diff --git a/harnesses/gas-estimation/cmd/script/config.go b/harnesses/gas-estimation/cmd/script/config.go
index 49349a05..66dc49a7 100644
--- a/harnesses/gas-estimation/cmd/script/config.go
+++ b/harnesses/gas-estimation/cmd/script/config.go
@@ -133,7 +133,10 @@ func chains() []Chain {
 			Slug:         "arbitrum",
 			ChainID:      42161,
 			RealizedRPC:  envDefault("GAS_REALIZED_RPC_ARBITRUM", "https://arbitrum-one-rpc.publicnode.com"),
-			VerifyRPC:    envDefault("GAS_REALIZED_RPC_VERIFY_ARBITRUM", "https://arbitrum.drpc.org"),
+			// dRPC's arbitrum public routes to 1rpc.io which rate-limits within 30s.
+			// arb1.arbitrum.io is Offchain Labs' own hosted RPC — genuinely different
+			// upstream from PublicNode (Node Fleet), sustains 5+ req/s in tests.
+			VerifyRPC:    envDefault("GAS_REALIZED_RPC_VERIFY_ARBITRUM", "https://arb1.arbitrum.io/rpc"),
 			OwlracleSlug: "arb",
 			BlockTimeSec: 1, // Arb One has sub-second nominal block time; realizer catch-up loop handles the burst.
 			SupportedSet: []Oracle{OraclePublicNode, OracleOwlracle, OracleEtherscan, OracleMetaMask},
diff --git a/harnesses/gas-estimation/cmd/script/oracles.go b/harnesses/gas-estimation/cmd/script/oracles.go
index d733a1b3..8fa1b0e3 100644
--- a/harnesses/gas-estimation/cmd/script/oracles.go
+++ b/harnesses/gas-estimation/cmd/script/oracles.go
@@ -7,6 +7,8 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/url"
+	"os"
 	"strconv"
 	"strings"
 	"sync"
@@ -21,6 +23,39 @@ const (
 	httpTimeout = 8 * time.Second
 )
 
+// proxyClient is a rotating-IP HTTP client built at process start
+// from HTTPS_PROXY / HTTP_PROXY. Used only by pollers that need to
+// evade per-IP throttles (currently: Owlracle, whose 10 req/h guest
+// ceiling per IP made every 60 s poll from a single VPS collapse
+// into HTTP 403 within an hour). Keep-alives are disabled so every
+// request opens a fresh TCP conn and the rotating proxy (webshare)
+// assigns a new upstream IP per request. If no proxy env is set,
+// proxyClient stays nil and pollers fall back to the default client.
+var proxyClient *http.Client
+
+func init() {
+	raw := os.Getenv("HTTPS_PROXY")
+	if raw == "" {
+		raw = os.Getenv("HTTP_PROXY")
+	}
+	if raw == "" {
+		return
+	}
+	parsed, err := url.Parse(raw)
+	if err != nil {
+		fmt.Printf("[proxy] parse %q: %v\n", raw, err)
+		return
+	}
+	proxyClient = &http.Client{
+		Timeout: httpTimeout,
+		Transport: &http.Transport{
+			Proxy:             http.ProxyURL(parsed),
+			DisableKeepAlives: true,
+		},
+	}
+	fmt.Printf("[proxy] wired via %s\n", parsed.Host)
+}
+
 // pollResult is what every oracle client returns. TargetBlock is the
 // block these predictions apply to (oracle-specific: Blocknative
 // gives an explicit next-block number, feeHistory returns the
@@ -228,7 +263,12 @@ type owlResp struct {
 
 func pollOwlracle(ctx context.Context, ep OracleEndpoint) pollResult {
 	req, _ := http.NewRequestWithContext(ctx, "GET", ep.URL, nil)
-	body, status, err := httpDo(ctx, req)
+	// Owlracle's guest tier is 10 req/h per IP. Our 4 chains × 60s
+	// poll = 240 req/h from a single VPS IP, which collapses to a
+	// permanent HTTP 403 within an hour. Route through the rotating
+	// proxy pool so each request gets a fresh upstream IP and stays
+	// under the per-IP guest ceiling.
+	body, status, err := httpDoProxied(ctx, req)
 	if err != nil {
 		return pollResult{Err: err}
 	}
@@ -491,6 +531,27 @@ func httpDo(ctx context.Context, req *http.Request) ([]byte, int, error) {
 	return raw, resp.StatusCode, nil
 }
 
+// httpDoProxied routes the request through proxyClient if a proxy is
+// configured (webshare rotating pool via HTTPS_PROXY env). Falls
+// back to httpDo when no proxy is set — useful for local runs
+// without proxy access.
+func httpDoProxied(ctx context.Context, req *http.Request) ([]byte, int, error) {
+	if proxyClient == nil {
+		return httpDo(ctx, req)
+	}
+	req.Header.Set("User-Agent", "OpenChainBench/1.0 (+https://openchainbench.com)")
+	resp, err := proxyClient.Do(req)
+	if err != nil {
+		return nil, 0, err
+	}
+	defer resp.Body.Close()
+	raw, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, resp.StatusCode, err
+	}
+	return raw, resp.StatusCode, nil
+}
+
 func parseHexU64(s string) (uint64, error) {
 	s = strings.TrimPrefix(s, "0x")
 	if s == "" {
diff --git a/harnesses/gas-estimation/cmd/script/realized.go b/harnesses/gas-estimation/cmd/script/realized.go
index 52cbd35c..2e51cfe6 100644
--- a/harnesses/gas-estimation/cmd/script/realized.go
+++ b/harnesses/gas-estimation/cmd/script/realized.go
@@ -217,14 +217,26 @@ func runRealizer(ctx context.Context, buf *Buffer, chain Chain) {
 			lastObserved = head
 			return
 		}
-		// Catch up at most 5 blocks per tick to bound the per-tick
-		// work even if the realizer lagged behind. On Polygon /
-		// Avalanche (2s block time, 12s realizer cadence) we expect
-		// to see 6 new blocks per tick on average, so the catch-up
-		// loop intentionally trails reality by a block or two —
-		// that's fine because the buffer holds predictions for
-		// pendingTTLBlocks worth of blocks anyway.
-		for n := lastObserved + 1; n <= head && n <= lastObserved+5; n++ {
+		// Per-chain catch-up cap. ETH (12s blocks) needs 5, Polygon
+		// and Avalanche (2s blocks) need ~10, Arbitrum L2 (sub-second
+		// blocks) sees ~48 new blocks per 12s realizer tick and falls
+		// hopelessly behind at cap=5 (buffers spike to 99 predictions
+		// per oracle, most get evicted before grading). Cap scales as
+		// `max(5, ceil(realizedPollInterval / BlockTimeSec) + 5)` —
+		// one tick of blocks plus a small slack. Hard ceiling of 60
+		// so a stalled realizer can never fill the tick window with
+		// fetches on the busiest chain.
+		catchupCap := uint64(5)
+		if chain.BlockTimeSec > 0 {
+			c := uint64(int(realizedPollInterval.Seconds())/chain.BlockTimeSec + 5)
+			if c > catchupCap {
+				catchupCap = c
+			}
+		}
+		if catchupCap > 60 {
+			catchupCap = 60
+		}
+		for n := lastObserved + 1; n <= head && n <= lastObserved+catchupCap; n++ {
 			processBlock(ctx, buf, n, chain)
 			lastObserved = n
 		}