diff --git a/spec/dcc/validate/xsd_performance_spec.rb b/spec/dcc/validate/xsd_performance_spec.rb
new file mode 100644
index 0000000..997af4f
--- /dev/null
+++ b/spec/dcc/validate/xsd_performance_spec.rb
@@ -0,0 +1,83 @@
+# frozen_string_literal: true
+
+require "spec_helper"
+
+# Regression guard for the schema memoization in `Dcc::Validate::Xsd`.
+#
+# Asserts no timing on purpose — a wall-clock benchmark would be
+# machine-dependent and flaky in CI. What these examples establish is that one
+# `Nokogiri::XML::Schema` gets built per distinct version, that a lookup returns
+# the schema stored under the version it asked for, and that looking the same
+# version up again — explicitly or via `:auto` — neither rebuilds it nor adds a
+# cache entry.
+RSpec.describe Dcc::Validate::Xsd do
+  let(:xml) { File.read(fixtures_path("dcclib", "valid.xml")) }
+  let(:cache) { described_class::SCHEMA_CACHE }
+  let(:builds) { [] }
+
+  # Records every real schema construction, letting each one through. Only the
+  # count is ever read, so store a marker rather than the argument — that
+  # argument is the resolved XSD, about 50 KB each time.
+  def track_builds
+    allow(Nokogiri::XML::Schema)
+      .to receive(:new).and_wrap_original do |original, *args, **kwargs|
+        builds << :built
+        original.call(*args, **kwargs)
+      end
+  end
+
+  # `SCHEMA_CACHE` is process-wide. Start cold, then hand the suite back
+  # exactly the entries it had.
+  around do |example|
+    warm = cache.dup
+    cache.clear
+    example.run
+  ensure
+    cache.replace(warm)
+  end
+
+  # The build count and the cache keys are asserted together on purpose. Two
+  # builds alone would also match one version built twice; the keys say which
+  # two were built. `.with` cannot discriminate here — the argument is the XSD
+  # document, not the version.
+  it "builds one schema per distinct version" do
+    track_builds
+
+    3.times { described_class.call(xml, version: "3.3.0") }
+    2.times { described_class.call(xml, version: "3.2.0") }
+
+    expect([builds.size, cache.keys]).to eq([2, %w[3.3.0 3.2.0]])
+  end
+
+  it "hands back the schema cached under the requested version" do
+    versions = %w[3.3.0 3.2.0]
+    versions.each { |version| described_class.call(xml, version: version) }
+
+    looked_up = versions.to_h { |v| [v, described_class.send(:schema_for, v)] }
+
+    expect(looked_up).to eq(versions.to_h { |v| [v, cache.fetch(v)] })
+  end
+
+  # Both examples warm the cache first and only then start counting, so any
+  # build recorded here is a cache miss. The keys are asserted alongside the
+  # count, so a rebuild that overwrote the entry fails too. Counting the cold
+  # build is the first example's job. This one pins the explicit-version path,
+  # the next one pins `:auto`.
+  it "reuses the cached schema when the same version is looked up again" do
+    described_class.call(xml, version: "3.3.0")
+    track_builds
+
+    3.times { described_class.call(xml, version: "3.3.0") }
+
+    expect([builds.size, cache.keys]).to eq([0, %w[3.3.0]])
+  end
+
+  it "reuses the cached schema when the version is auto-detected" do
+    described_class.call(xml, version: "3.3.0")
+    track_builds
+
+    3.times { described_class.call(xml) }
+
+    expect([builds.size, cache.keys]).to eq([0, %w[3.3.0]])
+  end
+end
diff --git a/spec/dcc/validate_spec.rb b/spec/dcc/validate_spec.rb
index dd5342c..d0fa682 100644
--- a/spec/dcc/validate_spec.rb
+++ b/spec/dcc/validate_spec.rb
@@ -31,13 +31,6 @@
       expect(result.errors.first.line).to be_a(Integer)
       expect(result.errors.first.source).to eq("xsd")
     end
-
-    it "caches the loaded schema" do
-      described_class.call(valid_xml)
-      cache_before = described_class.const_get(:SCHEMA_CACHE).size
-      described_class.call(valid_xml)
-      expect(described_class.const_get(:SCHEMA_CACHE).size).to eq(cache_before)
-    end
   end
 end