The current list

Entry 1 · Local time-to-answer separation

Fermi-Hubbard dynamics on 120 qubits

A 60-site Fermi-Hubbard hardware workflow produced local charge, spin, and double-occupancy observables and was compared with local MPS and observable-specific Majorana calculations.

Measured comparison. The quantum execution proxy was 272.50x shorter than the local chi=256 MPS wall time for this declared instance.

Quantum result. The hardware produced a full 120-qubit observable profile; raw mean double occupancy was 0.22862549 and readout-corrected mean double occupancy was 0.23067722.

Classical baselines

Claim boundary

• This is a local time-to-answer result, not a reproduction of the paper's headline practical-advantage claim.
• The chi=256 MPS baseline did not establish full convergence.
• The fast Majorana route demonstrates that observable-specific classical methods can change the ranking.

Entry 2 · Paper-aligned local separation

Non-Abelian SU(2) hadron dynamics on 120 active qubits

A Loop-String-Hadron implementation follows a differential hadron signal on a 60-site lattice and compares the quantum route with local circuit-MPS checks and published tensor-network and Pauli-propagation baselines.

Measured comparison. Both the local circuit checks and the paper-native baselines show a substantial runtime separation under their declared timing definitions.

Quantum result. The local hardware route produced charge-sector and differential-observable data for the 120-qubit circuit family.

Classical baselines

Claim boundary

• Hardware-only time is not cloud wall time and excludes several service overheads.
• The local scalar normalization remains distinct from the tracker's published hadron scalar.
• The result supports runtime separation and circuit or sector validation, not an independent precision reproduction of every published observable.

Entry 3 · Local runtime lower bound

Operator Loschmidt Echo on 80 qubits

A tracker-compatible 80-qubit extension estimates an Operator Loschmidt Echo from finite computational-basis samples and compares the complete mitigated hardware action with a bounded tracker-linked BP-TN calculation.

Measured comparison. The incomplete bond-dimension-64 classical delta half alone exceeded the complete Fire Opal action by more than 2.75x on this machine.

Quantum result. The measured delta/delta0 OLE ratio was 0.74028847 +/- 0.01663657; all eight sample ratios were positive.

Classical baselines

Claim boundary

• The classical calculation did not converge and no matched-accuracy ratio was obtained.
• This is a tracker-compatible 80-qubit extension with N_init=8, not an official tracker instance or an N_init=500 reproduction.
• The observation is local and does not cover every classical implementation or optimized compute platform.

Entry 4 · Diagnostic only

Random Graph Sampling on 70 data qubits

A complete 70-data-qubit non-Clifford circuit was sampled on IBM hardware, alongside an independent 70+8-qubit stabilizer-verification workflow and local classical scaling studies.

Measured comparison. Hardware returned 256 samples in 19 quantum-seconds, while a local Aer fit projects about 6.89 million years for one 70-qubit sample; sample counts and output quality are not matched.

Quantum result. The complete circuit returned 256 samples. The separate checked dataset retained 4,519 of 184,320 shots and gave a graph-state-prefix point estimate of 0.01217; its predeclared one-sided 95 percent lower-bound test failed. The original restricted-access IBM Boston execution reported substantially stronger effective performance than this independently accessible Kingston reproduction.

Classical baselines

Claim boundary

• The Tracker result used restricted access to IBM Boston, whereas this independent reproduction used the available IBM Kingston route; backend access, physical mapping, and calibration window are therefore not matched.
• Boston produced a substantially stronger workload-level result, but its historical calibration and complete raw fidelity-analysis record are not public, so the result does not establish that Boston was universally better hardware than Kingston.
• The 70-qubit classical runtime is extrapolated from measurements ending at 12 qubits, not measured at full width.
• The quantum samples have no validated full-distribution fidelity, and the separate predeclared 95 percent stabilizer test failed.
• The post-hoc 75 percent lower bound is an exploratory sensitivity result, not 75 percent fidelity and not evidence of quantum advantage.

Entry 5 · Local runtime lower bound

QOS-inspired PBMC68k feature generation on 60 qubits

A frozen 60-qubit QOS-inspired feature map generated 627 measured features for real PBMC68k cells on IBM Fez, reached the strongest held-out point score, and completed far sooner than the bounded local MPS attempt for the same specified feature target.

Measured comparison. Hardware generated the complete 60-qubit feature result in 26 quantum-seconds while local MPS remained incomplete after 2,577 seconds: a kernel-time lower bound greater than 99.1x; the complete Fire Opal route retained a lower bound greater than 5.0x.

Quantum result. Held-out balanced accuracy was 0.53125 (17/32), compared with 0.50000 (16/32) for the predeclared linear baseline and 0.43750 (14/32) for RBF. The exact McNemar p-value against linear was 1.0 and the paired-bootstrap 95 percent interval was -0.1875 to 0.25.

Classical baselines

Claim boundary

• The MPS route did not converge, so the same feature target was specified but numerical feature equality at a matched error tolerance was not established.
• The 26 quantum-seconds value was read from the Fire Opal dashboard; the archived get_result payload omitted the quantum-seconds field.
• The greater-than-99.1x ratio compares QPU-only dashboard time with local MPS wall time; the broader submission-to-retrieval comparison is a lower bound greater than 5.02x.
• The inexpensive classical linear and RBF classifiers do not require simulation of the 60-qubit feature map, so this is not an end-to-end speedup over ordinary classical machine learning.
• The held-out test contains only 32 cells; the one-cell hardware lead is not statistically significant and does not establish general predictive advantage.
• This is a local result under declared hardware and classical resources, not a claim against every tensor-network method, compute platform, or future implementation.