Quantum simulation advantage stops being a one-paper verdict: QSim-Arena agents cut certified runtime 21 to 185-fold in a week

The verdict on quantum simulation advantage becomes a continuously updated process rather than a one-time paper conclusion. The system runs one loop on each side of the crossover, enters published methods as contestants, and registers physically meaningful simulation instances from the literature as benchmark tasks. Earlier crossover estimates fixed one portfolio of methods on the date they were written, and any technique published afterwards or omitted before left them out of date; reported classically hard experiments have also been repeatedly narrowed by stronger classical methods. The baseline is the fault-tolerant crossover study of Sun et al., adopted only after the implementation reproduced its resource table for a hundred-site Ising chain, and a second model family rebuilt the first fault-tolerant table without the code, with the disagreements it raised fixed.

Agents apply published techniques to these instances and verify the results, so the estimate moves within days rather than at the next paper. The three techniques that entered — a commutator-scaling error bound, exact norms in place of upper bounds on them, and a symmetry-aware readout exploiting that every term of the four Hamiltonians commutes with the product of all Pauli operators — were all already published, so each gain is labelled catch-up rather than new. At sixteen qubits (twelve for the hydrogen chain at H6) and a stated physical error rate, certified runtime fell by factors of 21 (Hubbard), 23 (hydrogen chain), 25 (XXZ) and 185 (Schwinger), and measured runtime by a factor of four on every family; one step went the other way, with Hoeffding-bound counting of certified shots raising every certified cell.

Correction is built into the process: no lane verifies its own result, an adversarial lane attacks results, and a second model family re-prices or re-reads without access to the producer's code. Earlier practice depended on the author of an error noticing it; here another lane finds it, and retractions, decisions and rejected proposals stay in append-only logs. Among the recorded corrections, most were caught by a lane other than the one that made them, and two moved the record against the system's own interest: rigorous shot counting raised every certified quantum cost, and a registered prior-art search turned a candidate new bound into catch-up.

The system runs no quantum hardware: every quantum cost is projected while every classical cost is measured, and the reported comparison covers four Hamiltonian families at eight to twenty qubits and H2 to H6. A crossover is claimed only where both sides have eligible board entries at the same tolerance, and a boundary is given as the pair of grid sizes that brackets it, never interpolated.

The self-correction record is a list of cases, not a rate: until faults with known answers are planted by someone outside the project, the fraction of errors it misses is unknown, and most checks were not blind to the work they checked. The techniques that moved the estimate were all published before the run began, so the response to a paper appearing mid-run is shown by design, not yet by a case. The band between the two regions has no width yet, because two uncertainties, the date of the classical portfolio and the accuracy extrapolation of measured step counts, so far take one value each.