diff --git a/src/init.cpp b/src/init.cpp
index 9bd5e6616..1163f756e 100644
--- a/src/init.cpp
+++ b/src/init.cpp
@@ -1087,8 +1087,8 @@ static const int64_t g_nMinCoinCacheMB = 256; // never thrash below this working
 
 // Scheduled task: nudge nCoinCacheUsage toward "use all RAM except the reserve". If free RAM is below
 // the reserve we shrink the target (the next per-block flush releases the excess); if there is spare
-// RAM we grow it back toward the startup ceiling. Lock-free: it only reads system memory and writes
-// the aligned size_t threshold that the flush path reads.
+// RAM we grow it back toward the startup ceiling. nCoinCacheUsage is std::atomic<size_t>, so this
+// cross-thread write (vs the cs_main-held reads in FlushStateToDisk/VerifyDB) is well-defined, no lock needed.
 static void AdjustCoinCacheForMemoryPressure()
 {
     if (g_nMaxCoinCacheUsage == 0)
@@ -2072,7 +2072,25 @@ bool AppInit2(boost::thread_group& threadGroup, CScheduler& scheduler)
                 pcoinsdbview   = new CCoinsViewDB(nCoinDBCache, false, fReindex);
                 pcoinscatcher  = new CCoinsViewErrorCatcher(pcoinsdbview);
                 pcoinsTip      = new CCoinsViewCache(pcoinscatcher);
-                pnotarizations = new NotarizationDB(100*1024*1024, false, fReindex);
+                try {
+                    pnotarizations = new NotarizationDB(100*1024*1024, false, fReindex);
+                } catch (const std::exception& e) {
+                    // The notarizations (dPoW) DB is non-essential and node-regenerable. It has been seen to
+                    // snapshot/flush torn (0-byte log -> leveldb "Database I/O error" on reopen), which
+                    // previously aborted startup with a spurious "Error opening block database" and forced a
+                    // full resync. Wipe and recreate it instead of failing hard.
+                    LogPrintf("%s: notarizations DB failed to open (%s); moving aside and regenerating (non-fatal)\n", __FUNCTION__, e.what());
+                    // Move (do NOT delete) the old DB aside, so a transient open failure (fd
+                    // exhaustion, disk full, permissions) cannot permanently destroy notarization
+                    // history. If the recreate below also fails it propagates as fatal and the old
+                    // data survives in notarizations.corrupt for recovery.
+                    {
+                        boost::filesystem::path ndir = GetDataDir() / "notarizations";
+                        boost::filesystem::remove_all(ndir.string() + ".corrupt");
+                        boost::filesystem::rename(ndir, ndir.string() + ".corrupt");
+                    }
+                    pnotarizations = new NotarizationDB(100*1024*1024, false, true);
+                }
 
 
                 if (fReindex) {