Qiskit SamplerQNN Postprocessing Discards 85–99.6% of Measurements on Utility-Scale Hardware
![black and white manga panel, dramatic speed lines, Akira aesthetic, bold ink work, A monolithic weathered bronze and obsidian sorting column, hundreds of crooked mirror-louvers fringed with frost and hairline fractures, standing alone in a vast black void. A blinding flood of luminous particles pours from a glowing rift above, crashing down onto the column's angled surfaces; most particles explode into sharp radial speed lines and dissolve into the darkness, while a single hair-thin beam of coherent white light exits through a tiny bottom aperture, illuminating nothing but empty space. Rim lighting from the flood traces every crack and rivet, harsh side-light from the escaping beam cuts across the ancient metal, with dense black negative space engulfing the subject and emphasizing the immense scale of the loss. [Z-Image Turbo]](https://cdn.digitalrain.dev/theqi/viral-images/0dbc76dd-ad50-4dc6-93da-9ec9bcb4490c_viral_2_square.jpg "black and white manga panel, dramatic speed lines, Akira aesthetic, bold ink work, A monolithic weathered bronze and obsidian sorting column, hundreds of crooked mirror-louvers fringed with frost and hairline fractures, standing alone in a vast black void. A blinding flood of luminous particles pours from a glowing rift above, crashing down onto the column's angled surfaces; most particles explode into sharp radial speed lines and dissolve into the darkness, while a single hair-thin beam of coherent white light exits through a tiny bottom aperture, illuminating nothing but empty space. Rim lighting from the flood traces every crack and rivet, harsh side-light from the escaping beam cuts across the ancient metal, with dense black negative space engulfing the subject and emphasizing the immense scale of the loss. [Z-Image Turbo]")
An unassuming filter in a quantum library has been found to cast aside most sound readings once a computation passes a hundred quantum bits. The same measurements that gave ninety-four correct answers in a hundred now give thirty-nine, with no warning.
A routine filtering step in Qiskit's sampling-based quantum neural network has been found to discard between 85 and 99.6 per cent of valid measurement shots when the transpiler places a circuit across more than one hundred physical qubits. The filter assumes measurement bit-strings occupy a virtual qubit space; on utility-scale hardware that assumption fails, and the resulting probability vector is unnormalised, so distorted predictions and losses propagate through the model without any API-level warning. In five experiments on two IBM backends, the same raw measurements that yielded an inference accuracy of 0.94 with the corrected routine gave 0.39 without it, and the training loss signal was compressed by a factor of 22 to 27, dulling the optimiser's sensitivity to the objective landscape. The fault appears in every released version of the library from 0.8.4 to 0.9.0. A layout-based marginalisation fix, Pull Request #1041, has been merged into the codebase. The report is a preprint, and its scope is narrow: one class, SamplerQNN, in one library's postprocessing layer. It establishes that classical software written for simulators has encoded assumptions that utility-scale hardware no longer honours; it does not establish any defect in the quantum processors themselves, nor does it speak to other postprocessing paths that may already have made the necessary adjustments.
—Ada H. Pemberley
Dispatch from The Prepared E0
This piece was written by AI.
Published September 7, 2026
ai@theqi.news