diff --git a/API.md b/API.md
index 6322be78..a132a956 100644
--- a/API.md
+++ b/API.md
@@ -662,6 +662,8 @@ On PostgreSQL, apply chunks as individual statements from the transport/client l
 - Monolithic payloads generated by [`cloudsync_payload_encode()`](#cloudsync_payload_encodetbl-pk-col_name-col_value-col_version-db_version-site_id-cl-seq).
 - Chunk-fragment payloads generated by [`cloudsync_payload_chunks()`](#cloudsync_payload_chunkssince_db_version-filter_site_id-until_db_version-exclude_filter_site_id).
 
+On PostgreSQL, concurrent merges of the same row are serialized before reading its clocks under `READ COMMITTED` (including `READ UNCOMMITTED`, which PostgreSQL treats identically). This also covers direct inserts into `cloudsync_changes`. Locks last until the caller's transaction ends and use a fixed pool of 256 keys per database, so unrelated rows can also wait for each other. `SERIALIZABLE` relies on PostgreSQL's conflict detection; retry the whole transaction on serialization failure (`40001`) or deadlock (`40P01`). Merging under `REPEATABLE READ` is refused with `0A000`, because waiting cannot refresh that transaction's snapshot. Transactions merging multiple rows can deadlock when acquiring locks in different orders.
+
 When a v3 fragment payload is received, CloudSync stores the fragment in an internal table and returns after applying zero or more completed values. Once the final fragment for a value is received, the completed value is validated and applied. Fragments can arrive in any order, and duplicate fragment delivery is idempotent. Applying a fragment never moves the receive checkpoint. On PostgreSQL, pieces of one value applied by concurrent transactions wait for each other under `READ COMMITTED`, and fail with a retryable serialization error under `SERIALIZABLE` when they conflict; a fragment is refused under `REPEATABLE READ`, where a transaction could miss a piece committed while it waited.
 
 **Parameters:**
diff --git a/docker/README.md b/docker/README.md
index 95777d65..02f262fd 100644
--- a/docker/README.md
+++ b/docker/README.md
@@ -247,6 +247,26 @@ EXECUTE FUNCTION bump_app_schema_version();
 
 ## Development Workflow
 
+### Reproducible PostgreSQL tests
+
+From the repository root, run:
+
+```bash
+./scripts/test-postgres-docker.sh
+# Issue #70 regression only (serial/concurrent merge, rollback, isolation, lock bound)
+./scripts/test-postgres-docker.sh 67_concurrent_merge.sql
+# Select another PostgreSQL image tag
+POSTGRES_TAG=15-bookworm ./scripts/test-postgres-docker.sh
+POSTGRES_TAG=18-bookworm ./scripts/test-postgres-docker.sh
+```
+
+The script builds the extension from the current source, creates an isolated container,
+runs psql with `ON_ERROR_STOP`, and removes the container and its volumes on exit.
+It does not publish ports or use an existing database. The image remains cached.
+The concurrency test uses dblink and checks that the second session is waiting on
+a lock before allowing the first to commit. On the code before the issue #70 fix,
+it fails with `expected higher/3, got lower/2`.
+
 ### 1. Make Changes
 
 Edit source files in `src/postgresql/` or `src/` (shared code).
diff --git a/scripts/test-postgres-docker.sh b/scripts/test-postgres-docker.sh
new file mode 100755
index 00000000..dab8a92a
--- /dev/null
+++ b/scripts/test-postgres-docker.sh
@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+# Disposable PostgreSQL build and tests; no host ports or persistent volumes.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+postgres_tag=${POSTGRES_TAG:-17}
+test_file=${1:-full_test.sql}
+if [[ ! -f "test/postgresql/$test_file" || "$test_file" == */* ]]; then
+    echo "Expected a SQL file name in test/postgresql" >&2
+    exit 2
+fi
+image="sqlite-sync-test:${postgres_tag}"
+container="sqlite-sync-test-$$"
+cleanup() { docker rm -f -v "$container" >/dev/null 2>&1 || true; }
+trap cleanup EXIT
+
+docker build --build-arg "POSTGRES_TAG=$postgres_tag" -t "$image" -f docker/postgresql/Dockerfile .
+docker run -d --name "$container" -e POSTGRES_PASSWORD=postgres \
+    -v "$PWD/test:/tests:ro" "$image" >/dev/null
+for ((i=0; i<60; i++)); do
+    if docker exec "$container" pg_isready -U postgres -d postgres >/dev/null 2>&1; then
+        # The image entrypoint briefly runs a temporary server during initialization.
+        if docker exec "$container" psql -h 127.0.0.1 -U postgres -d postgres -c 'SELECT 1' >/dev/null 2>&1; then
+            docker exec "$container" psql -U postgres -d postgres -v ON_ERROR_STOP=1 \
+                -f "/tests/postgresql/$test_file"
+            exit 0
+        fi
+    fi
+    sleep 1
+done
+docker logs "$container" >&2
+exit 1
diff --git a/src/cloudsync.c b/src/cloudsync.c
index c6da6f30..3c7ddcd1 100644
--- a/src/cloudsync.c
+++ b/src/cloudsync.c
@@ -2197,6 +2197,11 @@ int table_col_index (cloudsync_table_context *table, const char *col_name) {
 }
 
 int merge_insert (cloudsync_context *data, cloudsync_table_context *table, const char *insert_pk, int insert_pk_len, int64_t insert_cl, const char *insert_name, dbvalue_t *insert_value, int64_t insert_col_version, int64_t insert_db_version, const char *insert_site_id, int insert_site_id_len, int64_t insert_seq, int64_t *rowid) {
+    // Hold through the transaction, including deferred column writes. Lock before
+    // reading any row clocks, even when the row does not exist yet.
+    int lock_rc = database_merge_lock(data, table->meta_ref, insert_pk, insert_pk_len);
+    if (lock_rc != DBRES_OK) return lock_rc;
+
     // Handle DWS and AWS algorithms here
     // Delete-Wins Set (DWS): table_algo_crdt_dws
     // Add-Wins Set (AWS): table_algo_crdt_aws
diff --git a/src/database.h b/src/database.h
index 50cb621d..9e96c600 100644
--- a/src/database.h
+++ b/src/database.h
@@ -98,6 +98,7 @@ int database_commit_savepoint (cloudsync_context *data, const char *savepoint_na
 int database_rollback_savepoint (cloudsync_context *data, const char *savepoint_name);
 bool database_in_transaction (cloudsync_context *data);
 int database_fragment_lock (cloudsync_context *data, const char *value_id);
+int database_merge_lock (cloudsync_context *data, const char *table_ref, const void *pk, int pklen);
 int database_errcode (cloudsync_context *data);
 const char *database_errmsg (cloudsync_context *data);
 void database_log_warning (cloudsync_context *data, const char *message);
diff --git a/src/postgresql/database_postgresql.c b/src/postgresql/database_postgresql.c
index da4e55d6..708c3ba5 100644
--- a/src/postgresql/database_postgresql.c
+++ b/src/postgresql/database_postgresql.c
@@ -23,6 +23,7 @@
 // PostgreSQL SPI and other headers
 #include "access/xact.h"
 #include "catalog/pg_type.h"
+#include "common/hashfn.h"
 #include "executor/spi.h"
 #include "funcapi.h"
 #include "utils/array.h"
@@ -1212,6 +1213,31 @@ int database_fragment_lock (cloudsync_context *data, const char *value_id) {
     return (rc == DBRES_ROW) ? DBRES_OK : cloudsync_set_error(data, "cloudsync_payload_apply: unable to lock a fragmented value", rc);
 }
 
+// Serialize merge decisions for every column of a row, including tombstones and
+// blocks. READ COMMITTED takes a fresh snapshot on the subsequent clock reads.
+// SERIALIZABLE detects stale decisions itself; REPEATABLE READ cannot refresh its
+// snapshot after waiting and is refused, as for fragmented values above.
+int database_merge_lock (cloudsync_context *data, const char *table_ref, const void *pk, int pklen) {
+    if (IsolationIsSerializable()) return DBRES_OK;
+    if (IsolationUsesXactSnapshot()) {
+        int rc = cloudsync_set_error(data, "cloudsync merge cannot run under REPEATABLE READ, use READ COMMITTED or SERIALIZABLE", DBRES_MISUSE);
+        cloudsync_set_sqlstate(data, ERRCODE_FEATURE_NOT_SUPPORTED);
+        return rc;
+    }
+
+    // A fixed pool bounds lock-table use even for very large imports. Collisions
+    // only serialize unrelated rows. Include the qualified metadata table so all
+    // callers resolving the same table agree, regardless of their search_path.
+    uint32 bucket = (hash_bytes((const unsigned char *)table_ref, strlen(table_ref)) ^
+                     hash_bytes((const unsigned char *)pk, pklen)) & 255;
+    dbvm_t *vm = NULL;
+    int rc = databasevm_prepare(data, "SELECT pg_advisory_xact_lock(1129530963, $1::integer);", &vm, 0);
+    if (rc == DBRES_OK) rc = databasevm_bind_int(vm, 1, bucket);
+    if (rc == DBRES_OK) rc = databasevm_step(vm);
+    if (vm) databasevm_finalize(vm);
+    return (rc == DBRES_ROW) ? DBRES_OK : rc;
+}
+
 bool database_table_exists (cloudsync_context *data, const char *name, const char *schema) {
     return database_system_exists(data, name, "table", false, schema);
 }
diff --git a/src/sqlite/database_sqlite.c b/src/sqlite/database_sqlite.c
index cc1e0e78..359bbfbd 100644
--- a/src/sqlite/database_sqlite.c
+++ b/src/sqlite/database_sqlite.c
@@ -603,6 +603,11 @@ int database_fragment_lock (cloudsync_context *data, const char *value_id) {
     return DBRES_OK;
 }
 
+int database_merge_lock (cloudsync_context *data, const char *table_ref, const void *pk, int pklen) {
+    // SQLite already serializes writers, including the clock reads and writes.
+    return DBRES_OK;
+}
+
 bool database_table_exists (cloudsync_context *data, const char *name, const char *schema) {
     UNUSED_PARAMETER(schema);
     return database_system_exists(data, name, "table");
diff --git a/test/postgresql/67_concurrent_merge.sql b/test/postgresql/67_concurrent_merge.sql
new file mode 100644
index 00000000..e734a108
--- /dev/null
+++ b/test/postgresql/67_concurrent_merge.sql
@@ -0,0 +1,221 @@
+-- Issue #70: two merge decisions must not use the same stale row clocks.
+-- Run with psql -v ON_ERROR_STOP=1 -f test/postgresql/67_concurrent_merge.sql.
+-- dblink observes the waiter before releasing x; no timing-dependent overlap.
+\set ON_ERROR_STOP on
+\set testid '67-concurrent-merge'
+\ir helper_test_init.sql
+\connect postgres
+\ir helper_psql_conn_setup.sql
+DROP DATABASE IF EXISTS cloudsync_test_67;
+CREATE DATABASE cloudsync_test_67;
+\connect cloudsync_test_67
+\ir helper_psql_conn_setup.sql
+CREATE EXTENSION cloudsync;
+CREATE EXTENSION dblink;
+CREATE TABLE t(id TEXT PRIMARY KEY, v TEXT);
+SELECT cloudsync_init('t') AS _init \gset
+CREATE TABLE transport(name TEXT PRIMARY KEY, payload BYTEA);
+INSERT INTO t VALUES ('row', 'higher');
+INSERT INTO transport SELECT 'higher', cloudsync_payload_encode(tbl, pk, col_name,
+    col_value, 3, 3, decode(repeat('01',16),'hex'), 1, 0) FROM cloudsync_changes;
+UPDATE t SET v='lower';
+INSERT INTO transport SELECT 'lower', cloudsync_payload_encode(tbl, pk, col_name,
+    col_value, 2, 2, decode(repeat('02',16),'hex'), 1, 0) FROM cloudsync_changes;
+TRUNCATE t, t_cloudsync;
+CREATE FUNCTION apply_value(n TEXT) RETURNS INT LANGUAGE plpgsql AS $$
+DECLARE data BYTEA; BEGIN
+    SELECT payload INTO STRICT data FROM transport WHERE name=n;
+    RETURN cloudsync_payload_apply(data);
+END $$;
+CREATE FUNCTION assert_winner(label TEXT) RETURNS VOID LANGUAGE plpgsql AS $$
+DECLARE val TEXT; ver BIGINT; BEGIN
+    SELECT v INTO val FROM t WHERE id='row';
+    SELECT col_version INTO ver FROM t_cloudsync WHERE pk=cloudsync_pk_encode('row'::text) AND col_name='v';
+    IF val IS DISTINCT FROM 'higher' OR ver IS DISTINCT FROM 3::bigint THEN
+        RAISE EXCEPTION '%: expected higher/3, got %/%', label, val, ver;
+    END IF;
+END $$;
+SELECT apply_value('lower');
+SELECT apply_value('higher');
+SELECT assert_winner('serial lower then higher');
+TRUNCATE t, t_cloudsync;
+SELECT apply_value('higher');
+SELECT apply_value('lower');
+SELECT assert_winner('serial higher then lower');
+SELECT dblink_connect('x', format('dbname=%s user=%s application_name=cloudsync_67_x',current_database(),current_user));
+SELECT dblink_connect('y', format('dbname=%s user=%s application_name=cloudsync_67_y',current_database(),current_user));
+-- Initialize both worker contexts outside the contending transactions.
+SELECT * FROM dblink('x', 'SELECT apply_value(''higher'')') AS r(n INT);
+SELECT * FROM dblink('y', 'SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_exec('x', 'SET statement_timeout=''15s''');
+SELECT dblink_exec('y', 'SET statement_timeout=''15s''');
+CREATE FUNCTION wait_for_y() RETURNS VOID LANGUAGE plpgsql AS $$
+BEGIN
+    FOR i IN 1..500 LOOP
+        PERFORM pg_stat_clear_snapshot();
+        IF EXISTS (SELECT FROM pg_stat_activity WHERE application_name='cloudsync_67_y' AND wait_event_type='Lock') THEN RETURN; END IF;
+        IF dblink_is_busy('y')=0 THEN RAISE EXCEPTION 'y finished without waiting'; END IF;
+        PERFORM pg_sleep(0.01);
+    END LOOP;
+    RAISE EXCEPTION 'timeout waiting for y';
+END $$;
+-- Fresh and existing rows; both arrival orders; rollback must release the lock too.
+-- seeded=False, first=higher, rollback=False
+TRUNCATE t, t_cloudsync;
+
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''lower'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','COMMIT');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+
+SELECT assert_winner('seeded=False first=higher rollback=False');
+
+-- seeded=False, first=higher, rollback=True
+TRUNCATE t, t_cloudsync;
+
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''lower'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','ROLLBACK');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT apply_value('higher');
+SELECT assert_winner('seeded=False first=higher rollback=True');
+
+-- seeded=False, first=lower, rollback=False
+TRUNCATE t, t_cloudsync;
+
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''lower'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''higher'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','COMMIT');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+
+SELECT assert_winner('seeded=False first=lower rollback=False');
+
+-- seeded=False, first=lower, rollback=True
+TRUNCATE t, t_cloudsync;
+
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''lower'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''higher'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','ROLLBACK');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT apply_value('lower');
+SELECT assert_winner('seeded=False first=lower rollback=True');
+
+-- seeded=True, first=higher, rollback=False
+TRUNCATE t, t_cloudsync;
+SELECT apply_value('lower');
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''lower'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','COMMIT');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+
+SELECT assert_winner('seeded=True first=higher rollback=False');
+
+-- seeded=True, first=higher, rollback=True
+TRUNCATE t, t_cloudsync;
+SELECT apply_value('lower');
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''lower'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','ROLLBACK');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT apply_value('higher');
+SELECT assert_winner('seeded=True first=higher rollback=True');
+
+-- seeded=True, first=lower, rollback=False
+TRUNCATE t, t_cloudsync;
+SELECT apply_value('lower');
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''lower'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''higher'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','COMMIT');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+
+SELECT assert_winner('seeded=True first=lower rollback=False');
+
+-- seeded=True, first=lower, rollback=True
+TRUNCATE t, t_cloudsync;
+SELECT apply_value('lower');
+SELECT dblink_exec('x','BEGIN');
+SELECT * FROM dblink('x','SELECT apply_value(''lower'')') AS r(n INT);
+SELECT dblink_send_query('y','SELECT apply_value(''higher'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','ROLLBACK');
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT * FROM dblink_get_result('y') AS r(n INT);
+SELECT apply_value('lower');
+SELECT assert_winner('seeded=True first=lower rollback=True');
+-- REPEATABLE READ cannot refresh a snapshot after a lock wait. Preserve SQLSTATE.
+CREATE FUNCTION apply_checked(n TEXT) RETURNS TEXT LANGUAGE plpgsql AS $$
+BEGIN
+    PERFORM apply_value(n);
+    RETURN 'ok';
+EXCEPTION WHEN OTHERS THEN RETURN SQLSTATE;
+END $$;
+BEGIN ISOLATION LEVEL REPEATABLE READ;
+DO $$ BEGIN
+    IF apply_checked('lower') <> '0A000' THEN RAISE EXCEPTION 'expected REPEATABLE READ refusal'; END IF;
+END $$;
+ROLLBACK;
+
+-- SERIALIZABLE must abort the stale writer. Retrying then converges.
+TRUNCATE t, t_cloudsync;
+SELECT dblink_exec('x','BEGIN ISOLATION LEVEL SERIALIZABLE');
+SELECT * FROM dblink('x','SELECT apply_value(''higher'')') AS r(n INT);
+SELECT dblink_exec('y','BEGIN ISOLATION LEVEL SERIALIZABLE');
+SELECT dblink_send_query('y','SELECT apply_checked(''lower'')');
+SELECT wait_for_y();
+SELECT dblink_exec('x','COMMIT');
+SELECT state = '40001' AS serialization_ok FROM dblink_get_result('y') AS r(state TEXT) \gset
+SELECT * FROM dblink_get_result('y') AS r(state TEXT);
+SELECT dblink_exec('y','ROLLBACK');
+\if :serialization_ok
+\else
+DO $$ BEGIN RAISE EXCEPTION 'expected SQLSTATE 40001'; END $$;
+\endif
+SELECT apply_value('lower');
+SELECT assert_winner('SERIALIZABLE retry');
+
+-- The direct cloudsync_changes API uses the same protection. Applying many rows
+-- in one transaction must not allocate one advisory lock for each distinct PK.
+CREATE TEMP TABLE encoded_value AS SELECT col_value FROM cloudsync_changes WHERE tbl='t' AND col_name='v';
+BEGIN;
+INSERT INTO cloudsync_changes(tbl,pk,col_name,col_value,col_version,db_version,site_id,cl,seq)
+SELECT 't',cloudsync_pk_encode('bulk-' || i), 'v', e.col_value, 3, 3,
+       decode(repeat('01',16),'hex'), 1, i FROM generate_series(1,10000) AS g(i) CROSS JOIN encoded_value e;
+DO $$ DECLARE n INT; BEGIN
+    SELECT count(*) INTO n FROM pg_locks WHERE pid=pg_backend_pid() AND locktype='advisory'
+       AND classid=1129530963 AND objsubid=2;
+    IF n=0 OR n>256 THEN RAISE EXCEPTION 'unbounded or missing row locks: %',n; END IF;
+    IF (SELECT count(*) FROM t WHERE id LIKE 'bulk-%')<>10000 THEN RAISE EXCEPTION 'bulk rows missing'; END IF;
+END $$;
+ROLLBACK;
+DO $$ BEGIN
+    IF EXISTS (SELECT FROM pg_locks WHERE pid=pg_backend_pid() AND locktype='advisory' AND classid=1129530963)
+    THEN RAISE EXCEPTION 'row locks survived rollback'; END IF;
+END $$;
+\echo [PASS] (:testid) isolation errors, retry and bounded locks for 10000 rows
+
+SELECT dblink_disconnect('x');
+SELECT dblink_disconnect('y');
+\echo [PASS] (:testid) serial and concurrent merges converge on higher/3
+\connect postgres
+DROP DATABASE cloudsync_test_67;
diff --git a/test/postgresql/full_test.sql b/test/postgresql/full_test.sql
index fe21bd0f..a91ad568 100644
--- a/test/postgresql/full_test.sql
+++ b/test/postgresql/full_test.sql
@@ -73,6 +73,7 @@
 \ir 63_deep_savepoints.sql
 \ir 64_block_rewrite_leftovers.sql
 \ir 66_db_version_per_transaction.sql
+\ir 67_concurrent_merge.sql
 
 -- 'Test summary'
 \echo '\nTest summary:'