Benchmark evidence archive

Key tables

The headline table of each study, in the order the evidence is organised: the statevector grid first, then MPS and dynamic circuits, then execution modes. The complete set of 132 tables and 42 figures is reproduced on the draft paper page.

Grid: statevector study

Table (overall) - Reconstructed performance on 252 matched conditions per configuration

Pair ratios compare positive displayed times with SuperFermion; memory is whole-worker peak RSS. The 51 condition tables are on the draft paper page.

MPS and dynamic circuits

Table (mps_status) - Replacement MPS study outcomes by included configuration

Table (mps_mcm_summary) - Selected rows from the completed dynamic circuit observations

The four case rows are summarised; the full 14-row table and the ladder, stress, QFT and bond sweeps are on the draft paper page.

Execution modes

Table (modes_outcomes) - Sampling outcomes and available comparator standings

"Faster" means below every external comparator with a recorded timing in that condition; missing capped or budget jobs count as neither wins nor losses. The 24 condition tables and the accuracy checks are on the draft paper page.

The evidence bundles

Every notebook and archive is documented, with the provenance recorded for each run, on the Draft paper page.

The cat rows behind the stabilizer and Clifford rungs are tabulated separately, together with the reason the MPS suite has no cat lane, under The sf-cat backend, on its own.

Does the paper match the raw runs?

The paper integrates three studies. The desktop and laptop archives were replayed against the paper's saved records: every statevector record must fall inside the rounding bounds of its exact desktop job, and every execution-mode value must be identical to the laptop CSV.

15/15 checks passed. The machine-readable report is benchmark_consistency_report.json.

What is deliberately not comparable

• The three studies keep separate timing contracts: statevector jobs report the primary compute stage, execution-mode rungs report ABBA window means, and MPS jobs report their own contracts. No cross-study compute total is introduced.
• Mode::Stab and Mode::Cliff compile to the same Pauli-frame tableau; matching rows at equal circuit and qubit count are by construction.
• Mode::Dm is a documented hold-out in the cat crate; the density-matrix section measures the sf-ir simulate_dm engine and is never framed as a cat-mode win.
• cat lanes run 4-thread RAYON; competitor lanes run shipped defaults; cirq is capped at n = 64 and stabilizer Aer at n = 256 (not_run_capped rows carry no numbers).
• The desktop study is CPU-only ideal statevector work: no GPU, noise or hardware results.

Downloads

• Draft paper page — every notebook and archive in the table above, with the provenance recorded for each study.
• Consistency report — the 15 checks above in machine-readable form.
• Draft paper — PDF (107 pages, 1.3 MB) · Markdown · LaTeX source.