The default --tpe-history-switch is unset: TPE never switches and retains the exact 0.3.0 estimator for the whole study. StudySpec accepts an absent or null sampler.config.history_switch; results and TPE terminal records write tpe_history_switch: null. A positive threshold explicitly enables bounded TPE.

Default identity gate

On AWS c6i.2xlarge, GCC 13.3, C++17 Release, the pinned HPO-BENCH replica compared every proposal ID, parameter, and hexadecimal objective value with 0.3.0 (6ccb6d4). Each problem uses instance 1, ordered batches of eight, 3,000 trials, and seeds 17, 48, and 79. The default is actually unset, not a large threshold.

Problem	Seeds	Compared proposal/value records
rotated_rastrigin20	3	9,000
ackley20	3	9,000
bbob_f15_d20	3	9,000
bbob_f21_d20	3	9,000
bbob_f24_d20	3	9,000
mixed_irrelevant15	3	9,000
rosenbrock20	3	9,000
rotated_ellipsoid20	3	9,000
Total	24 studies	72,000

The raw CSV, adjacent metadata, and SHA-256 sidecar record both binary hashes and every paired trace hash. All 48 traces and CSV/metadata hashes are verified on AWS. The measurement code revision is 1935897b2e2ba2bee1b782cd24c018c4b4058c5a; subsequent publication changes documentation and evidence only.

identity_replay.py defaults to this 24-study gate. Pass the pinned baseline and current identity executables via --baseline and --native, an output directory via --output, and optionally --jobs 8.

Bounded opt-in remains a tradeoff

For 10k–1M trials, --tpe-history-switch 1000 is a reasonable explicit opt-in when flat ask cost matters more than refinement quality. The unchanged round-three measurements give median-regret-ratio geomeans of 1.036492 at 3k versus 0.3.0 and 1.031570 at 10k versus the benchmark-only full-history variant: about 3.65% and 3.16% worse. The 3k rotated-ellipsoid ratio is 1.245558. These costs are accepted for opt-in, not asserted to be quality parity.

Opted-in 64D asks cost 656.450/663.061 microseconds at 100k/1M, exceeding the 375-microsecond W8 budget. The retained million-trial run takes 919.82 seconds with 21,152 KiB peak RSS. These flat ask/memory figures do not apply to the default full-history mode, whose memory and acquisition cost grow with history. Dense ordinal membership, streaming/output retention, and complete-line stops apply in both modes. No opt-in quality benchmark is rerun or tuned in this round.