent loops.
diff --git a/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.h b/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.h
index 5bb60da78..0cca9c821 100644
--- a/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.h
+++ b/MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.h
@@ -565,13 +565,12 @@ public:
     //  - these must be executed frame by frame rather than as a map
     //  - such a loop is treated as if they were a little nested network; this is done inside here
     //  - these little nested networks are defined in m_recurrentInfo[]
-    void Evaluate(const ComputationNodeBasePtr rootNode)
+    void Evaluate(const ComputationNodeBasePtr & rootNode)
     {
-        // prepare to compute with the subnetwork that this rootNode depends on, including
-        //  - auto-detecting recurrent loops
-        //  - collect input and learnable nodes
-        //  - calling Validate() on all nodes lazily, which sizes all matrices (column dimensions get updated to MB size)
-        BuildAndValidateSubNetwork(rootNode);
+        // caller must call BuildAndValidateSubNetwork() before
+        // TODO: Some places are hard to fix, e.g. encoder-decoder best-path functions. Those may be broken; this message will tell you.
+        if (!BuiltAndValidatedSubNetwork(rootNode))
+            LogicError("Evaluate for node %ls %ls: BuildAndValidateSubNetwork() has not been called on this node.");
 
         // determines order of evaluation, such that children get evaluated before their parent nodes
         std::list<ComputationNodeBasePtr>& allNodes = GetEvalOrder(rootNode, false);
@@ -692,6 +691,12 @@ public:
             }
         }
     }
+    template<class NODESET>
+    void Evaluate(const NODESET & nodes)
+    {
+        for (auto & node : nodes)
+            Evaluate(node);
+    }
 
     // propagate the features' MB size to all nodes of the network
     // TODO: This function should go. Resizing is now part of Validate() and EvaluateThisNode().
@@ -1233,6 +1238,24 @@ private:
 public:
     // prepares the network for computation
     void BuildAndValidateSubNetwork(const ComputationNodeBasePtr rootNode);
+    // and for a set of nodes
+    void StartEvaluateMinibatchLoop(const ComputationNodeBasePtr & rootNode)  // (ugly name; meant to be unique so we can rename if needed)
+    {
+        BuildAndValidateSubNetwork(rootNode);
+    }
+    template<class NODESET>
+    void StartEvaluateMinibatchLoop(const NODESET & nodes)  // (ugly name; meant to be unique so we can rename if needed)
+    {
+        for (auto & node : nodes)
+            StartEvaluateMinibatchLoop(node);
+    }
+    template<class NODESET>
+    void StartEvaluateMinibatchLoop(const NODESET & nodes1, const NODESET & nodes2) // often needed for two sets (training & evaluation criteria)
+    {
+        StartEvaluateMinibatchLoop(nodes1);
+        StartEvaluateMinibatchLoop(nodes2);
+    }
+    bool BuiltAndValidatedSubNetwork(const ComputationNodeBasePtr & rootNode);
 
     //this function will need to be called before actual validation and execution to 
     //predetermine how to share matrices to reduce memory usage.
diff --git a/MachineLearning/CNTKComputationNetworkLib/ComputationNode.h b/MachineLearning/CNTKComputationNetworkLib/ComputationNode.h
index 265baea67..885cb5b7b 100644
--- a/MachineLearning/CNTKComputationNetworkLib/ComputationNode.h
+++ b/MachineLearning/CNTKComputationNetworkLib/ComputationNode.h
@@ -503,9 +503,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
         virtual void /*IComputationNode::*/OnEvaluateBeginIteration()             // called before first iteration step of EvaluateThisNode()
         {
-            fprintf(stderr, "Trace: %ls %ls operation\n", NodeName().c_str(), OperationName().c_str());
+            fprintf(stderr, "OnEvaluateBeginIteration: %ls %ls operation\n", NodeName().c_str(), OperationName().c_str());
+        }
+        virtual void /*IComputationNode::*/OnEvaluateEndIteration()               // called after last iteration step of EvaluateThisNode()
+        {
+            fprintf(stderr, "OnEvaluateEndIteration: %ls %ls operation\n", NodeName().c_str(), OperationName().c_str());
         }
-        virtual void /*IComputationNode::*/OnEvaluateEndIteration() { }               // called after last iteration step of EvaluateThisNode()
 
     protected:
 
@@ -823,6 +826,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     template<class ElemType>
     class ComputationNode : public ComputationNodeBase // abstract class that cannot be instantiated
     {
+        typedef ComputationNodeBase Base;
     protected:
         //std containers such as list and map does not support class reference so we need to use pointer
         typedef shared_ptr<ComputationNode<ElemType>> ComputationNodePtr;
@@ -1249,6 +1253,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 #ifdef _DEBUG
         virtual void /*IComputationNode::*/OnEvaluateEndIteration()               // called after last iteration step of EvaluateThisNode()
         {
+            Base::OnEvaluateEndIteration();
             MaskMissingValuesColumnsToZero();
             if (m_functionValues.HasNan("OnEvaluateEndIteration"))
                 LogicError("%ls %ls operation unexpectedly produced NaN values.", NodeName().c_str(), OperationName().c_str());
@@ -1487,11 +1492,13 @@ public: \
     using Base::SaveToFile; using Base::UpdateFunctionAndGradientMBSize; using Base::SetInput; \
     using Base::Validate; using Base::ValidateUnaryMap; using Base::ValidateBinaryZip; using Base::ValidateUnaryReduce; using Base::ValidateBinaryReduce; using Base::ValidateInferBinaryChildren; using Base::ValidateInferInputSize
 
-#define UsingComputationNodeMembersBoilerplate \
+#define ComputationNodeBoilerplate \
 protected:    /* some boilerplate goes here */ \
     virtual const std::wstring OperationName() const override { return TypeName(); } \
-    virtual ComputationNodeBase * NewThis(DEVICEID_TYPE deviceId, const wstring & name) override { return new typename std::remove_reference<decltype(*this)>::type(deviceId, name); } \
-    UsingComputationNodeMembers
+    virtual ComputationNodeBase * NewThis(DEVICEID_TYPE deviceId, const wstring & name) override { return new typename std::remove_reference<decltype(*this)>::type(deviceId, name); }
+
+#define UsingComputationNodeMembersBoilerplate \
+    ComputationNodeBoilerplate; UsingComputationNodeMembers
 
 #pragma endregion base computation class
 
diff --git a/MachineLearning/CNTKEval/CNTKEvalTest/CNTKEvalTest.cpp b/MachineLearning/CNTKEval/CNTKEvalTest/CNTKEvalTest.cpp
index f7983cf30..9f236e11c 100644
--- a/MachineLearning/CNTKEval/CNTKEvalTest/CNTKEvalTest.cpp
+++ b/MachineLearning/CNTKEval/CNTKEvalTest/CNTKEvalTest.cpp
@@ -49,6 +49,7 @@ void DoCommand(const ConfigParameters& configRoot)
     DataReader<ElemType>* dataReader = new DataReader<ElemType>(readerConfig);
     eval.LoadModel(modelPath);
     dataReader->StartMinibatchLoop(mbSize, 0, epochSize);
+    eval.StartEvaluateMinibatchLoop(outputName);
     while (dataReader->GetMinibatch(inputMatrices))
     {
         void* data = (void*)arr->data();
diff --git a/MachineLearning/CNTKSGDLib/SGD.cpp b/MachineLearning/CNTKSGDLib/SGD.cpp
index b9aa6b3d5..9964f2a89 100644
--- a/MachineLearning/CNTKSGDLib/SGD.cpp
+++ b/MachineLearning/CNTKSGDLib/SGD.cpp
@@ -1145,7 +1145,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                     std::vector<ComputationNodeBasePtr> & labelNodes,
                     std::map<std::wstring, Matrix<ElemType>*>* inputMatrices)
     {
-        std::list<ComputationNodeBasePtr> nodes = net.GetNodesRequiringPreComputation();
+        std::list<ComputationNodeBasePtr> nodes = net.GetNodesRequiringPreComputation();    // this tests all HasComputed() flags
 
         if (nodes.size() == 0)
         {
@@ -1169,39 +1169,28 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             trainSetDataReader->StartMinibatchLoop(m_mbSize[0], 0);
         else                                    // using only one epoch
             trainSetDataReader->StartMinibatchLoop(m_mbSize[0], 0, m_epochSize);
-#if 1
-        size_t actualMBSize;
-        while (DataReaderHelpers::GetMinibatchIntoNetwork(*trainSetDataReader, net, nullptr, false, false, *inputMatrices, actualMBSize))
+        net.StartEvaluateMinibatchLoop(nodes);
+
+        // initialize
+        for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
+        {
+            auto node = static_pointer_cast<PreComputedNode<ElemType>>(*nodeIter);
+            node->MarkComputed(false/*begin accumulating*/);
+        }
+        size_t actualMBSizeDummy;
+        while (DataReaderHelpers::GetMinibatchIntoNetwork(*trainSetDataReader, net, nullptr, false, false, *inputMatrices, actualMBSizeDummy))
         {
             // TODO: move these into GetMinibatchIntoNetwork()  --but those are passed around; necessary? Can't we get them from 'net'?
             ComputationNetwork::UpdateEvalTimeStamps(featureNodes);
             ComputationNetwork::UpdateEvalTimeStamps(labelNodes);
 
-            for (auto & node : nodes)   // this loops over all pertinent PreComputeNodes
-                net.Evaluate(node);
+            net.Evaluate(nodes);
         }
-#else
-        while (trainSetDataReader->GetMinibatch(*inputMatrices))
-        {
-            // TODO: use GetMinibatchIntoNetwork(), should be easy
-            ComputationNetwork::UpdateEvalTimeStamps(featureNodes);
-            ComputationNetwork::UpdateEvalTimeStamps(labelNodes);
-
-            net.SetActualMiniBatchSizeFromFeatures();
-            trainSetDataReader->CopyMBLayoutTo(net.GetMBLayoutPtr());
-            net.VerifyActualNumParallelSequences(trainSetDataReader->GetNumParallelSequences());
-
-            // TODO: Exactly this loop should be INSIDE ComputationNetwork--pass the nodes array instead!
-            for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
-                net.Evaluate(*nodeIter);
-        }
-#endif
-
-        // mark done
+        // finalize
         for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
         {
             auto node = static_pointer_cast<PreComputedNode<ElemType>>(*nodeIter);
-            node->MarkComputed(true);
+            node->MarkComputed(true/*done accumulating*/);
         }
 
         return true;
@@ -1714,13 +1703,11 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
         int numMBsRun = 0;
 
-        size_t numEvalNodes = epochEvalErrors.size();
-
         // NOTE: the following two local matrices are not used in distGradAgg path
         // assume only one training criterion node for each epoch.
         // The criterion values are accumulated here over the minibatches (without having to pull them off the GPU).
         Matrix<ElemType> localEpochCriterion(1, 1, net.GetDeviceId());
-        Matrix<ElemType> localEpochEvalErrors(1, numEvalNodes, net.GetDeviceId());
+        Matrix<ElemType> localEpochEvalErrors(1, epochEvalErrors.size(), net.GetDeviceId());
 
         localEpochCriterion.SetValue(0);
         localEpochEvalErrors.SetValue(0);
@@ -1740,7 +1727,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (useGradientAggregation)
         {
             epochCriterion = double(0.0);
-            epochEvalErrors.assign(numEvalNodes, double(0.0));
+            epochEvalErrors.assign(epochEvalErrors.size(), double(0.0));
         }
 
         Profiler profiler(m_numMBsToCUDAProfile);
@@ -1752,13 +1739,13 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                                        m_enableDistributedMBReading &&
                                        trainSetDataReader->SupportsDistributedMBRead();
         if (useDistributedMBReading)
-        {
             trainSetDataReader->StartDistributedMinibatchLoop(tunedMBSize, epochNumber, g_mpi->CurrentNodeRank(), g_mpi->NumNodesInUse(), m_epochSize);
-        }
         else
-        {
             trainSetDataReader->StartMinibatchLoop(tunedMBSize, epochNumber, m_epochSize);
-        }
+        net.StartEvaluateMinibatchLoop(evaluationNodes);
+        net.StartEvaluateMinibatchLoop(criterionNodes);
+        if (m_needAdaptRegularization && m_adaptationRegType == AdaptationRegType::KL && refNode)
+            refNet.StartEvaluateMinibatchLoop(refNode);
 
         // TODO: what is this??
         AttemptUtteranceDerivativeFeatures(net, trainSetDataReader, featureNodes, inputMatrices);
@@ -1829,10 +1816,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
                 //compute eval node first since when gradient is computed the forward function values
                 //may be changed and need to be recomputed when gradient and function value share the same matrix
-                for (size_t i = 0; i < numEvalNodes; i++)
-                {
-                    net.Evaluate(evaluationNodes[i]);
-                }
+                net.Evaluate(evaluationNodes);
 
                 // only compute gradient when learning rate is large enough
                 if (learnRatePerSample > m_minLearnRate * 0.01)
@@ -1872,7 +1856,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                     // criteria are in FunctionValues()(0,0), we accumulate into another 1x1 Matrix (to avoid having to pull the values off the GPU)
                     Matrix<ElemType>::AddElementToElement(dynamic_pointer_cast<ComputationNode<ElemType>>(criterionNodes[0])->FunctionValues(),
                                                           0, 0, localEpochCriterion, 0, 0);
-                    for (size_t i = 0; i < numEvalNodes; i++)
+                    for (size_t i = 0; i < evaluationNodes.size(); i++)
                     {
                         Matrix<ElemType>::AddElementToElement(dynamic_pointer_cast<ComputationNode<ElemType>>(evaluationNodes[i])->FunctionValues(),
                                                               0, 0, localEpochEvalErrors, 0, i);
@@ -1882,14 +1866,14 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             else
             {
                 //distributed gradient aggregation
-                LazyInitDistGradAgg(learnableNodes, numEvalNodes, m_traceLevel);
+                LazyInitDistGradAgg(learnableNodes, evaluationNodes.size(), m_traceLevel);
 
                 //prepare the header
-                m_gradHeader->numEvalNode = numEvalNodes;
+                m_gradHeader->numEvalNode = evaluationNodes.size();
                 m_gradHeader->numSamples = actualMBSize;
                 m_gradHeader->numSamplesWithLabel = numSamplesWithLabel;
                 m_gradHeader->criterion = actualMBSize > 0 ? criterionNodes[0]->Get00Element() : 0.0;
-                for (size_t i = 0; i < numEvalNodes; i++)
+                for (size_t i = 0; i < evaluationNodes.size(); i++)
                     m_gradHeader->evalErrors[i] = actualMBSize > 0 ? evaluationNodes[i]->Get00Element() : 0.0;
 
                 m_distGradAgg->AggregateGradients(m_gradHeader, epochNumber);
@@ -1897,7 +1881,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                 aggregateNumSamples = m_gradHeader->numSamples;
                 aggregateNumSamplesWithLabel = m_gradHeader->numSamplesWithLabel;
                 epochCriterion += m_gradHeader->criterion;
-                for (size_t i = 0; i<numEvalNodes; i++)
+                for (size_t i = 0; i<epochEvalErrors.size(); i++)
                     epochEvalErrors[i] += m_gradHeader->evalErrors[i];
             }
 
@@ -1963,7 +1947,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                     {
                         timer.Restart();
                         epochCriterion = localEpochCriterion.Get00Element();
-                        for (size_t i = 0; i < numEvalNodes; i++)
+                        for (size_t i = 0; i < epochEvalErrors.size(); i++)
                             epochEvalErrors[i] = localEpochEvalErrors(0, i);
                         timer.Stop();
 
@@ -1991,7 +1975,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                         m_maxComputedEpochSize = numMBsRun * numSamplesLastMBs / m_numMBsToShowResult;
                     }
 
-                    for (size_t i = 0; i < numEvalNodes; i++)
+                    for (size_t i = 0; i < epochEvalErrors.size(); i++)
                     {
                         double evalError = (epochEvalErrors[i] - epochEvalErrorsLastMBs[i]) / numSamplesLastMBs;
                         string formatString = "EvalErr[%lu]PerSample = " + GeneratePaddedFloatOrExpFormat(0, 8, evalError) + "; ";
@@ -2012,7 +1996,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                     numSamplesLastMBs = 0;
 
                     epochCriterionLastMBs = epochCriterion;
-                    for (size_t i = 0; i < numEvalNodes; i++)
+                    for (size_t i = 0; i < epochEvalErrorsLastMBs.size(); i++)
                         epochEvalErrorsLastMBs[i] = epochEvalErrors[i];
 
                     if (std::isnan(epochCriterion))
@@ -2057,7 +2041,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         {
             // with parallelization, we have them in regular variables
             epochCriterion /= float(totalEpochSamples);
-            for (size_t i = 0; i< numEvalNodes; i++)
+            for (size_t i = 0; i< epochEvalErrors.size(); i++)
                 epochEvalErrors[i] /= totalEpochSamples;
         }
         else
@@ -2067,7 +2051,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             localEpochEvalErrors /= float(totalEpochSamples);
 
             epochCriterion = localEpochCriterion.Get00Element();
-            for (size_t i = 0; i < numEvalNodes; i++)
+            for (size_t i = 0; i < epochEvalErrors.size(); i++)
                 epochEvalErrors[i] = localEpochEvalErrors(0, i);
         }
 
@@ -2495,6 +2479,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
 #define EPSILON 1e-5
 
+    // this probes the automatic gradient computation with random inputs
     template<class ElemType>
     bool SGD<ElemType>::GradientCheck(ComputationNetwork& net,
                        const std::vector<ComputationNodeBasePtr> & criterionNodes,
@@ -2503,6 +2488,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     {
         vector<string> errMsgs;
 
+        net.StartEvaluateMinibatchLoop(criterionNodes[npos]);
+
         // gradient checking
         for (auto nodeIter = learnableNodes.begin(); nodeIter != learnableNodes.end(); nodeIter++)
         {
@@ -2524,7 +2511,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
                 node->UpdateEvalTimeStamp();
 
-                // use only the first criterion. Is
                 net.ComputeGradient<ElemType>(criterionNodes[npos]);
 
                 if (node->GradientValues().GetMatrixType() == MatrixType::SPARSE)
diff --git a/MachineLearning/CNTKSGDLib/SimpleEvaluator.h b/MachineLearning/CNTKSGDLib/SimpleEvaluator.h
index 666c7dbe7..142e0a327 100644
--- a/MachineLearning/CNTKSGDLib/SimpleEvaluator.h
+++ b/MachineLearning/CNTKSGDLib/SimpleEvaluator.h
@@ -122,6 +122,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                 evalResultsLastMBs.push_back((ElemType)0);
 
             dataReader->StartMinibatchLoop(mbSize, 0, testSize);
+            m_net.StartEvaluateMinibatchLoop(evalNodes);
 
             while (DataReaderHelpers::GetMinibatchIntoNetwork(*dataReader, m_net, nullptr, false, false, inputMatrices, actualMBSize))
             {
@@ -191,6 +192,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         }
 
         //returns error rate
+        // TODO: What does this function do?
         double EvaluateUnroll(IDataReader<ElemType>* dataReader, const size_t mbSize, double &evalSetCrossEntropy, const wchar_t* output = nullptr, const size_t testSize = requestDataSize)
         {
             std::vector<ComputationNodeBasePtr> & featureNodes = m_net.FeatureNodes();
@@ -211,6 +213,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             inputMatrices[L"numberobs"] = new Matrix<ElemType>(1, 1, m_net.GetDeviceId());
 
             dataReader->StartMinibatchLoop(mbSize, 0, testSize);
+            m_net.StartEvaluateMinibatchLoop(criterionNodes, evaluationNodes);
 
             double epochEvalError = 0;
             double epochCrossEntropy = 0;
@@ -415,9 +418,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             double evalResultsLastMBs = (double)0;
 
             for (auto ptr = dataReaders.begin(); ptr != dataReaders.end(); ptr++)
-            {
                 (*ptr)->StartMinibatchLoop(mbSize, 0, testSize);
-            }
+            // BUGBUG: Code below will fail because we now must call StartMinibatchLoop(), but I can't tell from below which nodes to call it for.
+            //for (auto & ptr : nets)
+            //    ptr->StartMinibatchLoop(xxx);
 
             bool bContinueDecoding = true;
             while (bContinueDecoding)
@@ -743,7 +747,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             }
         }
 
-        //return true if precomputation is executed.
+        // (only called by FindBestPath...())
         void ResetPreCompute()
         {
             //mark false
@@ -767,6 +771,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
             ComputationNetwork::UpdateEvalTimeStamps(featureNodes);
 
+            net.StartEvaluateMinibatchLoop(batchComputeNodes);  // TODO: Is this correct? There is no StartMinibatchLoop() for a reader.
+
             net.SetActualMiniBatchSizeFromFeatures();
             for (auto nodeIter = batchComputeNodes.begin(); nodeIter != batchComputeNodes.end(); nodeIter++)
                 net.Evaluate(*nodeIter);
diff --git a/MachineLearning/CNTKSGDLib/SimpleOutputWriter.h b/MachineLearning/CNTKSGDLib/SimpleOutputWriter.h
index 8bdb18041..c60e1ef2b 100644
--- a/MachineLearning/CNTKSGDLib/SimpleOutputWriter.h
+++ b/MachineLearning/CNTKSGDLib/SimpleOutputWriter.h
@@ -61,10 +61,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             //Matrix<ElemType> endOfFile =  Matrix<ElemType>((size_t)1,(size_t)1);
             //endOfFile(0,0)=0;
 
-            //evaluate with minibatches
+            // evaluate with minibatches
             dataReader.StartMinibatchLoop(mbSize, 0, numOutputSamples);
             dataReader.SetNumParallelSequences(1);
 
+            m_net.StartEvaluateMinibatchLoop(outputNodes);
+
             size_t totalEpochSamples = 0;
             std::map<std::wstring, void *, nocase_compare> outputMatrices;
 
@@ -107,7 +109,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             //clean up
             
         }
-        
 
         void WriteOutput(IDataReader<ElemType>& dataReader, size_t mbSize, std::wstring outputPath, const std::vector<std::wstring>& outputNodeNames, size_t numOutputSamples=requestDataSize)
         {
@@ -142,10 +143,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             std::map<std::wstring, Matrix<ElemType>*> inputMatrices;
             for (size_t i=0; i<featureNodes.size(); i++)
                 inputMatrices[featureNodes[i]->NodeName()] = &dynamic_pointer_cast<ComputationNode<ElemType>>(featureNodes[i])->FunctionValues();
-                        
-            //evaluate with minibatches
+
+            // evaluate with minibatches
             dataReader.StartMinibatchLoop(mbSize, 0, numOutputSamples);
 
+            m_net.StartEvaluateMinibatchLoop(outputNodes);
+
             size_t totalEpochSamples = 0;
             size_t numMBsRun = 0;
             size_t tempArraySize = 0;