explainability_subcard.md

Explainability Subcard

Field	Response
Intended Task/Domain:	Quantum error correction (QEC) for triangular color-code memory circuits
Model Type:	Convolutional Neural Network (CNN) — 3D fully-convolutional neural network pre-decoder that predicts physical space-time corrections which sparsify the input detector syndromes into a residual syndrome
Intended Users:	QEC researchers, QPU builders, QPU operators
Output:	Predicted space-time physical corrections (X- and Z-basis detector predictions across the color partitions) that sparsify the input syndrome into a residual syndrome
Describe how the model works:	This model is a neural network pre-decoder for triangular color-code memory circuits. It operates as a pre-processing stage in front of any global color-code decoder (we use Chromobius, which is open source). The network ingests raw detector syndromes across space and time (space-time syndrome volume) over the color-code lattice, and outputs predicted physical space-time corrections that sparsify the input syndrome into a residual syndrome before the final decoder runs. By reducing active syndromes, the pre-decoder lowers the logical error rate and speeds up downstream decoding. Trained on circuit-level depolarizing noise using a superdense (nearest-neighbor) color-code circuit with Z feed-forward; training data generated on-the-fly with cuStabilizer (Stim is used only to parse circuits into our internal representation, not to generate data).
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:	Not Applicable
Technical Limitations & Mitigation:	Color code only — not expected to generalize to other code families (e.g., surface code) without retraining. Distance and rounds may exceed the receptive field (RF = 13) since the network is fully convolutional; training at distance == RF (13) gave the best performance. Performance may degrade on hardware noise profiles differing substantially from the circuit-level depolarizing training distribution (e.g., biased noise, coherent errors). Best used as a pre-decoder — it can run as a standalone decoder but performs poorly alone; pairing it with a global decoder (e.g., Chromobius) delivers the reported logical error rate (LER) reductions. fp16 precision — the released checkpoint is half precision (trained in fp32); further INT8/FP8 quantization is available via the NVIDIA ModelOpt export pipeline as an inference-time optimization.
Verified to have met prescribed NVIDIA quality standards:	Yes
Performance Metrics:	At least 2x LER reduction vs. Chromobius alone, with the largest gains at larger distances and lower physical error rates: ~4.1x at d=13 (p=0.001, X), ~13.9x at d=17 (p=0.002, X), ~36x at d=21 (p=0.002, X), and ~356x at d=31 (p=0.002, X). Measured on triangular color-code memory circuits under circuit-level depolarizing noise at p ~ 0.0006-0.006, n_rounds = d, both X and Z bases.
Potential Known Risks:	Silent degradation on out-of-distribution noise with no runtime warning. At the highest evaluated physical error rate (p = 0.006), the LER improvement narrows substantially; p = 0.006 is above the color-code threshold, which is not a practical QEC operating regime. Weak as a standalone decoder — without a downstream global decoder (e.g., Chromobius) its logical accuracy is poor; it must be paired with a global decoder to realize the reported gains. There is a risk that post-export INT8 or FP8 quantization adds numerical approximation that reduces logical error rate performance, with the full impact at low physical error rates not yet characterized (quantization is an inference-time step on top of the fp16 checkpoint).
Licensing:	OpenMDW-1.1 License