Answers are grounded in this paper and cite their sources.
Kretschmer, Grewal, DeCross, Gerber, Gilmore, Gresh, Hunter-Jones, Mayer, Neyenhuis, Hayes and Aaronson (arXiv:2509.07255) gave a one-way communication task in which a message of n qubits provably does more than a message of a larger number of classical bits, and ran it on a trapped-ion processor with 12 qubits, where the measured score stands five standard errors above the bound for every classical message of 61 bits. Alice is given the classical description of a Haar-random state of n qubits, Bob a uniformly random Clifford measurement; only Alice's message reaches Bob, who returns a string z, and the score 2^n p(z) - 1 is formed with the probability p(z) that her state gives to z in his basis. For a message of m classical bits the mean score over the inputs is bounded by an explicit function of m and n. A literature search on the day of the run found that experiment and no run of the task on a superconducting processor. This deposit holds a registered run on a public superconducting processor. On the IBM Heron processor ibm_kingston the task is run on open chains of 9, 10 and 11 qubits. Alice's part of a circuit is a brickwork of 112, 198 and 220 CZ gates with one-qubit gates whose angles are fitted to her state (fidelity 0.89, 0.84 and 0.61); the score is formed against the state she was given, so the misfit lowers it. Bob's part is the shortest circuit found for his basis by a search on the stabilizer tableau (19, 24 and 29 CZ on average), followed by the read-out. Alice's part ends in the frame in which it began, so that Bob's part is the circuit his basis gives, whatever Alice was given. The 500, 400 and 200 pairs of inputs were drawn before the registration from bits measured on the processor, combined by XOR with the operating system's generator; every pair is measured 500, 1000 and 1500 times. Nothing is done about errors: no dynamical decoupling, no twirling, no correction, no discarded shot. Lines, thresholds, inputs, circuits, chains and analysis were fixed and hashed before the session (registration SHA-256 0589ffc08b0304181aab4fdd808c08b8e4643ef92c230730d71f92b742cc63f8). Results, as they fell: 1100 circuits, 950,000 shots, 11 jobs submitted as one batch, 277 s of QPU time. The scores are 0.5594 +- 0.0035, 0.4267 +- 0.0027 and 0.2858 +- 0.0023 for 9, 10 and 11 qubits. All twelve binding lines and the four attestation lines held: the scores lie more than five standard errors above the bounds for classical messages of 9, 10 and 11 bits (53, 63 and 45 standard errors in all) and of 14, 15 and 17 bits, and the outcomes scored against the inputs of another pair give zero. The largest numbers of bits whose bounds lie five standard errors below the scores, reported without a line, are 16, 21 and 21. With two corrections made after the verdict (Student's t for the finite number of pairs, and a crude bound on the effect of the 16-bit resolution of the input states) they are 16, 20 and 20. In the model of the registration, a device whose outcomes carry nothing of Alice's input passes the lines in none of 4000 simulated sessions. What is claimed are the registered lines. Read as a statement about memory, namely that a classical system in the place of the n qubits would have needed more than so many bits, the result rests on the theorem of arXiv:2509.07255 and on assumptions that the data do not test: nothing but the n qubits carries information from Alice's part of a circuit to Bob's part, the inputs are uniformly random and were fixed before the session, and the shots of one circuit are independent repetitions. The trapped-ion experiment does not need the last assumption, because it takes one outcome per fresh pair of inputs; here the first shot of every pair alone gives 0.58 +- 0.06, 0.50 +- 0.07 and 0.38 +- 0.10, that is 3.4, 3.5 and 2.1 standard errors above the bounds for n bits. A test that does not use the normal distribution excludes 15, 17 and 12 bits. The separation is small against that of the trapped-ion experiment, and twice as many bits as qubits are reached at 10 qubits only. The score fell in the last two jobs of the session by 2 to 4 standard errors. Nothing here is a claim of computational advantage: the scores themselves are computed on a laptop. The deposit holds the registration, the module (inputs, circuits of both parties, rule for the chain, estimator, classical bound), the scripts for the random inputs, the plans, the looks, the registration with its Monte Carlo, the session and the verdict, the raw bits of the processor and the bytes of the operating system's generator from which the inputs were formed, the input files, the plans of the session and of two preliminary looks, the logs of the looks and of the session with the bits of every shot of every circuit, the verdict (formed once) and its ledger, a rehearsal of the session on the noiseless simulator under the hashed registration, the write-up (oneway_paper_v1.pdf) and the scripts and figures that build it. Licence: CC BY 4.0.
No takes yet. Share an insight, caveat, or question.
Bernd Westrick (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: