IBM Iron–Sulfur Cluster Quantum Benchmark Audit Reveals Spin-State Discrepancies

IBM’s recent claim that quantum sampling can outperform classical selected‑configuration interaction (CI) methods for iron‑sulfur cluster chemistry has come under renewed scrutiny. A new analysis of the raw hardware data released by IBM for the flagship 2025 Science Advances paper shows that the quantum samples converge to electronic states with significant spin contamination, with ⟨S²⟩ values ranging from 4.7 to 7.0 rather than the target singlet value of zero. The same spin‑purity problem appears in both the 2Fe‑2S and 4Fe‑4S benchmarks, where the quantum results are hundreds to over a thousand millihartree from the published reference energies and, in the case of the 4Fe‑4S cluster, converge to a triplet state (S = 1) rather than the intended singlet.

IBM’s mitigation strategy, which expands the sampled determinant space by swapping α and β spin strings and uses an internal spin‑square diagnostic, does reduce the energy error by less than a nanohartree while quadrupling the subspace size. However, the mitigation does not enforce spin purity, and the default solver pipeline does not set the spin‑square target flag. The issue was reported on the Qiskit SQD GitHub tracker (issue 337) and closed the same day, with a maintainer noting that the flag only alters the determinant basis, not the final state’s spin. Independent re‑analysis of IBM’s own shot data, processed through their pipeline, confirms that the 2Fe‑2S samples achieve low ⟨S²⟩ but remain 248 mHa from the reference, while the 4Fe‑4S samples converge to a clean triplet state 1,438 mHa above the reference.

The audit, performed by the author and reviewed by multiple parties, identified five defects in IBM’s own work, including the spin‑sector artifact that led to the retraction of the strongest pro‑quantum claim. The audit trail, reproducibility map, and detailed timelines are publicly available, and the author invites others to reproduce the results or examine the state’s spin properties. The findings suggest that, at the current level of hardware and software, quantum sampling has not yet surpassed classical methods for these benchmark systems, and that spin contamination remains a critical issue for quantum chemistry applications.