Maverick03511/prepqc-exploitation-benchmarks

PrePQC Exploitation: public benchmark observations

Four separate small experiment records from the PrePQC Exploitation source repository, run on 2026-10-08. Each row is an aggregate of the stated job, operation, or fixed-seed trials; rows are not raw per-shot/per-case records or independent replications. These tables support reproduction and plotting; they are not a training set, a leaderboard, evidence of quantum speedup, or evidence that production ML-KEM was broken. The subsets have different inputs and success criteria and must not be pooled into a single score. The IBM and Garnet jobs used different providers and devices.

Subset	Rows	What each row means	Primary local result
qpu_counts	8	One outcome in one IBM ibm_fez job	QPU result record
garnet_basis_counts	16	One outcome in one Open Quantum IQM Garnet job	Garnet result record
mlkem_acvp_counts	5	One OpenSSL/NIST sample-vector operation category	ACVP result record
toy_ring_lwe_counts	3	One fixed-seed metric from the tiny synthetic ring	Toy method and result

IBM QPU: observed and ideal outcome counts

The marked circuit produced 11 in 232/256 shots (90.625%). The no-mark control produced 11 in 62/256 shots (24.21875%). ideal_probability comes from exact statevector simulation of the published OpenQASM circuits, multiplied by 256 in the chart. The ideal values are computed, while observed_count comes from the IBM workload pages and public result record. The two jobs have different gate counts; there was one job per condition, no per-shot export, and no device calibration or uncertainty estimate. This is a two-qubit workflow check, not a PQC attack.

IQM Garnet: four basis measurements

The preregistered basis experiment used four public synthetic circuits, one 256-shot job per condition on Open Quantum's IQM Garnet Public Plan. The no-mark X circuit returned 00 in 247/256 shots, versus 74/256 for marked X; the predeclared contrast was 67.58 percentage points. The marked Z circuit returned 11 in 235/256 shots. ideal_probability is computed from the published OpenQASM circuits; observed_count is the provider-reported aggregate readout. The CSV retains all four outcomes for each job, while the chart displays the prespecified focal outcome. The marked and unmarked circuits differ by an oracle CZ; there is only one job per condition, and physical mapping, transpiled circuits, calibration, and per-shot records were unavailable. This supports a basis-dependent signal in those submissions, not a coherence or fidelity estimate, QPU advantage, or PQC attack. Cite Open Quantum for the provider data.

OpenSSL ML-KEM: pinned sample-vector comparison

Installed OpenSSL 3.5.6 matched 240/240 sample cases from the NIST ACVP-Server corpus at commit 975de31. The five rows partition that total: 75 key generation, 75 encapsulation, 30 decapsulation, and 30 for each key-check operation. See the method and input hashes. This is not ACVP certification, an implementation comparison with a second library, or key recovery.

Tiny synthetic ring-LWE: fixed-seed sanity check

For seed 20261008, a search over all 729 candidate secrets accepted 20/20 constructed real challenges, rejected 20/20 paired uniform controls, and uniquely recovered the constructed secret in 20/20 real challenges. The ring is Z_17[x]/(x^6+1) with three public pairs per trial. The source code uses full negacyclic multiplication. This deliberately tiny synthetic task is not ML-KEM and does not establish a practical attack.

Files and provenance

• data/*.csv: numeric tables with units and identifiers; no keys, raw NIST vector bytes, IBM credentials, or CIC-PQC_OAV rows.
• plots/*.png: standalone 1200 × 680 charts generated from the CSVs.
• manifest.json: SHA-256 hashes of the public source records, circuits, and generated tables.
• Table builder and Windows chart builder: reproducible source code. Both run locally; neither makes network requests. The table builder requires Python 3.11+ standard library. The chart builder uses the Windows .NET charting assembly.

Rebuild from a checkout of the source repository:

python tools/build_benchmark_dataset.py
.\tools\plot_benchmarks.ps1
python tools/build_benchmark_dataset.py --check
.\tools\plot_benchmarks.ps1 -Check

The charts describe specific, bounded observations. They should not be extrapolated to production cryptographic security, QPU scaling, device fidelity, or field performance.

Citation

If you reuse these tables or charts, cite the dataset URL and the exact Git revision you used, along with the relevant original source above.