diff --git a/crates/webclaw-cli/examples/perf_corpus.rs b/crates/webclaw-cli/examples/perf_corpus.rs
new file mode 100644
index 0000000..aaa4c02
--- /dev/null
+++ b/crates/webclaw-cli/examples/perf_corpus.rs
@@ -0,0 +1,188 @@
+//! Deterministic extraction micro-benchmark over a fixed HTML corpus.
+//!
+//!   cargo run --release -p webclaw-cli --example perf_corpus -- capture
+//!   cargo run --release -p webclaw-cli --example perf_corpus -- bench [iters]
+//!
+//! `capture` fetches a fixed URL list via the real FetchClient and saves the
+//! RAW html to /tmp/webclaw-bench/corpus (shared across baseline/fixed runs).
+//! `bench` reads that corpus and times extract() and to_llm_text() in-process,
+//! so the only variable between runs is the extraction code itself.
+
+use std::fs;
+use std::path::Path;
+use std::time::Instant;
+
+use webclaw_core::{extract, to_llm_text};
+use webclaw_fetch::{BrowserProfile, FetchClient, FetchConfig};
+
+const CORPUS: &str = "/tmp/webclaw-bench/corpus";
+
+const URLS: &[&str] = &[
+    "https://example.com",
+    "https://en.wikipedia.org/wiki/Rust_(programming_language)",
+    "https://en.wikipedia.org/wiki/Web_scraping",
+    "https://en.wikipedia.org/wiki/PostgreSQL",
+    "https://news.ycombinator.com",
+    "https://developer.mozilla.org/en-US/docs/Web/HTTP",
+    "https://www.rust-lang.org",
+    "https://blog.rust-lang.org/2024/09/05/Rust-1.81.0/",
+    "https://docs.python.org/3/library/asyncio.html",
+    "https://github.com/tokio-rs/tokio",
+    "https://doc.rust-lang.org/book/ch01-00-getting-started.html",
+    "https://www.gnu.org/licenses/agpl-3.0.en.html",
+    "https://old.reddit.com/r/rust/",
+    "https://arstechnica.com/",
+    "https://www.theverge.com/",
+    "https://crates.io/",
+    "https://www.cloudflare.com/",
+    "https://stackoverflow.com/questions/tagged/rust",
+    "https://www.postgresql.org/docs/16/index.html",
+    "https://go.dev/",
+    "https://nodejs.org/en",
+    "https://www.djangoproject.com/",
+];
+
+#[tokio::main]
+async fn main() {
+    let mode = std::env::args().nth(1).unwrap_or_else(|| "bench".into());
+    match mode.as_str() {
+        "capture" => capture().await,
+        "bench" => {
+            let iters: usize = std::env::args().nth(2).and_then(|s| s.parse().ok()).unwrap_or(60);
+            bench(iters);
+        }
+        "snapshot" => {
+            let label = std::env::args().nth(2).unwrap_or_else(|| "baseline".into());
+            snapshot(&label);
+        }
+        other => {
+            eprintln!("unknown mode '{other}' (use capture|bench)");
+            std::process::exit(2);
+        }
+    }
+}
+
+async fn capture() {
+    fs::create_dir_all(CORPUS).unwrap();
+    let config = FetchConfig { browser: BrowserProfile::Chrome, ..FetchConfig::default() };
+    let client = FetchClient::new(config).expect("build client");
+    let mut ok = 0;
+    for (i, u) in URLS.iter().enumerate() {
+        let name = format!(
+            "{:02}_{}.html",
+            i + 1,
+            u.replace("https://", "").chars().map(|c| if c.is_alphanumeric() { c } else { '_' }).take(40).collect::<String>()
+        );
+        match client.fetch(u).await {
+            Ok(f) if f.html.len() > 1000 => {
+                fs::write(Path::new(CORPUS).join(&name), &f.html).unwrap();
+                println!("OK   {name}  ({} bytes)", f.html.len());
+                ok += 1;
+            }
+            Ok(f) => println!("SKIP {name}  (thin {} bytes)", f.html.len()),
+            Err(e) => println!("FAIL {name}  ({e})"),
+        }
+    }
+    println!("--- captured {ok} docs into {CORPUS}");
+}
+
+/// Write canonical extraction output per corpus doc so baseline/fixed runs can be diffed.
+fn snapshot(label: &str) {
+    let outdir = format!("/tmp/webclaw-bench/snapshots/{label}");
+    fs::create_dir_all(&outdir).unwrap();
+    let mut files: Vec<_> = fs::read_dir(CORPUS)
+        .expect("corpus dir missing — run `capture` first")
+        .filter_map(|e| e.ok().map(|e| e.path()))
+        .filter(|p| p.extension().map(|x| x == "html").unwrap_or(false))
+        .collect();
+    files.sort();
+    let mut n = 0;
+    for path in &files {
+        let html = fs::read_to_string(path).unwrap_or_default();
+        if html.is_empty() { continue; }
+        let stem = path.file_stem().unwrap().to_string_lossy().to_string();
+        let url = format!("https://corpus/{stem}");
+        match extract(&html, Some(&url)) {
+            Ok(ex) => {
+                let json = serde_json::to_string_pretty(&ex).unwrap_or_default();
+                let llm = to_llm_text(&ex, Some(&url));
+                fs::write(format!("{outdir}/{stem}.json"), json).unwrap();
+                fs::write(format!("{outdir}/{stem}.llm"), llm).unwrap();
+                n += 1;
+            }
+            Err(e) => fs::write(format!("{outdir}/{stem}.ERROR"), format!("{e}")).unwrap(),
+        }
+    }
+    println!("snapshot '{label}': wrote {n} docs to {outdir}");
+}
+
+fn percentile(sorted: &[u128], p: f64) -> u128 {
+    if sorted.is_empty() { return 0; }
+    let idx = ((sorted.len() as f64 - 1.0) * p).round() as usize;
+    sorted[idx]
+}
+
+fn bench(iters: usize) {
+    let mut files: Vec<_> = fs::read_dir(CORPUS)
+        .expect("corpus dir missing — run `capture` first")
+        .filter_map(|e| e.ok().map(|e| e.path()))
+        .filter(|p| p.extension().map(|x| x == "html").unwrap_or(false))
+        .collect();
+    files.sort();
+    if files.is_empty() {
+        eprintln!("no corpus files in {CORPUS}");
+        std::process::exit(1);
+    }
+
+    println!("# perf_corpus bench  docs={}  iters={}", files.len(), iters);
+    println!("{:<42} {:>10} {:>10} {:>10} {:>10}", "doc(KB)", "extract_us", "llm_us", "p50_us", "p90_us");
+
+    let mut grand_extract = 0u128;
+    let mut grand_llm = 0u128;
+    let mut grand_total_p50 = 0u128;
+
+    for path in &files {
+        let html = fs::read_to_string(path).unwrap_or_default();
+        if html.is_empty() { continue; }
+        let url = format!("https://corpus/{}", path.file_name().unwrap().to_string_lossy());
+
+        // warmup
+        for _ in 0..5 {
+            if let Ok(ex) = extract(&html, Some(&url)) {
+                std::hint::black_box(to_llm_text(&ex, Some(&url)));
+            }
+        }
+
+        let mut ex_times = Vec::with_capacity(iters);
+        let mut llm_times = Vec::with_capacity(iters);
+        let mut total_times = Vec::with_capacity(iters);
+        for _ in 0..iters {
+            let t0 = Instant::now();
+            let ex = match extract(&html, Some(&url)) { Ok(e) => e, Err(_) => continue };
+            let t1 = Instant::now();
+            let txt = to_llm_text(&ex, Some(&url));
+            let t2 = Instant::now();
+            std::hint::black_box(&txt);
+            ex_times.push((t1 - t0).as_micros());
+            llm_times.push((t2 - t1).as_micros());
+            total_times.push((t2 - t0).as_micros());
+        }
+        ex_times.sort();
+        llm_times.sort();
+        total_times.sort();
+        let ex_p50 = percentile(&ex_times, 0.50);
+        let llm_p50 = percentile(&llm_times, 0.50);
+        let tot_p50 = percentile(&total_times, 0.50);
+        let tot_p90 = percentile(&total_times, 0.90);
+        grand_extract += ex_p50;
+        grand_llm += llm_p50;
+        grand_total_p50 += tot_p50;
+
+        let label = format!("{} ({}KB)", path.file_stem().unwrap().to_string_lossy(), html.len() / 1024);
+        println!("{:<42} {:>10} {:>10} {:>10} {:>10}", label.chars().take(42).collect::<String>(), ex_p50, llm_p50, tot_p50, tot_p90);
+    }
+
+    println!("---");
+    println!("CORPUS_PASS_P50_SUM_US extract={grand_extract} llm={grand_llm} total={grand_total_p50}");
+    println!("(lower is better; total = one full extract+llm pass over the whole corpus at p50)");
+}