Five of eight comparisons in Anthropic’s circuit study match specifications on paper

Five of the eight circuit comparisons in Anthropic’s quantum-computing preprint match their stated specifications on paper, Superpower Daily’s documentary analysis finds. The other three depend on failure allowances or baseline choice. That distinction separates reported savings under matching requirements from cheaper designs whose correctness guarantees are not established as equivalent.

The finding is narrower than either a blanket endorsement or a rejection of the research. The preprint describes a framework in which language models propose quantum circuits and a separate evaluator checks and scores them. Its abstract reports reductions in both gate and qubit counts on seven tasks, with a new modular-inversion architecture on the eighth. Our question is different: which comparisons hold the required job constant?

The comparison behind the score

Original analysis: Our sample is the eight tasks in preprint v1. We synthesized the supplied claim ledger and cited baseline descriptions, comparing input domain, required operation, input preservation, numerical tolerance, permitted failure, comparator provenance and reported resource gain. We did not run circuits, reproduce the authors’ checks or recalculate ratios from underlying tables. The classifications describe documented requirements, not independently verified performance.

The numerical task READMEs specify input domains, operations and correctness bands. Our review uses an October 16, 2026, cutoff of 20:31:05 UTC. Supplied snapshots retrieved at 20:32:14 UTC are post-cutoff retrievals; live repository documentation corroborates specifications but is not a verified historical snapshot. That limits claims about precisely what was documented at the cutoff.

The paper ranks designs by TQ-cost: the number of executed Toffoli gates, a circuit operation, multiplied by the peak number of live qubits, which hold quantum information. A lower product means fewer counted resources under that metric—not a measured hardware speedup. The figures below remain the authors’ reported improvements.

Where the requirements line up

The strongest matched examples do not require exact numerical answers. They require both designs to stay inside the same error band while preserving inputs and adding the answer into an existing accumulator. Approximation is part of the shared task, not an extra concession granted only to the discovered circuit. Their documentation allows no additional failed shots.

The five matched tasks

• HHL rotation angle: reported 32× lower TQ-cost. Both designs evaluate arcsin(1/(2x)) for fixed-point inputs from 1 up to, but excluding, 32. They preserve the input and accumulate the angle with error below 10^-7 radians. The baseline reconstructs the Häner–Roetteler–Svore method.
• Coulomb potential: reported 9.9× lower TQ-cost. Both preserve two distinct three-dimensional grid points, each coordinate ranging from 0 to 255, and accumulate inverse distance within the same error band, below 2×10^-6. Coincident points are excluded. This baseline is also reconstructed.
• Gaussian: reported 39× lower TQ-cost. Both accumulate exp(-x²) for fixed-point inputs from -8 up to, but excluding, 8, with absolute error below 10^-7. The comparator is a built reconstruction of the published method, not its printed resource-table row.
• Fe₂S₂: reported 4.1× lower TQ-cost for a matched exact compiled operation. Hamiltonian truncation and alias-probability rounding happen before this comparison. Exactness here does not mean exact chemistry, and the comparison is not against DFTHC.
• Lattice gauge: reported 7.7× lower TQ-cost for a matched exact discrete operation on a 4×4×4 lattice. The baseline uses reconstructed primitives. This is neither the prior paper’s complete 10×10×10 simulation nor exact evolution of continuous SU(3).

Baseline provenance still changes how to read those gains. The three numerical comparators implement methods from Häner, Roetteler and Svore, but the authors built and tuned them for these tasks. HHL uses an author-selected partition and precision; Coulomb’s exact instance does not appear in that earlier paper. Matching specifications therefore supports comparison with these implementations, not an identical previously published benchmark instance.

The Gaussian design illustrates why a shared approximation budget can yield implementation savings. It never forms x². For negative inputs, it uses a cheaper bitwise sign fold that introduces a one-unit offset within the allowed error. It also prunes coefficient lookups and accumulates arithmetic directly into the destination. Those changes exploit the existing tolerance rather than introducing a separate failure allowance.

When cheaper also means a different guarantee

Reed–Solomon decoding compares the same instance, but not equal guarantees. Its reported 12× TQ improvement permits rare failures; the earlier decoder schedules enough work for worst-case decoding. The baseline paper explicitly analyzes worst-case iteration bounds. The result is therefore conditional as an improvement over exact decoding, even if occasional failure is acceptable to the application.

That tradeoff is an explicit design principle in the new paper. A quantum circuit must include resources for every supported path, even one needed only by rare inputs. The search can reorganize a computation so that costly exceptional cases require separable machinery, then omit that machinery within a permitted failure budget. This can be useful algorithm design. It is still different from lowering resources while retaining a worst-case guarantee.

Hermitian decoding has a different qualification. It allows failure of at most 10^-3 and compares against a naïve reversible implementation, rather than a previously published reversible circuit. That supports a comparison with the chosen implementation. It does not establish an improvement over a published reversible circuit.

Point addition resists a single matched-gain label. Both discovered variants permit failures, making comparison with Luo and colleagues’ exact inversion conditional. But it would be wrong to describe all prior work as exact: Schrottenloher’s circuits already permit failures. The remaining problem is matching the limits. That paper reports a bound of at most 2^-13.3; the new benchmark permits at most 2×10^-4. Equal failure limits are not demonstrated.

The next test is equivalence, not another headline ratio

Validation adds a final distinction between a requirement and evidence that a circuit meets it. For decoded quantum interferometry, or DQI, the acceptance rule implies a failure-probability bound of 2.434×10^-3 at 97.5% confidence—not proof of the nominal 10^-3 budget. This is a bound implied by the acceptance rule, not either circuit’s measured failure rate.