Model v3.0.0: honest error-correction units and qubit accounting
A primary-source audit of the noise model found three inconsistencies between how the error-correction literature reports its numbers and how the model consumed them. All three are now fixed; this is a MAJOR model version bump (2.5.0 → 3.0.0) because previously computed error-corrected scenarios return different (more honest) numbers.
Per-cycle vs per-logical-operation error rates
Both calibration sources report logical error per syndrome cycle (Google's Willow paper: ε_d is 'error per cycle of error correction'; the Bivariate Bicycle paper: 'p_L can be viewed as the logical error probability per syndrome cycle'). A lattice-surgery logical operation spans d syndrome cycles (surface code) or 11 cycles (gross code), so the surface-code path previously understated per-operation error by roughly a factor of d. For the gross code the old model made a second, partially compensating error (consuming the block-level rate as a per-qubit rate), so its net correction is small at full block occupancy — but up to 11× at low occupancy. The model now converts per-cycle rates to per-operation rates explicitly, with block-occupancy attribution, and exposes both quantities in the API response.
Run an error-corrected scenarioFull qLDPC physical-qubit accounting
The gross code entry previously counted only the 144 data qubits per block. Syndrome measurement requires 144 additional check ancillas ('requires n ancillary qubits... 288 physical qubits in total', Bravyi et al., Nature 2024), and logical operations add a 90-qubit logic processing unit (Yoder et al., Tour de gross, 2025). The catalog now counts all three: 378 physical qubits per block, 31.5 per logical qubit — consistent with how the surface-code path already counted ancillas.
Corrected suppression exponent for BB codes
The Bivariate Bicycle paper's own fit uses the circuit-level distance (d_circ ≤ 10 for the gross code), not the code distance (d = 12). The exponent is now d_circ/2 = 5 instead of 6, with the prefactor recalibrated to the published P_L(0.001) = 2×10⁻⁷. The two fits agree at p = 10⁻³ and diverge in extrapolation — the old exponent was ~10× optimistic at p = 10⁻⁴.
New regression fixture from an independent source
The test suite now pins Litinski's 'A Game of Surface Codes' worked examples (d = 13, ≈55,400 qubits, ≈4 h at p = 10⁻⁴; d = 27, ≈306,000 qubits, ≈7 h at p = 10⁻³) as a ground-truth fixture, alongside the per-cycle → per-operation conversion convention.
Every constant behind these fixes was verified against the primary papers (Bravyi et al. 2024, Yoder et al. 2025, Google Quantum AI 2024, Litinski 2019), and the derivations went through an independent adversarial review before shipping.
See the formulas and assumptionsSource: https://arxiv.org/abs/2408.13687