diff --git a/benchmark/scenarios/api-endpoint.js b/benchmark/scenarios/api-endpoint.js
new file mode 100644
index 0000000..53c3aef
--- /dev/null
+++ b/benchmark/scenarios/api-endpoint.js
@@ -0,0 +1,31 @@
+'use strict';
+
+// The suite had no scenario shaped like an actual API endpoint: `grep -r "res.json\|req.params\|req.query"`
+// over benchmark/scenarios returned nothing. That is the most common request shape in an Express
+// application, and the one where the framework's own work is a meaningful share of the request
+// rather than a rounding error next to zlib or JSON.parse.
+module.exports = {
+    name: 'routing/api-endpoint',
+    path: '/api/users/42/posts?fields=id,title,author&limit=10',
+    setup(app, express) {
+        const apiRouter = express.Router();
+
+        apiRouter.get('/users/:userId/posts', (req, res) => {
+            const fields = String(req.query.fields || '').split(',');
+            const limit = Number(req.query.limit) || 0;
+            const items = [];
+            for (let i = 0; i < limit; i++) {
+                items.push({ id: i, title: `post ${i}`, author: req.params.userId });
+            }
+
+            res.json({
+                userId: req.params.userId,
+                fields,
+                count: items.length,
+                items
+            });
+        });
+
+        app.use('/api', apiRouter);
+    }
+};
diff --git a/benchmark/scenarios/body-json-4kb.js b/benchmark/scenarios/body-json-4kb.js
new file mode 100644
index 0000000..3e5aaf4
--- /dev/null
+++ b/benchmark/scenarios/body-json-4kb.js
@@ -0,0 +1,29 @@
+'use strict';
+
+// body-json-512kb is a stress case: at half a megabyte JSON.parse dominates and the framework is
+// about 1% of the request, so that row cannot move. A few KB is what an API actually receives, and
+// at that size the body plumbing around the parse is visible. Both rows are kept - one measures the
+// parser, this one measures getting the bytes to it.
+const PAD = 4 * 1024;
+
+module.exports = {
+    name: 'middlewares/body-json-4kb',
+    path: '/abc',
+    wrk: {
+        script: 'post-json-4kb.lua',
+        connections: 200
+    },
+    verify: {
+        method: 'POST',
+        headers: {
+            'Content-Type': 'application/json'
+        },
+        body: JSON.stringify({ n: 1, pad: 'x'.repeat(PAD) })
+    },
+    setup(app, express) {
+        app.use(express.json());
+        app.post('/abc', (req, res) => {
+            res.send(`${req.body.pad.length}`);
+        });
+    }
+};
diff --git a/benchmark/scenarios/high-concurrency.js b/benchmark/scenarios/high-concurrency.js
new file mode 100644
index 0000000..7733930
--- /dev/null
+++ b/benchmark/scenarios/high-concurrency.js
@@ -0,0 +1,20 @@
+'use strict';
+
+// Every other scenario runs at 50-200 connections, where connection handling is free. uWS's
+// advantage over node:http at high connection counts is structural - per-connection memory and
+// accept handling - rather than anything in the request path, and nothing in the suite exercised it.
+// The handler is deliberately trivial so the row measures connections and not work.
+//
+// Caveat: wrk shares the runner with the server, so at this connection count part of what is
+// measured is the generator. Two wrk threads on a 4 vCPU runner leaves the server two.
+module.exports = {
+    name: 'connections/high-concurrency',
+    path: '/ping',
+    wrk: {
+        threads: 2,
+        connections: 1000
+    },
+    setup(app) {
+        app.get('/ping', (req, res) => res.send('pong'));
+    }
+};
diff --git a/benchmark/scenarios/realistic-stack.js b/benchmark/scenarios/realistic-stack.js
new file mode 100644
index 0000000..d03c1c9
--- /dev/null
+++ b/benchmark/scenarios/realistic-stack.js
@@ -0,0 +1,43 @@
+'use strict';
+
+const helmet = require('helmet');
+const cors = require('cors');
+const cookieParser = require('cookie-parser');
+const morgan = require('morgan');
+
+// middlewares-100 measures 100 no-ops, which no application runs. This is the stack a typical
+// Express service actually mounts, at the depth it actually mounts it.
+// morgan writes to a sink rather than stdout so the row measures the formatting work every request
+// pays, not the runner's terminal. 'combined' is used because that is the production default, and
+// it includes the remote address, which both frameworks have to resolve their own way.
+module.exports = {
+    name: 'middlewares/realistic-stack',
+    path: '/profile',
+    wrk: {
+        script: 'realistic-stack.lua',
+        connections: 200
+    },
+    // keep these in sync with realistic-stack.lua: the verify request and the load request are
+    // separate definitions, so cors would otherwise be exercised only by one of them
+    verify: {
+        method: 'GET',
+        headers: {
+            'Cookie': 'sid=abc123; theme=dark',
+            'Origin': 'https://example.com'
+        }
+    },
+    setup(app, express) {
+        app.use(helmet());
+        app.use(cors());
+        app.use(cookieParser());
+        app.use(express.json());
+        app.use(morgan('combined', { stream: { write: () => {} } }));
+
+        app.get('/profile', (req, res) => {
+            res.json({
+                sid: req.cookies.sid || null,
+                theme: req.cookies.theme || null
+            });
+        });
+    }
+};
diff --git a/benchmark/wrk-scripts/post-json-4kb.lua b/benchmark/wrk-scripts/post-json-4kb.lua
new file mode 100644
index 0000000..36dda00
--- /dev/null
+++ b/benchmark/wrk-scripts/post-json-4kb.lua
@@ -0,0 +1,7 @@
+wrk.method = "POST"
+wrk.path = "/abc"
+wrk.headers["Content-Type"] = "application/json"
+
+local kb = 1024
+local pad = string.rep("x", 4 * kb)
+wrk.body = '{"n":1,"pad":"' .. pad .. '"}'
diff --git a/benchmark/wrk-scripts/realistic-stack.lua b/benchmark/wrk-scripts/realistic-stack.lua
new file mode 100644
index 0000000..070835f
--- /dev/null
+++ b/benchmark/wrk-scripts/realistic-stack.lua
@@ -0,0 +1,4 @@
+wrk.method = "GET"
+wrk.path = "/profile"
+wrk.headers["Cookie"] = "sid=abc123; theme=dark"
+wrk.headers["Origin"] = "https://example.com"