Does deterministic=False remove all output-ordering guarantees in tf.data parallel map?

In a tf.data input pipeline that applies Dataset.map with num_parallel_calls greater than one, setting deterministic=False is a documented way to trade input-order guarantees for throughput. The goal is to maximize pipeline speed while still understanding exactly which ordering properties, if any, remain enforceable for downstream consumers that zip, compare, or align datasets positionally.

The relevant uncertainty is version-sensitive. The deterministic argument defaults to True and can also be controlled globally through tf.data.Options().deterministic, but naming and accepted arguments have shifted across TF 2.x releases. A separate control, tf.config.experimental.enable_op_determinism(), forces deterministic execution of individual ops and does not by itself restore dataset ordering. Because of this split, it is unclear whether any partial ordering guarantee survives when dataset-level determinism is disabled, or whether the behavior is strictly version-dependent.

Does deterministic=False eliminate every output-ordering guarantee for a parallel map, or do partial guarantees remain in specific configurations? Which inspection of the installed TensorFlow build is needed to confirm the current default and accepted arguments for the deterministic option? How should a reader separate dataset ordering determinism from op-level determinism when validating a pipeline?