diff --git a/.github/workflows/benchmarks.yml b/.github/workflows/benchmarks.yml
index 5bfc8a5..8a69f9d 100644
--- a/.github/workflows/benchmarks.yml
+++ b/.github/workflows/benchmarks.yml
@@ -122,7 +122,10 @@ jobs:
           BASE_BUILT: ${{ steps.build_base.outcome == 'success' }}
         run: |
           DLL=tests/EncDotNet.Benchmarks/bin/Release/net10.0/EncDotNet.Benchmarks.dll
-          ARGS=(--filter '*' --exporters json --warmupCount 3 --iterationCount 10 --launchCount 1)
+          # BenchmarkDotNet's default 500 ms iterations make each case take 10-15 s however
+          # fast the operation is. Shorter iterations, with more of them, halve the run time
+          # while keeping the median as precise (standard error about 2%).
+          ARGS=(--filter '*' --exporters json --warmupCount 3 --iterationCount 15 --iterationTime 200 --launchCount 1)
           if [ "$SCOPE" = "core" ]; then
             ARGS+=(--anyCategories Core)
           fi
diff --git a/scripts/DownloadBenchmarkCells.cs b/scripts/DownloadBenchmarkCells.cs
index 1b2442a..803f7bb 100644
--- a/scripts/DownloadBenchmarkCells.cs
+++ b/scripts/DownloadBenchmarkCells.cs
@@ -7,11 +7,17 @@
 using EncDotNet.Noaa;
 
 // ============================================================================
-// DownloadBenchmarkCells.cs — Downloads a fixed set of real NOAA ENC cells for
+// DownloadBenchmarkCells.cs — Downloads a few real NOAA ENC cells for
 // the benchmarks in tests/EncDotNet.Benchmarks.
 //
 // Usage: dotnet run scripts/DownloadBenchmarkCells.cs -- <output-dir> [cell...]
 //
+// Without cell names, the script picks one cell per usage band (harbour, approach
+// and coastal) from the live NOAA catalog: among active cells with update files,
+// the one at the 90th percentile of download size, so the set is substantial but
+// typical. Picking by rule rather than by name survives NOAA withdrawing or
+// renaming cells, which it does as it re-schemes its coverage.
+//
 // Each cell is extracted to <output-dir>/<cell>/ with its update files. Point
 // ENCDOTNET_ENC_ROOT at <output-dir> to include them in a benchmark run. The
 // script writes <output-dir>/cells.txt recording each cell's edition, update
@@ -24,9 +30,10 @@
 // Never commit the downloaded cells (see CONTRIBUTING.md).
 // ============================================================================
 
-// A harbour, an approach and a coastal cell, all with update files. Together they
-// cover the range of record counts and edge lengths in typical NOAA cells.
-string[] defaultCells = ["US5WA50M", "US4MD81M", "US3WA01M"];
+// Navigational purpose (usage band), the third character of the cell file name
+// (S-57 Appendix B.1, 5.6.4).
+(char Band, string Purpose)[] bands = [('5', "harbour"), ('4', "approach"), ('3', "coastal")];
+const double sizePercentile = 0.9;
 
 if (args.Length == 0)
 {
@@ -35,13 +42,43 @@
 }
 
 var outputDirectory = Path.GetFullPath(args[0]);
-var cellNames = args.Length > 1 ? args[1..] : defaultCells;
 Directory.CreateDirectory(outputDirectory);
 
 using var catalogClient = new EncProductCatalogClient();
 var catalog = await catalogClient.GetNoaaCatalogAsync();
 var cellsByName = catalog.Cells.ToDictionary(c => c.Name, StringComparer.OrdinalIgnoreCase);
 
+string[] cellNames;
+if (args.Length > 1)
+{
+    cellNames = args[1..];
+}
+else
+{
+    var picked = new List<string>();
+    foreach (var (band, purpose) in bands)
+    {
+        var candidates = catalog.Cells
+            .Where(c => c.Name.Length == 8 && c.Name[2] == band)
+            .Where(c => c.UpdateNumber > 0 && c.ZipfileSize > 0)
+            .Where(c => !c.Status.Contains("cancel", StringComparison.OrdinalIgnoreCase))
+            .OrderBy(c => c.ZipfileSize)
+            .ThenBy(c => c.Name, StringComparer.Ordinal)
+            .ToList();
+        if (candidates.Count == 0)
+        {
+            Console.Error.WriteLine($"warning: no active {purpose} cell with updates in the NOAA catalog.");
+            continue;
+        }
+
+        var cell = candidates[(int)((candidates.Count - 1) * sizePercentile)];
+        Console.WriteLine($"Picked {cell.Name} as the {purpose} cell ({candidates.Count} candidates, zip {cell.ZipfileSize:N0} bytes).");
+        picked.Add(cell.Name);
+    }
+
+    cellNames = [.. picked];
+}
+
 using var httpClient = new HttpClient();
 var manifest = new List<string>();
 var missing = 0;
@@ -96,4 +133,4 @@
 }
 
 File.WriteAllLines(Path.Combine(outputDirectory, "cells.txt"), manifest);
-return missing == cellNames.Length ? 1 : 0;
+return cellNames.Length == 0 || missing == cellNames.Length ? 1 : 0;
diff --git a/tests/EncDotNet.PerformanceTests/PerformanceCells.cs b/tests/EncDotNet.PerformanceTests/PerformanceCells.cs
index 37cab38..500dc4c 100644
--- a/tests/EncDotNet.PerformanceTests/PerformanceCells.cs
+++ b/tests/EncDotNet.PerformanceTests/PerformanceCells.cs
@@ -13,7 +13,7 @@ internal static class PerformanceCells
     private static readonly Lazy<Inputs> LazyHarbourX4 = new(() => new Inputs(SyntheticCellProfile.Harbour.Scale(4)));
 
     /// <summary>
-    /// Gets the harbour-scale cell, shaped like NOAA US5WA50M.
+    /// Gets the harbour-scale cell, shaped like an earlier edition of NOAA US5WA50M.
     /// </summary>
     public static Inputs Harbour => LazyHarbour.Value;
 
diff --git a/tests/Shared/SyntheticCellProfile.cs b/tests/Shared/SyntheticCellProfile.cs
index f3f435e..770bcaf 100644
--- a/tests/Shared/SyntheticCellProfile.cs
+++ b/tests/Shared/SyntheticCellProfile.cs
@@ -6,7 +6,10 @@ namespace EncDotNet.Testing;
 /// <remarks>
 /// The built-in profiles are calibrated against real NOAA cells: record counts, the
 /// distribution of edge lengths, attributes per feature and spatial pointers per feature
-/// follow the measured averages of the cell named in each profile's remarks.
+/// follow the measured averages of the cell named in each profile's summary. Those were
+/// earlier editions; NOAA has since re-schemed its coverage into smaller cells (edition 36
+/// of US4MD81M is about 320 KB), so the larger profiles now stand for the bigger cells an
+/// application may still meet rather than for a typical current one.
 /// </remarks>
 /// <param name="Name">The profile name, used in the generated data set name and in benchmark output.</param>
 /// <param name="Seed">The seed for the generator; the same profile always yields the same bytes.</param>
@@ -35,7 +38,7 @@ public sealed record SyntheticCellProfile(
     int AreaFeatures)
 {
     /// <summary>
-    /// A harbour-scale cell shaped like NOAA US5WA50M (about 400 KB).
+    /// A harbour-scale cell shaped like an earlier edition of NOAA US5WA50M (about 400 KB).
     /// </summary>
     public static SyntheticCellProfile Harbour { get; } = new(
         Name: "Harbour",
@@ -52,7 +55,7 @@ public sealed record SyntheticCellProfile(
         AreaFeatures: 211);
 
     /// <summary>
-    /// An approach-scale cell shaped like NOAA US4MD81M (about 3.5 MB).
+    /// An approach-scale cell shaped like an earlier edition of NOAA US4MD81M (about 3.5 MB).
     /// </summary>
     public static SyntheticCellProfile Approach { get; } = new(
         Name: "Approach",