diff --git a/.github/workflows/benchmarks.yml b/.github/workflows/benchmarks.yml
index 29c5183..25b7fb9 100644
--- a/.github/workflows/benchmarks.yml
+++ b/.github/workflows/benchmarks.yml
@@ -37,7 +37,7 @@ jobs:
       - name: HS256 smoke regression (same gate as main CI)
         run: .venv/bin/python -m pytest tests/test_benchmark_jwt_libraries.py::test_benchmark_hs256_smoke_vs_competitors -q
 
-      - name: Run full comparison script
+      - name: Run full comparison script (default pem keys)
         run: |
           mkdir -p benchmark-results
           .venv/bin/python scripts/compare_jwt_libraries.py \
@@ -47,7 +47,17 @@ jobs:
             --warmup 50 \
             --markdown benchmark-results/ci-bench.md
 
+      - name: Run asymmetric comparison (cached competitor keys)
+        run: |
+          .venv/bin/python scripts/compare_jwt_libraries.py \
+            --algorithms RS256,EdDSA \
+            --iterations 200 \
+            --rounds 2 \
+            --warmup 50 \
+            --competitor-key-mode cached \
+            --markdown benchmark-results/ci-bench-cached.md
+
       - uses: actions/upload-artifact@v4
         with:
           name: benchmark-md
-          path: benchmark-results/ci-bench.md
+          path: benchmark-results/
diff --git a/README.md b/README.md
index 6428c45..a4b6624 100644
--- a/README.md
+++ b/README.md
@@ -129,6 +129,8 @@ python -m venv .venv
 
 The script covers HMAC, RSA, RSA-PSS, ECDSA, and EdDSA algorithms. Unsupported library/algorithm combinations are reported as `0` throughput. For a quicker smoke test, pass something like `--algorithms HS256,RS256,EdDSA --iterations 100 --rounds 1`.
 
+**Benchmark fairness:** the default `--competitor-key-mode pem` keeps pre-parsed `EncodingKey`/`DecodingKey` for OxyJWT while competitors often receive PEM bytes (see [Benchmarks](docs-site/docs/benchmarks.md)). For asymmetric comparisons, also run with `--competitor-key-mode cached`.
+
 Benchmark outputs are ignored by git because results depend on the machine, Python version, compiler flags, and CPU state.
 
 The default Rust crypto backend is `aws_lc_rs`, chosen for stronger performance on RSA and ECDSA in local benchmarks. You can still build with `rust_crypto` for comparison:
@@ -152,20 +154,19 @@ OxyJWT implements JWT/JWS signing and verification. JWE encryption is not part o
 
 See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup and pull request expectations. Report security issues privately via [SECURITY.md](SECURITY.md).
 
-## 🚀 Performance Benchmarks
+## Performance benchmarks
 
-OxyJWT is built for absolute speed. By bypassing the Python GIL and leveraging Rust's cryptographic primitives, it completely destroys standard Python libraries in both symmetric and asymmetric cryptography.
+OxyJWT is optimized for throughput on typical JWT workloads (especially HMAC). See [docs-site/docs/benchmarks.md](docs-site/docs/benchmarks.md) for smoke vs extended vs full workflows and key-preparation modes.
 
-Below is a performance comparison measured in **Operations per second (ops/sec)** (higher is better):
+The table below is a **historical snapshot** (default script settings, `pem` competitor keys). RS256 encode numbers are not comparable to `--competitor-key-mode cached`; re-run the script on your hardware before drawing conclusions.
 
-| Algorithm | Operation | ⚡ OxyJWT | PyJWT | Authlib | python-jose |
-| :--- | :--- | :--- | :--- | :--- | :--- |
-| **HS256** | Encode | **620,270** | 140,670 | 99,408 | 99,507 |
-| **HS256** | Decode | **361,073** | 109,272 | 94,823 | 51,838 |
-| **RS256** | Encode | **1,934** | 35 | 35 | 35 |
-| **RS256** | Decode | **58,752** | 27,200 | 26,085 | 23,046 |
-| **EdDSA** | Encode | **69,105** | 17,518 | 15,014 | N/A |
-| **EdDSA** | Decode | **31,666** | 10,741 | 10,317 | N/A |
-| **ES256** | Encode | **46,559** | 19,632 | 16,199 | 19,723 |
+| Algorithm | Operation | OxyJWT | PyJWT | Authlib | python-jose |
+| :--- | :--- | ---: | ---: | ---: | ---: |
 
-*Tested against standard Python ecosystem libraries. OxyJWT consistently dominates across all algorithms.*
\ No newline at end of file
+| **HS256** | Encode | 620,270 | 140,670 | 99,408 | 99,507 |
+| **HS256** | Decode | 361,073 | 109,272 | 94,823 | 51,838 |
+| **RS256** | Encode | 1,934 | 35 | 35 | 35 |
+| **RS256** | Decode | 58,752 | 27,200 | 26,085 | 23,046 |
+| **EdDSA** | Encode | 69,105 | 17,518 | 15,014 | N/A |
+| **EdDSA** | Decode | 31,666 | 10,741 | 10,317 | N/A |
+| **ES256** | Encode | 46,559 | 19,632 | 16,199 | 19,723 |
\ No newline at end of file
diff --git a/scripts/compare_jwt_libraries.py b/scripts/compare_jwt_libraries.py
index 8207737..cac1aa9 100644
--- a/scripts/compare_jwt_libraries.py
+++ b/scripts/compare_jwt_libraries.py
@@ -506,7 +506,11 @@ def parse_args() -> argparse.Namespace:
         "--competitor-key-mode",
         choices=("pem", "cached"),
         default="pem",
-        help="Use PEM bytes for competitors or preloaded cryptography key objects where supported.",
+        help=(
+            "pem (default): competitors get PEM str/bytes; OxyJWT uses pre-parsed "
+            "EncodingKey/DecodingKey. cached: preloaded cryptography objects for "
+            "fairer RSA/EC/EdDSA vs PyJWT/Authlib."
+        ),
     )
     parser.add_argument(
         "--algorithms",