diff --git a/benchmarks/run.sh b/benchmarks/run.sh
index 4c9b758..fd8ef6f 100755
--- a/benchmarks/run.sh
+++ b/benchmarks/run.sh
@@ -7,6 +7,7 @@ durability="${DURABILITY:-async}"
 threads="${THREADS:-$(getconf _NPROCESSORS_ONLN)}"
 heap="${HEAP:-4g}"
 history="${HISTORY:-64}"
+runs="${RUNS:-3}"
 results="results/scale-$scale-$durability-history-$history"
 
 ../mvnw -q -f ../pom.xml -Pbenchmarks -DskipTests install
@@ -17,9 +18,11 @@ jvm=(-Xms"$heap" -Xmx"$heap" -XX:+UseParallelGC
 	--add-opens java.base/java.util=ALL-UNNAMED
 	-cp "$classpath")
 
-mkdir -p "$results"
-for store in hstore hypergraphdb; do
-	java "${jvm[@]}" io.hstore.bench.Comparison run --store "$store" --scale "$scale" --durability "$durability" \
-		--threads "$threads" --history "$history" --out "$results/$store.json"
+rm -rf "$results" && mkdir -p "$results"
+for run in $(seq 1 "$runs"); do
+	for store in hstore hypergraphdb; do
+		java "${jvm[@]}" io.hstore.bench.Comparison run --store "$store" --scale "$scale" --durability "$durability" \
+			--threads "$threads" --history "$history" --out "$results/$store-$run.json"
+	done
 done
-java "${jvm[@]}" io.hstore.bench.Comparison report "$results/hstore.json" "$results/hypergraphdb.json" | tee "$results/report.md"
+java "${jvm[@]}" io.hstore.bench.Comparison report "$results"/hstore-*.json "$results"/hypergraphdb-*.json | tee "$results/report.md"
diff --git a/benchmarks/src/main/java/io/hstore/bench/Comparison.java b/benchmarks/src/main/java/io/hstore/bench/Comparison.java
index dd256b9..b74ec3a 100644
--- a/benchmarks/src/main/java/io/hstore/bench/Comparison.java
+++ b/benchmarks/src/main/java/io/hstore/bench/Comparison.java
@@ -55,7 +55,7 @@ public static void main(String[] args) throws Exception {
             default -> {
                 IO.println("""
                         usage: Comparison run --store hstore|hypergraphdb [--scale N] [--durability async|sync] [--threads N] [--history N] --out FILE
-                               Comparison report FILE...""");
+                               Comparison report FILE... (runs of two stores; cells show the median and range)""");
                 System.exit(2);
             }
         }
diff --git a/benchmarks/src/main/java/io/hstore/bench/Report.java b/benchmarks/src/main/java/io/hstore/bench/Report.java
index c270bc1..6a3e675 100644
--- a/benchmarks/src/main/java/io/hstore/bench/Report.java
+++ b/benchmarks/src/main/java/io/hstore/bench/Report.java
@@ -11,7 +11,7 @@
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
-import java.util.Objects;
+import java.util.stream.Stream;
 
 final class Report {
 
@@ -22,12 +22,17 @@ private Report() {
     }
 
     static String markdown(String[] args) {
-        List<Run> runs = Arrays.stream(args).skip(1).map(Path::of).map(Report::load).toList();
-        if (runs.size() < 2) {
-            throw new IllegalArgumentException("report needs at least two result files");
+        Map<String, List<Run>> byStore = new LinkedHashMap<>();
+        Arrays.stream(args).skip(1).map(Path::of).map(Report::load)
+                .forEach(run -> byStore.computeIfAbsent(run.store(), _ -> new ArrayList<>()).add(run));
+        if (byStore.size() != 2) {
+            throw new IllegalArgumentException("report needs runs of exactly two stores, got " + byStore.keySet());
         }
-        Run baseline = runs.getLast();
-        Run subject = runs.getFirst();
+        List<List<Run>> stores = List.copyOf(byStore.values());
+        List<Run> subjects = stores.getFirst();
+        List<Run> baselines = stores.getLast();
+        Run subject = subjects.getFirst();
+        Run baseline = baselines.getFirst();
         StringBuilder out = new StringBuilder();
         Map<String, Json> header = subject.header();
         out.append("Dataset: %s nodes, %s hyperedges, %s incidences; durability %s; %s threads.%n"
@@ -37,30 +42,53 @@ static String markdown(String[] args) {
                 text(header.get("java")), number(header.get("maxHeap")) / (1 << 20)));
         out.append("Caches: %s uses a %s; %s uses a %s.%n%n".formatted(subject.store(), text(subject.header().get("cache")),
                 baseline.store(), text(baseline.header().get("cache"))));
-        out.append("| Workload | %s | %s | Ratio | Results agree |%n".formatted(subject.store(), baseline.store()));
+        out.append("Each cell is the median of %d runs of %s and %d runs of %s, with the range in brackets.%n%n"
+                .formatted(subjects.size(), subject.store(), baselines.size(), baseline.store()));
+        out.append("| Workload | %s | %s | Ratio of medians | Results agree |%n".formatted(subject.store(), baseline.store()));
         out.append("|---|---:|---:|---:|:---:|%n".formatted());
         for (Map.Entry<String, Map<String, Json>> entry : subject.results().entrySet()) {
             String workload = entry.getKey();
-            Map<String, Json> mine = entry.getValue();
-            Map<String, Json> theirs = baseline.results().get(workload);
-            if (theirs == null) {
+            if (baselines.stream().anyMatch(run -> !run.results().containsKey(workload))) {
                 continue;
             }
             boolean disk = workload.equals("disk");
-            boolean single = number(mine.get("operations")) == 1;
-            double a = disk ? number(mine.get("operations")) : single ? number(mine.get("millis")) : rate(mine);
-            double b = disk ? number(theirs.get("operations")) : single ? number(theirs.get("millis")) : rate(theirs);
-            double ratio = disk || single ? b / a : a / b;
+            boolean single = number(entry.getValue().get("operations")) == 1;
+            double[] mine = values(subjects, workload, disk, single);
+            double[] theirs = values(baselines, workload, disk, single);
+            double ratio = disk || single ? median(theirs) / median(mine) : median(mine) / median(theirs);
             boolean agree = disk || workload.startsWith("ingest") || workload.equals("reopen")
-                    || Objects.equals(number(mine.get("checksum")), number(theirs.get("checksum")));
-            out.append("| `%s` — %s | %s | %s | **%.2f×** | %s |%n".formatted(workload, text(mine.get("description")),
-                    format(a, disk, single), format(b, disk, single), ratio, agree ? "yes" : "**no**"));
+                    || Stream.concat(subjects.stream(), baselines.stream())
+                    .map(run -> number(run.results().get(workload).get("checksum"))).distinct().count() == 1;
+            out.append("| `%s`: %s | %s | %s | **%.2f×** | %s |%n".formatted(workload, text(entry.getValue().get("description")),
+                    cell(mine, disk, single), cell(theirs, disk, single), ratio, agree ? "yes" : "**no**"));
         }
         out.append("%nRatios above 1 favour %s: throughput ratios divide %s by %s; latency and size ratios divide %s by %s.%n"
                 .formatted(subject.store(), subject.store(), baseline.store(), baseline.store(), subject.store()));
         return out.toString();
     }
 
+    private static double[] values(List<Run> runs, String workload, boolean disk, boolean single) {
+        return runs.stream().mapToDouble(run -> {
+            Map<String, Json> result = run.results().get(workload);
+            return disk ? number(result.get("operations")) : single ? number(result.get("millis")) : rate(result);
+        }).sorted().toArray();
+    }
+
+    private static double median(double[] sorted) {
+        int middle = sorted.length / 2;
+        return sorted.length % 2 == 1 ? sorted[middle] : (sorted[middle - 1] + sorted[middle]) / 2;
+    }
+
+    private static String cell(double[] sorted, boolean disk, boolean single) {
+        String median = format(median(sorted), disk, single);
+        return sorted.length == 1 ? median
+                : median + " (" + bare(sorted[0], disk) + " to " + bare(sorted[sorted.length - 1], disk) + ")";
+    }
+
+    private static String bare(double value, boolean disk) {
+        return disk ? "%.1f".formatted(value / (1 << 20)) : "%,.0f".formatted(value);
+    }
+
     private static String format(double value, boolean disk, boolean single) {
         if (disk) {
             return "%.1f MiB".formatted(value / (1 << 20));
diff --git a/docs/benchmarks.md b/docs/benchmarks.md
index db4ddc9..88c4be4 100644
--- a/docs/benchmarks.md
+++ b/docs/benchmarks.md
@@ -176,7 +176,10 @@ SCALE=4 DURABILITY=async benchmarks/run.sh
 SCALE=4 DURABILITY=async HISTORY=1 benchmarks/run.sh
 ```
 
-`run.sh` builds the `benchmarks` module (`-Pbenchmarks`) and runs each store in its own JVM. It writes
-`benchmarks/results/scale-<s>-<durability>-history-<h>/{hstore,hypergraphdb}.json` and the Markdown report
-`report.md` in the same directory. `THREADS` and `HEAP` override the thread count and heap size. Close other
+`run.sh` builds the `benchmarks` module (`-Pbenchmarks`) and runs each store in its own JVM, `RUNS` times
+(default 3), alternating the two stores so slow periods hit both. It writes
+`benchmarks/results/scale-<s>-<durability>-history-<h>/{hstore,hypergraphdb}-<run>.json`, plus a Markdown report
+`report.md` in the same directory. Each cell in the report is the median over the runs, with the fastest and slowest
+run in brackets, and "Results agree" requires every run of both stores to return the same checksum. `THREADS`
+and `HEAP` override the thread count and heap size. Close other
 workloads first: the numbers are only meaningful on an otherwise idle machine.
diff --git a/docs/development.md b/docs/development.md
index 168b7c2..972751f 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -211,7 +211,7 @@ SCALE=1 DURABILITY=sync  benchmarks/run.sh
 `core/src/java` and `storage/bdb-je/src/java` with `javac --release 11` against Berkeley DB JE 5.0.73, and
 installs `org.hypergraphdb:hgdb` and `org.hypergraphdb:hgbdbje` version `1.4-99485a1`.
 
-`run.sh` builds with `-Pbenchmarks`, runs each store in its own JVM on the same deterministic dataset (seeded
+`run.sh` builds with `-Pbenchmarks`, runs each store `RUNS` times (default 3) in its own JVM on the same deterministic dataset (seeded
 generator, `SCALE` × 50,000 nodes and × 100,000 hyperedges), writes `benchmarks/results/scale-N-durability/*.json`
 and prints a Markdown report whose *Results agree* column cross-checks every query checksum between the two
 engines. Methodology and published numbers: [benchmarks.md](benchmarks.md).