The scheduler will govern your access to living compute

What the work claims

HPC-quantum-computing platforms expose several quantum processing units that differ in size, topology, native gates, and noise. Because errors compound along a compiled circuit, the same high-level circuit compiled for different devices produces different executables with different expected fidelities, the probability that the executed circuit returns the correct result. Choosing well is worth real accuracy, but the obvious way to choose well, compiling every circuit on every device and picking the best, costs k times D compilations for k circuits and D devices, and compilation is expensive.¹

The authors replace exhaustive compilation with prediction. A graph neural network estimates, before any device-specific compilation, the expected fidelity of each circuit on each available QPU. A Quantum Meta-Scheduler then assigns circuits using a composite score that blends predicted fidelity against each device's current load share, controlled by a single fidelity weight w between 0 and 1.¹ The circuit is compiled only for the device it is assigned to.

Measured against that design, the numbers are strong. On the held-out test split the predictor achieves a mean absolute error of 1.04 percent and a root mean squared error of 2.00 percent in expected fidelity, and the prediction-based dispatch path saves about 1.72 times the wall-clock time of exhaustive selection in the densest circuit bin, with the saving growing as devices are added.¹ This is a methods paper with a working system, not a position piece, and it should be read as such.

How it works

The fidelity estimator is a graph neural network trained on timed circuits whose labels come from compiling and scoring real circuits on the target devices. The scheduler sits in a continuous dispatch loop: at each tick it snapshots the shared queue and the per-device state, computes each device's load share, the fraction of all dispatched circuits that device has received, and assigns each circuit to the device maximizing w times the predicted fidelity minus (1 minus w) times the load share.¹

The w parameter is the interesting part. Because both terms are normalized to the same 0-to-1 scale, the ratio w over (1 minus w) acts as an explicit exchange rate: a device may absorb that much extra load share per unit of fidelity advantage before circuits are redirected to a noisier but emptier device. At w near 1 the platform is a quality maximizer that piles work onto the best hardware; at w near 0 it is a load balancer indifferent to fidelity. The authors show experimentally that individual circuits switch devices at different values of w, so the allocation morphs gradually across the whole range rather than collapsing at a threshold.¹

Two practical footnotes matter for anyone porting this design. First, the prediction errors are small on average but long-tailed: the maximum observed regret on an individual circuit is 0.331, so aggregate accuracy is not a per-circuit guarantee. Second, the training labels are not free: compiling and scoring roughly 12,000 circuits on all three devices cost 904.5 seconds one-off, amortized after about 91,000 dispatched circuits at the measured mean saving of 9.96 milliseconds per circuit.¹ The predictor is itself a piece of infrastructure with a maintenance bill.

Where a skeptic should push

The single most load-bearing assumption is that expected fidelity is a stable, learnable property of the pair (circuit, device). That holds when device noise is quasi-stationary and characterized by calibration snapshots. The paper's evaluation uses an emulated environment derived from real IQM devices, and the authors are careful to note that the predictor's worst errors fall on circuits whose true fidelities sit in the 0.1 to 0.4 range, exactly the band where assignment mistakes hurt most.¹ On real, drifting hardware the estimate degrades the moment the calibration does.

A second push: the paper optimizes execution fidelity and parallelism, and explicitly nothing else. Latency, energy, user priority, and cost do not enter the composite score. That is a reasonable scope limit for a systems paper, but it means the w knob governs a two-way trade that real platforms will immediately complicate, and each added term becomes another silent policy.

Third, keep the magnitudes in view. The 1.04 percent MAE and the roughly 1.7 times scheduling overhead saving are real but modest; the brute-force baseline is expensive, not prohibitive. The contribution is architectural, a clean separation between prediction, policy, and dispatch, more than a headline speedup.

What a scheduling knob means for tissue platforms

Why does a quantum systems paper belong in a conversation about computing on living neural tissue? Because the awkward resource problem is identical. Any platform that gives multiple users access to a pool of organoids on multi-electrode arrays faces exactly this: substrates that differ in health, maturity, and excitability; telemetry that predicts, at least probabilistically, which substrate will complete which experiment usefully; and an operator who must decide who gets the good tissue. At least one commercial organoid cloud already sells time-shared access to a small pool of organoids alongside dedicated tiers, which is to say the multi-tenant living substrate is not hypothetical.²

The opportunity is that fidelity-aware scheduling is the first concrete template for how to run such a pool without arbitrariness. The prediction layer doubles as a welfare sensor: a learned estimate of how reliably a substrate will perform an experiment is, from another angle, a rolling estimate of that substrate's physiological state. If platforms were required to keep allocation logs, the same telemetry that schedules work would yield a longitudinal health record of each organoid, auditable after the fact. That is a governance instrument that costs almost nothing extra, because the measurements already exist for scheduling.

The threat is the same machinery read from the tenant side. A quality-maximizing scheduler systematically starves substrates it predicts will perform poorly. On quantum hardware that is an efficiency decision. On living tissue it is an algorithmic triage decision: the organoid that predicts badly gets fewer runs, less stimulation, and earlier retirement, and no ethics review ever sees it because it is just a scheduling parameter. Set w high enough and the platform quietly rations its living components by predicted usefulness, a fate that would land on the least mature or least healthy tissue first. The knobs that quantum computing is now learning to expose and tune are exactly the knobs someone will eventually have to govern for wetware, and the paper's normalization of w as an ordinary hyperparameter is the cautionary part.

There is also a deeper mismatch that cuts against a naive transplant. Quantum noise is stochastic; tissue change is adaptive. A predictor trained on calibration data assumes the substrate's statistics are stable, and living neural tissue violates that assumption on principle, since stimulation and learning reshape the very properties being predicted. The scheduler's core assumption, stationarity, is the one assumption living substrates guarantee to break, which means the wetware version of this system needs a feedback loop the quantum version does not: predictions must be continuously revalidated against a substrate that remembers its history. Whether that is a solvable engineering problem or a fundamental ceiling is an open question, and it is the right question to ask before any organoid cloud advertises fidelity-aware allocation.

The bottom line

Established: a graph neural network can predict pre-compilation circuit fidelity with about 1 percent mean absolute error, and a single-weight scheduler can trade that fidelity against load balance transparently, at roughly 1.7 times lower dispatch overhead than exhaustive selection in the tested regime.¹ Hypothesis, stated here rather than in the paper: the same architecture will be applied to multi-tenant platforms built on living neural tissue, because the resource problem is the same and the commercial pressure will be identical. What would confirm it is the first wetware platform shipping health-aware allocation; what would break the analogy is evidence that adaptive substrates defeat the stationarity assumption so thoroughly that prediction-based triage is worse than round-robin. Either outcome would be worth knowing before the scheduler, not after it, becomes the governor.