Radiation

rad_assess

Correlated-error footprint, burst rate and logical-error floor for a chip in an environment: the one error class QEC handles worst, because surface codes assume independence.

Fidelity

reduced_order_chain_graded_against_G4CMP

Validation

reference

TWO PUBLISHED DEVICES, which is the unusual part: Google's gap-engineered Willow (arXiv:2506.18228) and Li et al.'s 63q sapphire device (arXiv:2402.04245). The phonon transport kernel is separately graded against G4CMP (Geant4 11.4.1) at 10.2% of distribution mass misplaced, against a 6.64% statistical noise floor.

measured

muon_burst_rate_per_hr

1.02x of the measured 9.3/hr at 0.3 mm, with NO free parameter: textbook sea-level flux and a known die area. But the die thickness is NOT published and this channel is steep in it (6.03x at 0.725 mm), so the pass is CONDITIONAL on the thin value.

gamma_burst_rate_per_hr

0.29x of the measured 41.4/hr at 0.3 mm, served as a BAND over the cached spectra. NO SINGLE THICKNESS SATISFIES BOTH CHANNELS (consistent=False), so the shortfall reads as a site flux difference of about 3.4x rather than a model error -- which is an inference about someone else's building, not a validated rate.

median_burst_qubits

15 of 60 measured, band 11-19, and the chain reproduces it. But it carries a 24% KERNEL-SHAPE SYSTEMATIC in a known direction at Willow's 0.5 mm transport thickness (linear between 0.381 and 0.6 mm): G4CMP's own kernel is more peaked near the strike (1.81x at 0.381 mm), and substituting it shrinks the median by that much, so served sizes are biased HIGH. Correcting toward the reference would put the median at or below the band's lower edge.

large_burst_share

0.054 measured, Poisson band 0.029-0.089. Same kernel-shape systematic as the median.

p90_burst_qubits

SERVED BUT NOT VALIDATED: no device publishes a measured p90. Its kernel-shape systematic at Willow's 0.5 mm is 20%, inside the +/-27% band the median is graded against.

footprint_systematic_by_thickness

the kernel-shape systematic depends on transport thickness: median / p90 overestimate 0% / 0% at 0.1 mm, 0% / 0% at 0.2 mm, 25% / 33% at 0.381 mm, 23% / 9% at 0.6 mm, 49% / 43% at 1 mm; none on thin dies, about double on 1 mm ones. Each footprint carries the figure for its own thickness.

kernel_vs_g4cmp

10.2% of distribution mass misplaced against a 6.64% statistical noise floor, so 1.5x the floor and not the 1.3x first recorded.

eta_trap_correction

375 G4CMP configurations, 90% conformal band +/-9.2%, which is the ALL-REGION number and not the tighter interior one.

note

The correlated floor served here is Google's MEASURED d=29 repetition-code floor (arXiv:2408.13687), not a prediction from the caller's design, and GOOGLE DOES NOT ATTRIBUTE THAT FLOOR TO RADIATION. Do not read this endpoint as explaining their floor.

Refuses when

• detection_threshold_keV omitted -> burst_rate refused; it is a DEVICE property and the chain cannot predict it
• trap geometry outside the eta box -> trap_correction refused rather than extrapolated
• thickness outside 0.05-2.0 mm -> burst_rate and footprint refused; inside it, interpolated between Geant4 runs and the response names the two runs
• unknown source -> HTTP 422 from the request model
• qubits off the die, duplicate positions, or n_qubits disagreeing with qubit_xy_mm -> footprint_qubits refused, with the die size a generated grid would need; never clamped. Any layout other than Willow's 60 measured sites reads BORROWED
• inputs away from a graded point -> that section reads calibration.status = BORROWED, naming the axis it departs on and the device whose normalisation is reused. NOT a refusal: the number is still returned.