diff --git a/.claude/skills/benchmark/SKILL.md b/.claude/skills/benchmark/SKILL.md
index 8f6a2e78a..7bcb69942 100644
--- a/.claude/skills/benchmark/SKILL.md
+++ b/.claude/skills/benchmark/SKILL.md
@@ -55,6 +55,12 @@ go run . -benchnum 5
 
 Multi-component snapshot with configurable parallelism:
 
+```bash
+make benchmark_stress
+```
+
+Or manually:
+
 ```bash
 cd benchmark/stress
 ./prepare_data.sh
@@ -69,7 +75,30 @@ EC_STRESS_COMPONENTS=50 EC_STRESS_WORKERS=20 go run .
 
 Defaults: 10 components, 35 workers.
 
-## Step 5: Profile if needed
+## Step 5: Compare against baseline
+
+The stress benchmark has regression detection. After running:
+
+```bash
+cd benchmark/stress
+./compare.sh benchmark-output.txt
+```
+
+This compares current results against `baseline.json` using thresholds from
+`thresholds.json` (default: 15% RSS, 20% ns/op). Exits non-zero on regression.
+
+## Step 6: Regenerate baseline
+
+After intentional performance changes, update the stored baseline:
+
+```bash
+make generate-baseline
+```
+
+This runs the stress benchmark, parses results, and writes `benchmark/stress/baseline.json`
+with current metrics, commit SHA, date, and Go version.
+
+## Step 7: Profile if needed
 
 Use the CLI's built-in profiling:
 
@@ -79,7 +108,7 @@ ec validate image --trace=cpu ...     # pprof CPU profile
 ec validate image --trace=mem ...     # heap profile
 ```
 
-## Step 6: Report results
+## Step 8: Report results
 
 Output is in standard Go benchmark format (ns/op, memory stats). Summarize:
 - Benchmark type run (simple/stress)
diff --git a/.github/workflows/benchmark.yaml b/.github/workflows/benchmark.yaml
new file mode 100644
index 000000000..655982aa2
--- /dev/null
+++ b/.github/workflows/benchmark.yaml
@@ -0,0 +1,168 @@
+# Copyright The Conforma Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+---
+name: Stress Benchmark
+
+"on":
+  pull_request:
+    branches:
+      - main
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+
+  stress:
+    name: Stress Benchmark
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    env:
+      # Tuned for 4 vCPU / 16 GB CI runners to complete within 5 minutes.
+      # Code defaults are 10 components / 35 workers.
+      EC_STRESS_COMPONENTS: "10"
+      EC_STRESS_WORKERS: "10"
+    steps:
+      - name: Harden Runner
+        uses: step-security/harden-runner@b09bb98e06d4d774595224525879c09bc6e98c40 # v2.20.1
+        with:
+          egress-policy: audit
+          disable-telemetry: true
+
+      - name: Checkout repository
+        uses: actions/checkout@d23441a48e516b6c34aea4fa41551a30e30af803 # v6.1.0
+
+      - name: Restore Cache
+        uses: actions/cache/restore@caa296126883cff596d87d8935842f9db880ef25 # v5.1.0
+        with:
+          key: main
+          path: '**'
+
+      - name: Setup Go environment
+        uses: actions/setup-go@924ae3a1cded613372ab5595356fb5720e22ba16 # v6.5.0
+        with:
+          go-version-file: go.mod
+          cache: false
+
+      - name: Setup ORAS
+        uses: oras-project/setup-oras@1d808f7d7f6995cc68b7bf507bfe5c5446e1dc9d # v2.0.1
+
+      - name: Prepare benchmark data
+        run: |
+          cd benchmark/stress
+          ./prepare_data.sh
+
+      - name: Build stress benchmark
+        run: go build -o benchmark/stress/stress ./benchmark/stress
+
+      - name: Run stress benchmark
+        id: bench
+        continue-on-error: true
+        run: |
+          set -o pipefail
+          cd benchmark/stress
+          ./stress 2>benchmark-stderr.txt | tee benchmark-output.txt
+
+      - name: Compare against baseline
+        id: compare
+        if: steps.bench.outcome == 'success'
+        run: |
+          cd benchmark/stress
+          if [[ -f baseline.json ]]; then
+            ./compare.sh benchmark-output.txt
+          else
+            echo "No baseline found, skipping comparison."
+          fi
+
+      - name: Write job summary
+        if: always()
+        run: |
+          if [[ "${{ steps.bench.outcome }}" == "failure" && -s benchmark/stress/benchmark-stderr.txt ]]; then
+            {
+              echo "### Stderr"
+              echo '```'
+              tail -50 benchmark/stress/benchmark-stderr.txt
+              echo '```'
+            } >> "$GITHUB_STEP_SUMMARY"
+          fi
+
+          if [[ ! -f benchmark/stress/benchmark-output.txt ]]; then
+            echo "## Stress Benchmark" >> "$GITHUB_STEP_SUMMARY"
+            echo "Benchmark did not produce output." >> "$GITHUB_STEP_SUMMARY"
+            exit 0
+          fi
+
+          line=$(grep '^BenchmarkStress' benchmark/stress/benchmark-output.txt || true)
+          if [[ -z "$line" ]]; then
+            echo "## Stress Benchmark" >> "$GITHUB_STEP_SUMMARY"
+            echo "No benchmark results found in output." >> "$GITHUB_STEP_SUMMARY"
+            exit 0
+          fi
+
+          read -r ns_op peak_rss alloc heap < <(BENCH_LINE="$line" python3 -c "
+          import os, re
+          line = os.environ['BENCH_LINE']
+          def val(p):
+              m = re.search(p, line)
+              return m.group(1) if m else '0'
+          print(val(r'([\d.]+)\s+ns/op'), val(r'([\d.]+)\s+peak-RSS-bytes'), val(r'([\d.]+)\s+allocated-bytes/op'), val(r'([\d.]+)\s+heap-bytes-from-system'))
+          ")
+
+          secs=$(awk -v val="${ns_op:-0}" 'BEGIN {printf "%.1f", val / 1000000000}')
+          rss_mb=$(awk -v val="${peak_rss:-0}" 'BEGIN {printf "%.0f", val / 1048576}')
+          alloc_mb=$(awk -v val="${alloc:-0}" 'BEGIN {printf "%.0f", val / 1048576}')
+          heap_mb=$(awk -v val="${heap:-0}" 'BEGIN {printf "%.0f", val / 1048576}')
+
+          has_baseline=false
+          if [[ -f benchmark/stress/baseline.json ]]; then
+            has_baseline=true
+            bl_rss=$(python3 -c "import json; print(json.load(open('benchmark/stress/baseline.json'))['peak_rss_bytes'])")
+            bl_ns=$(python3 -c "import json; print(json.load(open('benchmark/stress/baseline.json'))['ns_per_op'])")
+            bl_rss_mb=$(awk -v val="$bl_rss" 'BEGIN {printf "%.0f", val / 1048576}')
+            bl_secs=$(awk -v val="$bl_ns" 'BEGIN {printf "%.1f", val / 1000000000}')
+            rss_change=$(awk -v cur="$peak_rss" -v base="$bl_rss" 'BEGIN {printf "%+.1f", ((cur - base) / base) * 100}')
+            time_change=$(awk -v cur="$ns_op" -v base="$bl_ns" 'BEGIN {printf "%+.1f", ((cur - base) / base) * 100}')
+          fi
+
+          {
+            echo "## Stress Benchmark"
+            echo ""
+            if [[ "$has_baseline" == "true" ]]; then
+              echo "| Metric | Current | Baseline | Change | Description |"
+              echo "|--------|---------|----------|--------|-------------|"
+              echo "| Components | ${EC_STRESS_COMPONENTS} | | | Snapshot components validated |"
+              echo "| Workers | ${EC_STRESS_WORKERS} | | | Parallel validation workers |"
+              echo "| Execution time | ${secs}s | ${bl_secs}s | ${time_change}% | Wall-clock time per iteration |"
+              echo "| Peak RSS | ${rss_mb} MB | ${bl_rss_mb} MB | ${rss_change}% | Max physical memory used |"
+              echo "| Allocated memory | ${alloc_mb} MB | | | Total Go heap allocations |"
+              echo "| Heap from system | ${heap_mb} MB | | | Heap memory requested from OS |"
+            else
+              echo "| Metric | Value | Description |"
+              echo "|--------|-------|-------------|"
+              echo "| Components | ${EC_STRESS_COMPONENTS} | Snapshot components validated |"
+              echo "| Workers | ${EC_STRESS_WORKERS} | Parallel validation workers |"
+              echo "| Execution time | ${secs}s | Wall-clock time per iteration |"
+              echo "| Peak RSS | ${rss_mb} MB | Max physical memory used |"
+              echo "| Allocated memory | ${alloc_mb} MB | Total Go heap allocations |"
+              echo "| Heap from system | ${heap_mb} MB | Heap memory requested from OS |"
+            fi
+            if [[ "${{ steps.compare.outcome }}" == "failure" ]]; then
+              echo ""
+              echo "> **⚠️ Performance regression detected.** Update the baseline with \`make generate-baseline\` if this is expected."
+            fi
+          } >> "$GITHUB_STEP_SUMMARY"
diff --git a/AGENTS.md b/AGENTS.md
index 1d1cbbd6a..79fadb343 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -59,6 +59,26 @@ Acceptance tests require `/etc/hosts` entries:
 127.0.0.1 rekor.localhost
 ```
 
+## Benchmarks
+
+Two benchmarks live under `benchmark/`:
+
+- **simple/** — Single-component validation against the `@redhat` policy collection.
+- **stress/** — Multi-component validation with configurable parallelism
+  (`EC_STRESS_COMPONENTS`, `EC_STRESS_WORKERS`).
+
+```bash
+make benchmark              # Run simple benchmark
+make benchmark_stress       # Run stress benchmark (via pattern rule)
+make generate-baseline      # Run stress benchmark and write baseline.json
+```
+
+The stress benchmark has regression detection: `benchmark/stress/baseline.json` stores
+reference metrics (peak RSS, ns/op) and `benchmark/stress/thresholds.json` defines
+percentage thresholds. CI runs `benchmark/stress/compare.sh` to compare each run
+against the baseline and fails the check on regression. Update the baseline with
+`make generate-baseline` after intentional performance changes.
+
 ## Single-File Verification
 
 ```bash
diff --git a/Makefile b/Makefile
index 9f094e69e..efe3d34eb 100644
--- a/Makefile
+++ b/Makefile
@@ -197,6 +197,28 @@ benchmark_data: benchmark/simple/data.tar.gz ## Prepare data for benchmark
 .PHONY: benchmark
 benchmark: benchmark_simple ## Run benchmarks
 
+.PHONY: generate-baseline
+generate-baseline: benchmark/stress/data.tar.gz ## Generate stress benchmark baseline
+	@cd benchmark/stress && \
+	go run . 2>benchmark-stderr.txt | tee benchmark-output.txt && \
+	python3 -c "\
+	import re, json, sys; \
+	line = [l for l in open('benchmark-output.txt') if l.startswith('BenchmarkStress')]; \
+	line or sys.exit('No BenchmarkStress results found'); \
+	line = line[0]; \
+	def val(p): \
+	    m = re.search(p, line); \
+	    return m.group(1) if m else ''; \
+	ns = val(r'([\d.]+)\s+ns/op'); rss = val(r'([\d.]+)\s+peak-RSS-bytes'); \
+	(ns and rss) or sys.exit('Failed to parse benchmark metrics'); \
+	json.dump({'peak_rss_bytes': int(float(rss)), 'ns_per_op': int(float(ns)), \
+	  'components': int('$${EC_STRESS_COMPONENTS:-10}'), 'workers': int('$${EC_STRESS_WORKERS:-10}'), \
+	  'commit': '$(shell git rev-parse --short HEAD)', 'date': '$(shell date -u +%Y-%m-%d)', \
+	  'go_version': '$(shell go env GOVERSION | sed "s/^go//")' \
+	}, open('baseline.json','w'), indent=2); print()" && \
+	rm -f benchmark-output.txt benchmark-stderr.txt && \
+	echo "Baseline written to benchmark/stress/baseline.json"
+
 .PHONY: tools-ci
 tools-ci: ## Ensure all tools build cleanly
 	@echo "• tkn:" && \
diff --git a/benchmark/README.md b/benchmark/README.md
index fe72b6d6e..733e72b53 100644
--- a/benchmark/README.md
+++ b/benchmark/README.md
@@ -17,3 +17,25 @@ times.
 - **stress/** — Multi-component validation with configurable parallelism. Set
   `EC_STRESS_COMPONENTS` (default 10) and `EC_STRESS_WORKERS` (default 35) to
   control the workload.
+
+## Baseline and regression detection
+
+The stress benchmark stores a performance baseline in
+`stress/baseline.json` (peak RSS and ns/op) along with configurable
+regression thresholds in `stress/thresholds.json`. The CI workflow
+compares each run against the baseline and fails the check when a metric
+exceeds its threshold.
+
+To regenerate the baseline after an intentional change:
+
+```
+make generate-baseline
+```
+
+This runs the stress benchmark locally, parses the results, and writes a
+new `baseline.json` with the current commit SHA, date, Go version, and
+worker/component counts.
+
+Thresholds are expressed as percentages (e.g., 15 means a 15% increase
+triggers a failure). Adjust them in `stress/thresholds.json` as
+optimizations land.
diff --git a/benchmark/stress/baseline.json b/benchmark/stress/baseline.json
new file mode 100644
index 000000000..9260dea7d
--- /dev/null
+++ b/benchmark/stress/baseline.json
@@ -0,0 +1,9 @@
+{
+  "peak_rss_bytes": 2250485760,
+  "ns_per_op": 2567888013,
+  "components": 10,
+  "workers": 10,
+  "commit": "fc37eb13",
+  "date": "2026-08-11",
+  "go_version": "1.26.3"
+}
diff --git a/benchmark/stress/compare.sh b/benchmark/stress/compare.sh
new file mode 100755
index 000000000..f0462703d
--- /dev/null
+++ b/benchmark/stress/compare.sh
@@ -0,0 +1,108 @@
+#!/bin/bash
+# Copyright The Conforma Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# Compares current benchmark results against a stored baseline and exits
+# non-zero if any metric regresses beyond the configured threshold.
+set -o errexit
+set -o nounset
+set -o pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+BASELINE="${SCRIPT_DIR}/baseline.json"
+THRESHOLDS="${SCRIPT_DIR}/thresholds.json"
+BENCHMARK_OUTPUT="${1:-${SCRIPT_DIR}/benchmark-output.txt}"
+
+if [[ ! -f "$BASELINE" ]]; then
+    echo "No baseline found, skipping comparison."
+    exit 0
+fi
+
+if [[ ! -f "$THRESHOLDS" ]]; then
+    echo "No thresholds file found, skipping comparison."
+    exit 0
+fi
+
+if [[ ! -f "$BENCHMARK_OUTPUT" ]]; then
+    echo "No benchmark output found at ${BENCHMARK_OUTPUT}"
+    exit 1
+fi
+
+line=$(grep '^BenchmarkStress' "$BENCHMARK_OUTPUT" || true)
+if [[ -z "$line" ]]; then
+    echo "No BenchmarkStress results found in output."
+    exit 1
+fi
+
+read -r current_ns current_rss baseline_ns baseline_rss threshold_rss threshold_time < <(
+    BENCH_LINE="${line}" BASELINE_PATH="${BASELINE}" THRESHOLDS_PATH="${THRESHOLDS}" python3 -c "
+import json, os, re, sys
+line = os.environ['BENCH_LINE']
+def extract(pattern):
+    m = re.search(pattern, line)
+    return m.group(1) if m else ''
+ns = extract(r'([\d.]+)\s+ns/op')
+rss = extract(r'([\d.]+)\s+peak-RSS-bytes')
+if not ns or not rss:
+    print('Failed to parse benchmark metrics from output.', file=sys.stderr)
+    sys.exit(1)
+b = json.load(open(os.environ['BASELINE_PATH']))
+t = json.load(open(os.environ['THRESHOLDS_PATH']))
+print(ns, rss, b['ns_per_op'], b['peak_rss_bytes'], t['peak_rss_percent'], t['ns_per_op_percent'])
+"
+)
+
+if awk -v b="$baseline_rss" -v t="$baseline_ns" 'BEGIN {exit !(b==0 || t==0)}'; then
+    echo "Baseline contains zero values, cannot compute regression."
+    exit 1
+fi
+
+rss_change=$(awk -v cur="$current_rss" -v base="$baseline_rss" 'BEGIN {printf "%.1f", ((cur - base) / base) * 100}')
+time_change=$(awk -v cur="$current_ns" -v base="$baseline_ns" 'BEGIN {printf "%.1f", ((cur - base) / base) * 100}')
+
+baseline_rss_mb=$(awk -v val="$baseline_rss" 'BEGIN {printf "%.0f", val / 1048576}')
+current_rss_mb=$(awk -v val="$current_rss" 'BEGIN {printf "%.0f", val / 1048576}')
+baseline_secs=$(awk -v val="$baseline_ns" 'BEGIN {printf "%.1f", val / 1000000000}')
+current_secs=$(awk -v val="$current_ns" 'BEGIN {printf "%.1f", val / 1000000000}')
+
+echo ""
+echo "=== Benchmark Comparison ==="
+echo ""
+printf "%-20s %10s %10s %10s %10s\n" "Metric" "Baseline" "Current" "Change" "Threshold"
+printf "%-20s %10s %10s %9s%% %9s%%\n" "Peak RSS" "${baseline_rss_mb} MB" "${current_rss_mb} MB" "$rss_change" "$threshold_rss"
+printf "%-20s %10s %10s %9s%% %9s%%\n" "Execution time" "${baseline_secs}s" "${current_secs}s" "$time_change" "$threshold_time"
+echo ""
+
+failed=0
+
+rss_exceeded=$(awk -v change="$rss_change" -v thresh="$threshold_rss" 'BEGIN {print (change > thresh) ? 1 : 0}')
+time_exceeded=$(awk -v change="$time_change" -v thresh="$threshold_time" 'BEGIN {print (change > thresh) ? 1 : 0}')
+
+if [[ "$rss_exceeded" == "1" ]]; then
+    echo "FAIL: Peak RSS regressed by ${rss_change}% (threshold: ${threshold_rss}%)"
+    failed=1
+fi
+
+if [[ "$time_exceeded" == "1" ]]; then
+    echo "FAIL: Execution time regressed by ${time_change}% (threshold: ${threshold_time}%)"
+    failed=1
+fi
+
+if [[ "$failed" == "0" ]]; then
+    echo "PASS: No regressions detected."
+fi
+
+exit "$failed"
diff --git a/benchmark/stress/stress.go b/benchmark/stress/stress.go
index da7538484..ddb5d1fc7 100644
--- a/benchmark/stress/stress.go
+++ b/benchmark/stress/stress.go
@@ -165,6 +165,7 @@ func ec(dir string, components, workers int) func() {
 			strconv.Itoa(workers),
 			"--effective-time",
 			"2024-12-10T00:00:00Z",
+			"--allow-past-effective-time",
 		}); err != nil {
 			panic(err)
 		}
diff --git a/benchmark/stress/thresholds.json b/benchmark/stress/thresholds.json
new file mode 100644
index 000000000..660f6f93a
--- /dev/null
+++ b/benchmark/stress/thresholds.json
@@ -0,0 +1,4 @@
+{
+  "peak_rss_percent": 15,
+  "ns_per_op_percent": 20
+}