diff --git a/src/ping.rs b/src/ping.rs
index ba60250..46849c7 100644
--- a/src/ping.rs
+++ b/src/ping.rs
@@ -6,17 +6,18 @@
 //! the host answers pings — a different thing, often throttled or blocked —
 //! and would need raw sockets or an external `ping` binary.
 //!
-//! Flow: each round samples one random host from every segment, probes them
-//! concurrently, and keeps the ones that pass the filters — repeating until
-//! `count` valid IPs are gathered (or the `max_probe` budget is hit), then
-//! returns the lowest-latency ones.
+//! Flow: probe random hosts across all segments concurrently, keeping the ones
+//! that pass the filters, until `count` valid IPs are gathered (or the
+//! `max_probe` budget is hit). Each segment is sampled in proportion to how
+//! often it has yielded valid IPs, so persistently failing segments fade out
+//! and the budget flows to productive ones. Returns the lowest-latency IPs.
 
 use std::collections::HashSet;
 use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
 use std::time::{Duration, Instant};
 
 use anyhow::Result;
-use futures::stream::{self, StreamExt};
+use futures::stream::{FuturesUnordered, StreamExt};
 use ipnet::IpNet;
 use rand::Rng;
 use tokio::net::TcpStream;
@@ -122,49 +123,87 @@ pub async fn run(
 
     let mut collected: Vec<PingResult> = Vec::new();
     let mut probed = 0usize;
-
     let (port, times, to, concurrency) = (cfg.port, cfg.times, cfg.timeout, cfg.concurrency.max(1));
 
-    // Lazily generate fresh random hosts, round-robin across all segments. A
-    // continuous stream (vs discrete rounds) keeps `concurrency` probes always
-    // in flight, so a slow unreachable IP only blocks its own slot.
+    // Per-segment adaptive sampling. CF segments are wildly uneven — many are
+    // unroutable from a given vantage point. Each segment is drawn in
+    // proportion to its Laplace-smoothed valid-hit rate, so a segment that
+    // keeps failing is sampled less and the budget flows to productive ones.
+    let mut attempts = vec![0u32; nets.len()];
+    let mut hits = vec![0u32; nets.len()];
     let mut seen: HashSet<IpAddr> = HashSet::new();
     let mut rng = rand::thread_rng();
-    let mut idx = 0usize;
-    let candidates = std::iter::from_fn(move || {
-        let max_tries = nets.len().saturating_mul(50).max(1);
-        for _ in 0..max_tries {
-            let net = &nets[idx % nets.len()];
-            idx = idx.wrapping_add(1);
-            let ip = random_host(net, &mut rng);
-            if seen.insert(ip) {
-                return Some(ip);
+
+    // Keep `concurrency` probes in flight, refilling one as each finishes so the
+    // weights learned so far steer every new draw.
+    let mut inflight = FuturesUnordered::new();
+    for _ in 0..concurrency {
+        match draw_candidate(&nets, &attempts, &hits, &mut seen, &mut rng) {
+            Some((seg, ip)) => {
+                attempts[seg] += 1;
+                inflight.push(probe(seg, ip, port, times, to));
             }
+            None => break,
         }
-        None // every segment effectively exhausted
-    });
+    }
 
-    let mut probes = stream::iter(
-        candidates.map(|ip| async move { (ip, probe_ip(ip, port, times, to).await) }),
-    )
-    .buffer_unordered(concurrency);
-
-    while let Some((ip, outcome)) = probes.next().await {
+    while let Some((seg, ip, outcome)) = inflight.next().await {
         probed += 1;
         let status = cfg.classify(&outcome);
         let latency = (outcome.received > 0).then_some(outcome.avg_latency);
         if matches!(status, ProbeStatus::Ok) {
+            hits[seg] += 1;
             collected.push(PingResult { ip, latency: outcome.avg_latency });
         }
         on_progress(Progress::Probing { ip, status, latency, valid: collected.len(), probed });
         if collected.len() >= cfg.count || probed >= cfg.max_probe {
             break;
         }
+        // Refill the freed slot, biased by the updated weights.
+        if let Some((seg, ip)) = draw_candidate(&nets, &attempts, &hits, &mut seen, &mut rng) {
+            attempts[seg] += 1;
+            inflight.push(probe(seg, ip, port, times, to));
+        }
     }
 
     Ok(finalize(collected, cfg.count))
 }
 
+/// Choose a segment index with probability proportional to its Laplace-smoothed
+/// valid-hit rate `(hits+1)/(attempts+2)`. Untried segments start at 0.5 and
+/// fade toward 0 the longer they go without yielding a valid IP.
+fn weighted_segment(attempts: &[u32], hits: &[u32], rng: &mut impl Rng) -> usize {
+    let weight = |i: usize| (hits[i] as f64 + 1.0) / (attempts[i] as f64 + 2.0);
+    let total: f64 = (0..attempts.len()).map(weight).sum();
+    let mut pick = rng.gen_range(0.0..total);
+    for i in 0..attempts.len() {
+        pick -= weight(i);
+        if pick < 0.0 {
+            return i;
+        }
+    }
+    attempts.len() - 1
+}
+
+/// Draw a fresh, never-probed host, choosing the segment by weight. Returns the
+/// segment index and IP, or None if no new host turned up (segments exhausted).
+fn draw_candidate(
+    nets: &[IpNet],
+    attempts: &[u32],
+    hits: &[u32],
+    seen: &mut HashSet<IpAddr>,
+    rng: &mut impl Rng,
+) -> Option<(usize, IpAddr)> {
+    for _ in 0..64 {
+        let seg = weighted_segment(attempts, hits, rng);
+        let ip = random_host(&nets[seg], rng);
+        if seen.insert(ip) {
+            return Some((seg, ip));
+        }
+    }
+    None
+}
+
 /// Sort by latency ascending and keep the best `count`.
 fn finalize(mut results: Vec<PingResult>, count: usize) -> Vec<PingResult> {
     results.sort_by_key(|r| r.latency);
@@ -172,6 +211,19 @@ fn finalize(mut results: Vec<PingResult>, count: usize) -> Vec<PingResult> {
     results
 }
 
+/// Probe one IP and tag the result with its segment index, so the scheduler can
+/// credit the right segment. One named future type keeps `FuturesUnordered`
+/// happy (distinct async blocks would be distinct types).
+async fn probe(
+    seg: usize,
+    ip: IpAddr,
+    port: u16,
+    times: usize,
+    to: Duration,
+) -> (usize, IpAddr, ProbeOutcome) {
+    (seg, ip, probe_ip(ip, port, times, to).await)
+}
+
 /// Probe one IP `times` times, returning how many succeeded and the average
 /// latency over the successful probes.
 async fn probe_ip(ip: IpAddr, port: u16, times: usize, to: Duration) -> ProbeOutcome {