2026-09-23 local rerun (one untimed warm-up, three timed runs per engine).
Command: python3 scripts/benchmark.py --accounts 5000 --operations 500 --repeat 3 --quiet
The harness restored identical fixtures before each run and verified parity before reporting speedup.

accounts=5000 operations=500 (plus 3 rejected edge cases)
legacy=4.723607s modern=0.040941s speedup=115.4x
identical outputs and account balances: {'PROCESSED': 500, 'REJECTED': 3, 'TOTAL_CENTS': 499999475}

These timings describe this deliberately inefficient synthetic COBOL baseline on this host. The earlier three-run measurement in benchmark-5000x500.txt was 128.1x; runtime varies between runs.