Memory-benchmark fixture provenance
==================================================

longmemeval_oracle_synthetic.json is a SYNTHETIC, format-faithful fixture
authored for this repository's unit and integration-smoke tiers. It follows
the published LongMemEval instance format (question_id / question_type /
question / answer / question_date / haystack_session_ids / haystack_dates /
haystack_sessions / answer_session_ids; Wu et al., "LongMemEval:
Benchmarking Chat Assistants on Long-Term Interactive Memory", ICLR 2025),
but every question, answer, and conversation turn in it was written from
scratch for this fixture.

It contains NO excerpt of the LongMemEval dataset and is NOT attributed to
it; do not treat these three questions as LongMemEval data or compare
numbers produced from them with published LongMemEval results. If this
fixture is ever replaced with a real 3-question excerpt of the MIT-licensed
LongMemEval dataset, replace this NOTICE with the dataset's MIT attribution
§4.

The fixture is oracle-variant-shaped: each question's haystack contains its
evidence session(s) plus at most one distractor, so the offline tiers stay
tiny and cheap.
