In a distributed transactional cache with L1 enabled I can observe the following.
Prepare cache by adding an entry with Cache.put( k, v1 ).
1. Node B starts with adding a changed value. Cache.put( k, v2 )
2. Node B TxDistributionInterceptor.visitPrepareCommand flushL1Caches sends invalidations.
3. Node A calls Cache.get( k ) retrieves v1 and stores this value in L1.
4. Node B proceeds with transaction.
The result is that Node A answers subsequent Cache.get(k) with v1 and Node B answers with v2.

It seems the invalidation is either send to early or must be synchronized in some way with the transaction.

Cache config:
<namedCache name="entity">
<jmxStatistics enabled="true" />
<clustering mode="dist">
<stateTransfer fetchInMemoryState="false" timeout="20000" />
<async />
<l1 enabled="true" />
<hash numOwners="1"/>
</clustering>
<locking isolationLevel="READ_COMMITTED"
lockAcquisitionTimeout="15000" useLockStriping="false" />
<eviction maxEntries="10000" strategy="LRU" />
<expiration maxIdle="100000" wakeUpInterval="5000"/>
<storeAsBinary storeKeysAsBinary="true" storeValuesAsBinary="false" enabled="false" />
<transaction transactionMode="TRANSACTIONAL" autoCommit="false" lockingMode="OPTIMISTIC"/>
</namedCache>