Imagine you're trying to read the dial on a bathroom scale without stepping on it too hard. You lean in just enough to see the number, but your lean itself nudges the needle. In superconducting quantum computing, dispersive readout is that lean — you probe the state of a transmon qubit via a coupled cavity, but the probing drive itself can kick population into levels you never wanted to touch. This paper builds a converged open-system model to quantify exactly how much dirt gets left behind, and to test where the standard rotating-wave approximation (RWA) breaks down. The committed claim: at a readout cavity occupation of 80 photons, the first excited leakage level (p2) settles at 1.02 × 10⁻³ and the next (p3) at 4.10 × 10⁻⁵, and these populations are persistent — they survive a 73 ns drive-off wait with less than 3% decay per level, following bare-level relaxation rates rather than cavity-photon decay. This is not transient dressing; it is genuine bare-state occupation deposited by the measurement process itself. The distinction matters because dressed-state artifacts would vanish when the drive turns off, while bare-level leakage lingers and contaminates subsequent gates. Methodologically, the model uses a displaced-frame treatment of the cavity quantized mode, with a nearest-neighbour RWA for both the drive and qubit-cavity coupling. Channel-removal controls isolate which dissipation channels produce which leakage: dephasing is the main source of p2, while relaxation drives p3 and above. These attributions hold across different bath models, though absolute population values shift by up to 15%. That robustness across bath models is a useful sanity check — it means the qualitative story isn't an artifact of a specific noise model. The paper then stress-tests the RWA drive truncation against the full charge-operator result in dissipation-free benchmarks. At switch-off for 80-photon occupation, the RWA underestimates p2 by about 10% and p3 by 24%, with discrepancies growing at deeper levels. One culprit: a parity-allowed non-nearest-neighbour charge matrix element the RWA drive operator simply drops. This is a concrete, named mechanism for the approximation's failure, not a vague hand-wave about higher-order terms. An attempt to go beyond RWA with bare-basis Lindblad dissipators fails a basic consistency test — it doesn't leave the ground state stationary at zero drive. A dressed-basis (secular) construction passes this test, which points the way forward but also marks the boundary of what this paper delivers. The beyond-RWA regime remains open territory. The practical upshot for the quantum computing community: if you're designing high-fidelity readout for transmon processors, you now have converged numbers for the leakage floor your readout protocol imposes. At 80-photon occupation, p2 ≈ 10⁻³ is not negligible for error budgets targeting 10⁻⁴ gate fidelities. And the RWA approximation you're probably using underestimates this by 10–24% depending on the level. These are the kinds of systematic errors that become load-bearing as hardware improves. This is a careful, methodical paper — not flashy, not claiming a breakthrough, but filling in a quantitative gap that the field needs filled. It identifies a persistent readout artifact, attributes it to specific channels, tests the dominant approximation's limits with named mechanisms, and acknowledges where the model stops being reliable. That's what converged simulation work looks like when done honestly.
