Noncomputational errors pose a significant challenge for quantum error correction (QEC). To address this, we have released Deltakit‑Stim, an extension of the Stim tool to enable the efficient simulation and modelling of these errors, narrowing a long‑standing gap in QEC tooling.
One important class of noncomputational errors is leakage, where a qubit exits the computational subspace {|0>,|1>} in an uncontrolled manner. Although leakage events may occur rarely, their impact can be severe: once a qubit has leaked, it can corrupt every qubit it subsequently interacts with until it relaxes back into the computational subspace. The result is a correlated cascade of errors across both space and time, violating a foundational assumption of most QEC schemes – namely, that errors occur locally.
Another important noncomputational error is atom loss, which is the unexpected disappearance of individual neutral atoms or ions from their intended traps. Atom loss is a dominant error mechanism in AMO-based qubit platforms (see, for example, arXiv:2506.20661), and its damaging effects similarly extend across the spacetime of a quantum computation.
Both leakage and loss can often be heralded – detected indirectly through measurement – but the precise spacetime location where the error originated is typically unknown. This distinguishes these noncomputational errors from the well‑understood erasure channel, where exact spacetime heralding can dramatically reduce QEC overheads.
Why noncomputational errors are hard to simulate
Despite their importance, simulating noncomputational errors such as leakage has remained challenging. While tools like Stim (Quantum 5 2021) have enabled tremendous progress by supporting fast, large‑scalesimulation of QEC circuits with circuit‑level noise, they focus on local, stochastic Pauli error models. This means that they cannot directly represent the highly correlated error processes induced by leakage or loss – precisely the processes that can dominate real error‑correction performance.
To address this, we have released Deltakit‑Stim, an extension of Stim that natively models noncomputational errors. It introduces explicit leakage and relaxation channels, allowing qubits to leave and return to the computational subspace at specified points in a circuit. In addition, Deltakit‑Stim supports a range of leakage transport mechanisms associated with two‑qubit gates, enabling leakage states to move around the system through qubit interactions. Noncomputational errors can also be heralded via an extension of Stim’s measurement infrastructure.
At present, the implementation supports a depolarising model of leakage (qubits interacting with leaked qubits incur a fully depolarising error), with additional noise models planned for future releases.
How it works under the hood
At the level of implementation, Stim is a fast Clifford simulator because its memory model and instruction set operate in parallel across many shots at once (i.e. many runs of a QEC circuit). The central idea behind our extension to Stim is to preserve its efficient shot-level parallelism while introducing the possibility of noncomputational states. These two goals are partly in tension: one would like to operate uniformly across many shots in parallel, yet noncomputational errors require specialised handling on a small subset of those shots. To address this, we augment the simulation with an additional register that tracks whether each qubit is inside or outside the computational subspace. Subsequent errors can then be conditioned on the state of this register and evaluated efficiently using Stim’s rare event iteration infrastructure (as discussed in https://algassert.com/post/2200).
In addition to shot‑level noise simulation, Deltakit‑Stim produces detector error models (DEMs) that represent conditional error mechanisms – an essential feature for capturing the correlated structure induced by leakage. DEMs enumerate individual error processes, mapping their probabilities to the detectors (deterministic products of measurements) flipped by those errors.
For purely stochastic Pauli noise, Stim can construct a DEM using a single reverse pass over the circuit, exploiting the fact that the detecting regions are known during reversal. Leakage complicates this picture: noncomputational errors are correlated through time, and their heralded effects can aggregate probabilistically in the DEM. This makes a single reverse pass insufficient.
To overcome this, Deltakit‑Stim builds a lightweight data structure during a fast forward pass that captures the noncomputational structure of the noise model. This structure is then used to extend the reverse pass, allowing the tooling to record the effect of conditional, time‑correlated errors with efficiency comparable to upstream Stim.
Enabling adaptive decoding
The resulting adaptive DEMs can be consumed by adaptive decoders such as Riverlane’s Local Clustering Decoder (Nature Communications 16, 2025), a hardware decoder implemented in FPGA. In these decoders, heralded leakage events trigger immediate updates to the decoder’s internal error‑model prior – represented in a processing grid distributed across the FPGA’s fabric – leading to substantially improvederror‑correction performance, see the paper for detailed numerics. An example of this adaptive response during a wiggled quantum memory experiment is highlighted in red in the illustration below.
Try it out
Deltakit‑Stim makes it possible to simulate and decode realistic QEC circuits affected by leakage and loss – without sacrificing performance or scalability. If you’re exploring noncomputational error models, interested in how QEC circuits perform in the presence of noncomputational errors or building adaptive decoders, we invite you to give it a try here.