diff --git a/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/FilterMutectCalls.java b/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/FilterMutectCalls.java
index 6c6bae4871b..a59512e5948 100644
--- a/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/FilterMutectCalls.java
+++ b/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/FilterMutectCalls.java
@@ -104,7 +104,7 @@ public final class FilterMutectCalls extends MultiplePassVariantWalker {
     private static final int NUMBER_OF_LEARNING_PASSES = 2;
 
     @Override
-    protected int numberOfPasses() { return NUMBER_OF_LEARNING_PASSES + 1; }
+    protected int numberOfPasses() { return NUMBER_OF_LEARNING_PASSES + 2; }    // {@code NUMBER_OF_LEARNING_PASSES} passes for learning, one for the threshold, and one for calling
 
     @Override
     public boolean requiresReference() { return true;}
@@ -142,9 +142,9 @@ protected void nthPassApply(final VariantContext variant,
                                 final FeatureContext featureContext,
                                 final int n) {
         ParamUtils.isPositiveOrZero(n, "Passes must start at the 0th pass.");
-        if (n < NUMBER_OF_LEARNING_PASSES) {
+        if (n <= NUMBER_OF_LEARNING_PASSES) {
             filteringEngine.accumulateData(variant, referenceContext);
-        } else if (n == NUMBER_OF_LEARNING_PASSES) {
+        } else if (n == NUMBER_OF_LEARNING_PASSES + 1) {
             vcfWriter.add(filteringEngine.applyFiltersAndAccumulateOutputStats(variant, referenceContext));
         } else {
             throw new GATKException.ShouldNeverReachHereException("This walker should never reach (zero-indexed) pass " + n);
@@ -156,6 +156,10 @@ protected void afterNthPass(final int n) {
         if (n < NUMBER_OF_LEARNING_PASSES) {
             filteringEngine.learnParameters();
         } else if (n == NUMBER_OF_LEARNING_PASSES) {
+            // it's important for filter parameters to stay the same and only learn the threshold in the final pass so that the
+            // final threshold used corresponds exactly to the filters
+            filteringEngine.learnThreshold();
+        }else if (n == NUMBER_OF_LEARNING_PASSES + 1) {
             final Path filteringStats = IOUtils.getPath(
                 filteringStatsOutput != null ? filteringStatsOutput
                     : outputVcf + FILTERING_STATS_EXTENSION);
diff --git a/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/Mutect2FilteringEngine.java b/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/Mutect2FilteringEngine.java
index c01324039e4..3a977b981b4 100644
--- a/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/Mutect2FilteringEngine.java
+++ b/src/main/java/org/broadinstitute/hellbender/tools/walkers/mutect/filtering/Mutect2FilteringEngine.java
@@ -63,7 +63,7 @@ public Mutect2FilteringEngine(M2FiltersArgumentCollection MTFAC, final VCFHeader
 
     /**
      * Maximum probability that a potential variant is not a true somatic mutation.  Variants with error probabilities
-     * below this threshold are called; variants with error probabilities above are filtered.
+     * at or below this threshold are called; variants with error probabilities above are filtered.
      */
     public double getThreshold() { return thresholdCalculator.getThreshold(); }
 
@@ -159,6 +159,11 @@ public void learnParameters() {
         filteringOutputStats.clear();
     }
 
+    public void learnThreshold() {
+        thresholdCalculator.relearnThresholdAndClearAcumulatedProbabilities();
+        filteringOutputStats.clear();
+    }
+
     /**
      * Create a filtered variant and record statistics for the final pass of {@link FilterMutectCalls}
      */
@@ -177,8 +182,9 @@ public VariantContext applyFiltersAndAccumulateOutputStats(final VariantContext
                 }
             });
 
-            // TODO: clarify this logic
-            if (errorProbability > EPSILON && errorProbability > getThreshold() - EPSILON) {
+            // error probability must exceed threshold, and just in case threshold is bad, probabilities close to 1 must be filtered
+            // and probabilities close to 0 must not be filtered
+            if (( errorProbability > Math.min(1 - EPSILON, Math.max(EPSILON, getThreshold())))) {
                 vcb.filter(entry.getKey().filterName());
             }
         }