diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index b9adede..3ac5504 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -48,6 +48,20 @@ jobs:
           echo "8aca8db96f1b94770f1b0d72b6dddcb1ebb8123cb3712530b08cc387b349a3d8  $RUNNER_TEMP/actionlint.tar.gz" | sha256sum --check
           tar -xzf "$RUNNER_TEMP/actionlint.tar.gz" -C "$RUNNER_TEMP" actionlint
           "$RUNNER_TEMP/actionlint" -color
+      - name: Format
+        shell: bash
+        run: |
+          curl --fail --silent --show-error --location --output "$RUNNER_TEMP/fourmolu.zip" https://github.com/fourmolu/fourmolu/releases/download/v0.20.1.0/fourmolu-0.20.1.0-linux-x86_64.zip
+          echo "bdd97063b445145b39191259c428bca489fa30578b82d9066fc7377db390f1d0  $RUNNER_TEMP/fourmolu.zip" | sha256sum --check
+          unzip -q "$RUNNER_TEMP/fourmolu.zip" -d "$RUNNER_TEMP"
+          git ls-files -z '*.hs' | xargs -0 "$RUNNER_TEMP/fourmolu-0.20.1.0-linux-x86_64/fourmolu" --mode check --record-brace-space true
+      - name: HLint
+        shell: bash
+        run: |
+          curl --fail --silent --show-error --location --output "$RUNNER_TEMP/hlint.tar.gz" https://github.com/ndmitchell/hlint/releases/download/v3.10/hlint-3.10-x86_64-linux.tar.gz
+          echo "ccabc8802a58154699a3583b8dddc5ea2e6d65753a62c45c0e80088ebb16b42b  $RUNNER_TEMP/hlint.tar.gz" | sha256sum --check
+          tar -xzf "$RUNNER_TEMP/hlint.tar.gz" -C "$RUNNER_TEMP"
+          "$RUNNER_TEMP/hlint-3.10/hlint" src test app
 
   haskell:
     name: Haskell
@@ -57,21 +71,68 @@ jobs:
         with:
           persist-credentials: false
       - uses: haskell-actions/setup@cd0d9bdd65b20557f41bea4dbe43d0b5fbbfe553 # v2.11.0
+        id: setup
         with:
           ghc-version: '9.14.1'
           cabal-version: '3.18.1.0'
-      - name: Build
+      - name: Resolve the build plan
         run: |
           cabal update
-          cabal build all --enable-tests
+          cabal build all --enable-tests --dry-run
+      - uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        id: store
+        with:
+          path: ${{ steps.setup.outputs.cabal-store }}
+          key: cabal-store-${{ runner.os }}-ghc-${{ steps.setup.outputs.ghc-version }}-${{ hashFiles('dist-newstyle/cache/plan.json') }}
+          restore-keys: cabal-store-${{ runner.os }}-ghc-${{ steps.setup.outputs.ghc-version }}-
+      - name: Build
+        run: cabal build all --enable-tests
+      - uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        if: steps.store.outputs.cache-hit != 'true'
+        with:
+          path: ${{ steps.setup.outputs.cabal-store }}
+          key: ${{ steps.store.outputs.cache-primary-key }}
       - name: Test
         run: cabal test all --test-show-details=direct
+      - name: Type fixtures
+        run: cabal exec -- sh test/types.sh ghc
+      - name: Confidence bounds against decimal arithmetic
+        run: |
+          mkdir -p "$RUNNER_TEMP/confidence"
+          cabal exec -- ghc -XGHC2021 -Wall -Werror -package invar -outputdir "$RUNNER_TEMP/confidence" -o "$RUNNER_TEMP/confidence/reference" test/confidence/Main.hs
+          python3 test/confidence/check.py "$RUNNER_TEMP/confidence/reference"
+
+  workers:
+    name: Workers
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: haskell-actions/setup@cd0d9bdd65b20557f41bea4dbe43d0b5fbbfe553 # v2.11.0
+        id: setup
+        with:
+          ghc-version: '9.14.1'
+          cabal-version: '3.18.1.0'
+      - name: Resolve the build plan
+        run: |
+          cabal update
+          cabal build all --enable-tests --dry-run
+      - uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        with:
+          path: ${{ steps.setup.outputs.cabal-store }}
+          key: cabal-store-${{ runner.os }}-ghc-${{ steps.setup.outputs.ghc-version }}-${{ hashFiles('dist-newstyle/cache/plan.json') }}
+          restore-keys: cabal-store-${{ runner.os }}-ghc-${{ steps.setup.outputs.ghc-version }}-
+      - name: Build the core
+        run: cabal build exe:invar
       - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:
           python-version: '3.12'
+          cache: pip
+          cache-dependency-path: worker/requirements.txt
       - name: Install the pinned numerical dependencies
-        run: python -m pip install --no-cache-dir --requirement worker/requirements.txt
-      - name: Test the shared worker modules and the hf package with the built core; mlx and vllm packages run on their native hosts
+        run: python -m pip install --requirement worker/requirements.txt
+      - name: Test the shared worker modules and the hf package with the built core, in three concurrent groups; mlx and vllm packages run on their native hosts
         shell: bash
         env:
           HF_HUB_OFFLINE: '1'
@@ -84,25 +145,34 @@ jobs:
           core_directory="$(dirname "$core")"
           export PATH="$core_directory:$PATH"
           python -B -m unittest worker.tests.invocation worker.tests.probepacked worker.tests.probestore
-          python -B -m unittest discover -s worker/tests/hf -t . -p '*.py'
-      - name: Format
-        shell: bash
-        run: |
-          curl --fail --silent --show-error --location --output "$RUNNER_TEMP/fourmolu.zip" https://github.com/fourmolu/fourmolu/releases/download/v0.20.1.0/fourmolu-0.20.1.0-linux-x86_64.zip
-          echo "bdd97063b445145b39191259c428bca489fa30578b82d9066fc7377db390f1d0  $RUNNER_TEMP/fourmolu.zip" | sha256sum --check
-          unzip -q "$RUNNER_TEMP/fourmolu.zip" -d "$RUNNER_TEMP"
-          git ls-files -z '*.hs' | xargs -0 "$RUNNER_TEMP/fourmolu-0.20.1.0-linux-x86_64/fourmolu" --mode check --record-brace-space true
-      - name: HLint
-        shell: bash
-        run: |
-          curl --fail --silent --show-error --location --output "$RUNNER_TEMP/hlint.tar.gz" https://github.com/ndmitchell/hlint/releases/download/v3.10/hlint-3.10-x86_64-linux.tar.gz
-          echo "ccabc8802a58154699a3583b8dddc5ea2e6d65753a62c45c0e80088ebb16b42b  $RUNNER_TEMP/hlint.tar.gz" | sha256sum --check
-          tar -xzf "$RUNNER_TEMP/hlint.tar.gz" -C "$RUNNER_TEMP"
-          "$RUNNER_TEMP/hlint-3.10/hlint" src test app
-      - name: Type fixtures
-        run: cabal exec -- sh test/types.sh ghc
-      - name: Confidence bounds against decimal arithmetic
-        run: |
-          mkdir -p "$RUNNER_TEMP/confidence"
-          cabal exec -- ghc -XGHC2021 -Wall -Werror -package invar -outputdir "$RUNNER_TEMP/confidence" -o "$RUNNER_TEMP/confidence/reference" test/confidence/Main.hs
-          python3 test/confidence/check.py "$RUNNER_TEMP/confidence/reference"
+          # The two slowest modules run alone; every other module, including new ones, runs in the third group.
+          slow=(distillation overlap)
+          rest=()
+          for path in worker/tests/hf/*.py; do
+            name="$(basename "$path" .py)"
+            case " __init__ ${slow[*]} " in
+              *" $name "*) ;;
+              *) rest+=("worker.tests.hf.$name") ;;
+            esac
+          done
+          groups=("${slow[@]}" rest)
+          pids=()
+          for group in "${slow[@]}"; do
+            python -B -m unittest "worker.tests.hf.$group" > "$RUNNER_TEMP/$group.log" 2>&1 &
+            pids+=("$!")
+          done
+          python -B -m unittest "${rest[@]}" > "$RUNNER_TEMP/rest.log" 2>&1 &
+          pids+=("$!")
+          failed=()
+          for index in "${!groups[@]}"; do
+            wait "${pids[index]}" || failed+=("${groups[index]}")
+          done
+          for group in "${groups[@]}"; do
+            echo "::group::hf tests: $group"
+            cat "$RUNNER_TEMP/$group.log"
+            echo "::endgroup::"
+          done
+          if [ "${#failed[@]}" -gt 0 ]; then
+            echo "Failed hf test groups: ${failed[*]}"
+            exit 1
+          fi
diff --git a/worker/tests/hf/overlap.py b/worker/tests/hf/overlap.py
index bb46100..a2daaaf 100644
--- a/worker/tests/hf/overlap.py
+++ b/worker/tests/hf/overlap.py
@@ -168,20 +168,21 @@ def test_two_optimizer_steps_per_update_are_counted_in_the_history(self):
         steps, = [item["state"]["steps"] for item in history["artifacts"]]
         self.assertTrue(steps and all(count == 2 for count in steps))
 
-    def test_a_default_two_cycle_run_goes_through_the_runtime_scores_the_reference_and_equals_an_explicit_staleness_zero(self):
+    def test_a_default_two_cycle_run_journals_its_declaration_scores_the_reference_and_repeats_exactly(self):
+        runs = {}
         for mode in ("serial", "resident"):
             with self.subTest(mode=mode):
-                records, default = self.train("default" + mode, 2, **{"inference-mode": mode})
-                _, explicit = self.train("explicit" + mode, 2, staleness=0, **{"inference-mode": mode})
-                self.assertEqual(journaled(default)[0]["entry"], "declaration")
+                records, runs[mode] = self.train("default" + mode, 2, **{"inference-mode": mode})
+                self.assertEqual(journaled(runs[mode])[0]["entry"], "declaration")
                 published = [value["policy"] for value in records if value.get("phase") == "published"]
                 self.assertEqual(len({self.settings["policy"], *published}), 3)
-                self.assertEqual([value is not None for value in references(default)], [False, False, True, True])
-                comparison = compared({"run": default, **independent(self.root, default / "generation2", "default" + mode)},
-                                      {"run": explicit, **independent(self.root, explicit / "generation2", "explicit" + mode)})
-                self.assertTrue(comparison["equal"])
-                self.assertTrue(comparison["schedule_equal"])
-                self.assertTrue(comparison["execution"]["declared_equal"])
+                self.assertEqual([value is not None for value in references(runs[mode])], [False, False, True, True])
+        _, repeated = self.train("repeated", 2, staleness=0)
+        comparison = compared({"run": runs["serial"], **independent(self.root, runs["serial"] / "generation2", "defaultserial")},
+                              {"run": repeated, **independent(self.root, repeated / "generation2", "repeated")})
+        self.assertTrue(comparison["equal"])
+        self.assertTrue(comparison["schedule_equal"])
+        self.assertTrue(comparison["execution"]["declared_equal"])
 
     def test_staleness_is_a_schedule_difference(self):
         _, fresh = self.train("fresh", 1, staleness=0)