Pillar D · State as of 2026-10-05

By how much does the winner win? A unit of measure for the advantage margin

As of October 5, 2026, Rosetta Quantum’s ledger adopts one unit for the margin of victory in any quantum-vs-classical comparison: the time-to-solution ratio at matched budget — classical time over quantum time, same instance, same declared quality floor, budgets sealed before the run — published with its uncertainty, plus a twin energy unit whenever power data exists. The measured record is why win/lose is not enough: Google’s 2019 “10,000 years” became 86.4 classical seconds against Sycamore’s 600 for the same 3 million samples, at higher fidelity (Zhao et al., arXiv:2406.18889); the Gaussian-boson-sampling margin shrank “nine orders of magnitude” by algorithms alone (Bulmer et al., Sci. Adv. 8, eabl9236, 2022); and Rosetta’s own quantum arm posted a +0.073 margin at p = 0.0949 — a margin that licenses nothing. From today, every new comparison row in the ledger will carry a ratio, not a medal.

→ Leer en español

State as of: 2026-10-05

Status as of: October 5, 2026.

What does a bare win/lose verdict hide?

The size of the win, the direction it is drifting, and the unit it was measured in — almost everything a decision actually needs. The first peer-reviewed history of the field classifies each advantage experiment as refuted or unrefuted (LaRose, Quantum 10, 2198, September 2026), and that binary is the right spine for a history — but a claim can keep its “unrefuted” label while the margin underneath it shrinks by orders of magnitude, and the label never moves. The paper that disciplined speedup claims a decade earlier (Rønnow et al., Science 345, 420, 2014) already warned, verbatim, that “quantum speedup is elusive and can depend on the question posed”. Rosetta Quantum’s own scoreboard is binary too: the advantaged-solve counter defined in our unit post still reads 0, and a counter says nothing about how far from the line the losing runs land, or how fast they are closing on the crossover point. This post fixes the unit under the counter: when a row moves, the ledger will say by how much.

What unit will the ledger print?

The time-to-solution ratio at matched budget: R = classical time-to-solution ÷ quantum time-to-solution, on the same instance, at the same declared quality floor, under a budget-parity table sealed before the run. R above 1.0 means the quantum side was faster; below 1.0, it lost; and a ratio will only count as advantage when its uncertainty interval clears 1.0 under a test and an α declared in advance. Time to solution is not our invention — it is the yardstick Rønnow et al. built their speedup definitions around when they benchmarked a 503-qubit D-Wave Two against classical annealers (Science 345, 420, 2014). What Rosetta adds is the denominator discipline. First, sealed parity: our own N=90 match ran under a parity manifest hashed before the run — 12 free parameters on the quantum side against 0 classical, seal PARITY-N90, hash 0a134c23 — and the anatomy of that match is what a denominator that was not starved looks like. Second, honest clocks: our hybrid audit covered a published pipeline whose QPU share of wall-clock was 0.7% (Lozano, arXiv:2605.17623) — a ratio whose clock quietly times only the quantum kernel is not a measurement, it is marketing with digits.

What does the measured record look like in this unit?

Inverted in one flagship case, eroded nine orders of magnitude in the other, and negative by the vendor’s own arithmetic in the quadratic class. Case one, random circuit sampling: Sycamore sampled its benchmark circuit in 200 seconds and the 2019 paper estimated 10,000 years for the classical equivalent (Arute et al., Nature 574, 505, 2019) — taken at face value, R ≈ 1.6×10⁹ (our arithmetic, declared, on those two published times). In June 2024, 1,432 A100 GPUs generated 3 million uncorrelated samples of the same task in 86.4 seconds against Sycamore’s 600 for that sample count, at a higher linear cross-entropy score — “7 times faster”, and in the authors’ words “the first unambiguous experimental evidence to refute Sycamore’s claim of quantum advantage” (Zhao et al., arXiv:2406.18889). In the unit: R ≈ 0.14 (86.4 over 600, our arithmetic on their published times). The full seven-year erosion is charted in our supremacy retrospective and we do not re-litigate it here.

Case two, Gaussian boson sampling, is the opposite shape and the honesty test for the unit: the margin survives, smaller. Simulating state-of-the-art GBS fell from a 600-million-year estimate to several months on a ≈100,000-core supercomputer — “a nine orders of magnitude improvement” over previous estimates, verbatim (Bulmer et al., arXiv:2108.01622; published as Sci. Adv. 8, eabl9236, 2022). A machine that runs in minutes against a simulation of ≈73 days still holds R somewhere around 10⁴–10⁵ (our arithmetic, declared, on the authors’ figures and the machines’ published run times) — still a quantum win on the clock, nine orders smaller than claimed, with its validation caveats filed in our boson-sampling post. A medal hides exactly this difference: “unrefuted” covers both a 10⁵ margin and a vanished one.

Case three is a projection with the sign already flipped, using the proponents’ own figures: a quadratic-speedup optimization run on realistic error-corrected hardware would take “roughly a day and a million physical qubits” where the classical rival needs “about four CPU-minutes” (Sanders et al., PRX Quantum 1, 020312, 2020) — R ≈ 1/360, our declared arithmetic, before any such machine exists. The ceiling logic for that whole class is in our not-going-to-accelerate post. And some reports yield no ratio at all: the t-VMC re-run of a 256-atom analog quench cost ≈2,100 GPU-hours against seconds of machine time for the same observable, but the figure is flagged by its own authors as not a classical lower bound (Wiersema et al., arXiv:2609.01719, September 2026; read in our analog-vs-universal post). When a ratio cannot be computed honestly, the ledger row will say “no ratio: asymmetric report” rather than invent one.

Case	Task	The margin, in the unit	Direction today	Source
Sycamore RCS, 2019→2024	random circuit sampling, 53 qubits	2019 estimate: 10,000 yr vs 200 s (R ≈ 1.6×10⁹, our arithmetic); 2024 rerun: 86.4 s vs 600 s → R ≈ 0.14	classical 7× faster, higher XEB	Arute et al., Nature 574, 505 (2019) · Zhao et al., arXiv:2406.18889
Gaussian boson sampling	photonic sampling, up to 100 modes	600M yr → ≈73 days classical: “nine orders of magnitude” erosion; surviving R ≈ 10⁴–10⁵ (our arithmetic)	quantum still faster; margin cut 10⁹×	Bulmer et al., Sci. Adv. 8, eabl9236 (2022)
Quadratic class on FT hardware	combinatorial optimization	“roughly a day” quantum vs “about four CPU-minutes” classical → R ≈ 1/360 (our arithmetic)	projected negative, proponents’ own figures	Sanders et al., PRX Quantum 1, 020312 (2020)
Analog quench, 256 atoms	quench dynamics observable	≈2,100 GPU-hours vs seconds — authors: not a lower bound	no ratio: asymmetric report, no sealed parity	Wiersema et al., arXiv:2609.01719 (2026)
Rosetta quantum arm, N=90	allosteric-pocket ranking	delta +0.073, p = 0.0949 vs pre-registered α = 0.05; crossings: 0	noise — licenses nothing	seal RQ-EXP-QMARGIN-001 (2026-08-03)
Energy, same RCS bench	kWh per 3M samples	13.7 kWh classical vs 4.3 kWh Sycamore (2024) → 0.29 kWh in system-level rerun	energy margin flipped sign within months	Zhao et al., arXiv:2406.18889 · arXiv:2407.00769

When is a margin just noise?

When the test that would separate it from the line was not declared before the run — or was declared, and failed. The worked example is ours, and it stays on the board on purpose: under sealed parity, Rosetta’s quantum arm beat its classical-only rival by +0.073 mean delta with 54 of 89 proteins improved — and the pre-registered paired test returned p = 0.0949 against a declared α of 0.05, so the margin licenses nothing and the crossing counter stayed at 0 (the full anatomy, seal RQ-EXP-QMARGIN-001). The same discipline applies to rates: four weeks into our prospective scoreboard the weekly sample is n = 5, and its two readings (3/5 and 2/5) carry Wilson intervals of 23–88% and 12–77% — 65 points wide, both still containing the 32.6% retrospective rate (our prospective-validation post). A margin whose interval straddles the line is a reading, not a result. So the unit ships with three conditions, all of them future-tense commitments of the ledger: the test, the α and the quality floor will be declared before the run; the interval will have to clear 1.0, not touch it; and the ratio will have to hold on re-measurement — the seven-year RCS chain above is the standing reminder that R moves when the classical side wakes up.

Is time the only unit worth printing?

No — and the same flagship bench just showed that the sign of a margin can depend on the unit. The 2024 GPU re-run that beat Sycamore’s clock 7× lost on the meter: it burned 13.7 kWh against the 4.3 kWh the quantum run consumed, both figures published by the re-run’s own authors (Zhao et al., arXiv:2406.18889) — a ≈3.2× energy margin for the quantum side, our arithmetic. Months later, a system-level simulation of an equivalent task brought the classical energy down to 0.29 kWh against that same 4.3 (Zhao et al., arXiv:2407.00769) — the energy margin flipped sign by engineering alone, ≈15× the other way, our arithmetic again. One bench, one year, two units, two signs; the full accounting is in our energy post. The ledger’s twin unit will therefore be energy per solved problem, printed whenever the power data exists — which, outside the RCS chain, it mostly does not.

What will govern from today?

Four forward rules, all effective October 5, 2026, all applying to new rows only. Every new quantum-vs-classical comparison row in Rosetta’s ledger will carry R with its interval and the hash of the parity table it ran under. Quarterly refreshes will re-measure the ratio, not just re-check the verdict. Rows where no honest ratio can be computed will say “no ratio: asymmetric report” — the Wiersema case above will be the template. And the advantaged-solve counter will stay the headline number of the house: the ratio is what will move underneath it, and the first R that clears 1.0 and holds will be published with the same prominence as the zeros have been.

What we don’t know

We don’t know whether one scalar survives every task class: for sampling benchmarks there is no verifiable solved instance, the quality floor is itself contested, and the 7× figure compares sample generation at a stated cross-entropy score, not a solved problem (Zhao et al., arXiv:2406.18889). We don’t know how the unit should price quality when one side wins time and the other wins fidelity — the 2024 re-run dominated both at once, so the declared-quality-floor rule exists in our book but has not yet been exercised on a hard case. We don’t know the energy margin for most of the record, because outside the random-circuit-sampling chain almost no one publishes kWh — the gap we flagged in our energy post is still open. And we don’t know whether our own +0.073 is a small real effect or noise; the pre-registered test cannot tell, and the re-match condition is already sealed in the anatomy post.