3 open, 2 answered, 5 ruled out. Each one stated its falsifier before the work started.
What drives the ~1/hour correlated-error floor?
Is the residual ~1/hour burst rate behind a below-threshold processor's error floor mostly cosmic hadrons — so shallow underground sites fix it — or muons and environmental gammas, which they do not?
Ends if: If hadrons dominate events above 1 MeV across every plausible die thickness and mounting orientation, the underground-mitigation reading stands; record it and stop.
Where outsiders could help: Species-resolved Geant4 runs on public cryostat geometry, an audit of the published flux and attenuation figures, and the counting statistics on a handful of floor events.
Read more →How tight is the structural floor for dissipative cats?
The ideal-code floor is a valid lower bound on logical error for the dissipative cat but runs 6–14× low. Is there an effective gap that closes the distance?
Ends if: An effective gap for the dissipative cat fails to close the floor-to-rate distance across the κ₂/γ sweep, or is shown not to exist.
Where outsiders could help: Exact Liouvillian numerics (QuTiP) and open-systems theory.
Read more →What must a device paper report to be checkable?
Published superconducting devices reach a small fraction of the design rules a reader could check them against. Is the binding constraint reporting convention rather than physics?
Ends if: A minimal reporting set fails to lift median reach toward the saturated value across a large corpus, or lifts the counts without changing any verdict.
Where outsiders could help: Annotating published device papers against the rule set.
Read more →Do frequency-collision rules need gate direction?
Does the standard seven-condition collision taxonomy give the right count on working hardware when the coupling map is read as undirected?
Ends if: If the undirected and directed readings give the same violation count on measured IBM Eagle calibration data, direction does not matter and the rules stand as written.
Outcome: They do not agree: 27 violations undirected, 4 with real gate direction, on a working processor. One condition fires on 57 of 144 (39.6%) of a device's real gates, and is now off by default.
Read more →Can a pre-fab model predict radiation bursts?
Can a model built from the environment, deposited energy and phonon transport reproduce a published processor's measured burst rate and footprint without fitting to it?
Ends if: Each stage had a numeric gate set before it ran; any stage that misses its gate is reported as failed, not refit.
Outcome: Muon burst rate 9.5/hr against 9.3 measured with nothing fitted — but only at a 0.3 mm die, and no single thickness fits the muon and gamma channels together. Of the last five build stages, two passed their gates and two failed; the failures ship as published known misses rather than refits.
Read more →Feedforward cancellation of strike frequency shifts
Radiation strikes shift qubit frequencies coherently for about a millisecond. Can a feedforward frame correction, driven by a burst model, cancel the shift?
Ends if: If an existing, cheaper mechanism already removes the coherent part, or what remains is incoherent, there is nothing for feedforward to do.
Outcome: Killed before build. A spin echo in the code cycle already cuts the excess detection from 17% to 2% above background in published data, and what survives the echo is incoherent, which no frame update can fix.
A neural decoder on public Willow data
Does a neural surface-code decoder beat the classical decoders shipped with Google's public Willow dataset?
Ends if: If the best configuration does not beat correlated matching on a held-out patch, the program stops.
Outcome: It does not: the best configuration is 28.75% worse than correlated matching at d=5 (n=100,000). An earlier +13.24% was measured against plain matching mislabelled as correlated, and is withdrawn.
A neural surrogate for RSFQ cells
Does a large graph model predict RSFQ cell behaviour better than a classical baseline on the same data?
Ends if: If a gradient-boosted tree on the same inputs matches or beats it, the model is not worth serving.
Outcome: The tree won on every head, and the neural model had been predicting "works" for every design it was shown. Its headline figures are withdrawn and the tree is what ships.
Read more →Room-temperature spin registers
Is there a room-temperature spin platform where coupling and readout work at the same distance?
Ends if: If every candidate's coupling window and readout window fail to overlap, the direction is closed.
Outcome: All eight candidates died, six to prior art and two to a fabrication dependency. The coupling and readout windows sit a factor of 5–10 apart in distance.
Read more →Learned readout discrimination
Does a neural network read weak-measurement records better than a simple linear model?
Ends if: If a linear model on a handful of features matches the network, the learned approach adds nothing.
Outcome: A linear model on 8 time-binned means scored 0.820 held-out accuracy, against 0.805 for a network on the raw 64-sample record.
Read more →