Two Readers. One Writer.

Reports #010 and #011 showed that a correct earlier read is not enough to authorize a later write. Report #012 asks whether that rule survives contact with a real transactional database.

The unsafe control reproduced the duplicate.

Node A and Node B used separate PostgreSQL connections. Both first observed ABSENT / version=100. After a synchronization barrier, both executed an unconditional mutation and inserted an effect.

The PostgreSQL adapter bound the observation to mutation.

The safe workers carried the observed version into the database predicate:

UPDATE operations
SET state = 'committed',
    version = version + 1,
    updated_at = clock_timestamp()
WHERE id = 'op-1'
  AND state = 'absent'
  AND version = 100
RETURNING state, version;

Only a worker receiving a returned row could insert its effect, and that insert occurred in the same database transaction.

One connection won. The other became evidence.

In the canonical run, Node A won the conditional mutation:

Node B had already read the same version 100 snapshot, but by the time its conditional UPDATE was evaluated the row had changed:

That zero-row result was classified as PRECONDITION_FAILED, not as permission to retry blindly.

RECONCILE turned the failed write into current knowledge.

The stale writer rolled back its attempted transaction and reread the authoritative database state:

The final database invariant therefore held with exactly one committed effect.

TTP now has a real storage adapter.

The database-specific invariant is simple:

Interpretation boundary

This benchmark uses a real PostgreSQL 17.6 service container and two real independent client connections, but the business tables and effects are synthetic. It does not certify exactly-once delivery, cross-database atomicity, distributed consensus, failover behavior, every PostgreSQL isolation level, external APIs, payment rails, blockchains or arbitrary agent safety.

Verdict: two independent PostgreSQL readers saw the same valid snapshot. Unconditional writes created two effects. A version-bound conditional mutation created one winner, rejected the stale writer and preserved exactly one effect.