diff --git a/extract/padelnomics_extract/src/padelnomics_extract/proxy.py b/extract/padelnomics_extract/src/padelnomics_extract/proxy.py
index fc264a1..07acbbd 100644
--- a/extract/padelnomics_extract/src/padelnomics_extract/proxy.py
+++ b/extract/padelnomics_extract/src/padelnomics_extract/proxy.py
@@ -157,6 +157,13 @@ def make_tiered_cycler(tiers: list[list[str]], threshold: int, proxy_failure_lim
       per-proxy dead tracking removes broken individuals; tier-level threshold
       catches systemic failure even before any single proxy hits the limit.
 
+    Stale-failure protection:
+      With parallel workers, some threads may fetch a proxy just before the tier
+      escalates and report failure after. record_failure(proxy_url) checks which
+      tier the proxy belongs to and ignores the tier-level circuit breaker if the
+      proxy is from an already-escalated tier. This prevents in-flight failures
+      from a dead tier instantly exhausting the freshly-escalated one.
+
     Returns a dict of callables:
       next_proxy()          -> str | None  — URL from active tier (skips dead), or None
       record_success(proxy_url=None) -> None  — resets consecutive failure counter
@@ -174,6 +181,15 @@ def make_tiered_cycler(tiers: list[list[str]], threshold: int, proxy_failure_lim
     assert isinstance(tiers, list), f"tiers must be a list, got {type(tiers)}"
     assert proxy_failure_limit >= 0, f"proxy_failure_limit must be >= 0, got {proxy_failure_limit}"
 
+    # Reverse map: proxy URL -> tier index. Used in record_failure to ignore
+    # "in-flight" failures from workers that fetched a proxy before escalation —
+    # those failures belong to the old tier and must not count against the new one.
+    proxy_to_tier_idx: dict[str, int] = {
+        url: tier_idx
+        for tier_idx, tier in enumerate(tiers)
+        for url in tier
+    }
+
     lock = threading.Lock()
     cycles = [itertools.cycle(t) for t in tiers]
     state = {
@@ -245,6 +261,15 @@ def make_tiered_cycler(tiers: list[list[str]], threshold: int, proxy_failure_lim
             if idx >= len(tiers):
                 # Already exhausted — no-op
                 return False
+
+            # Ignore failures from proxies that belong to an already-escalated tier.
+            # With parallel workers, some threads fetch a proxy just before escalation
+            # and report back after — those stale failures must not penalise the new tier.
+            if proxy_url is not None:
+                proxy_tier = proxy_to_tier_idx.get(proxy_url)
+                if proxy_tier is not None and proxy_tier < idx:
+                    return False
+
             state["consecutive_failures"] += 1
             if state["consecutive_failures"] < threshold:
                 return False