QECTOR Decoder Workbench

Windows build · v0.5.2 - Backend: qector-decoder-v3 v0.7.0

Professional quantum error-correction analysis platform built on qector_decoder_v3 (compiled Rust core + public Python layer). Features 16 decoders, 10 code families (including qLDPC and color codes), batch/streaming decode, hardware detection, a resilient self/auto-debug backend, a 56-tool MCP server, and multi-format documentation export.

A professional desktop application for quantum error correction (QEC) research, decoder evaluation, and reproducible benchmarking, running fully offline on Windows.

Table of Contents

• Overview
• Download & Quick Start
• System Requirements
• Features
• License
• Citation
• Related

Overview

The Windows portable build bundles its own Python runtime, the scientific stack, and the qector_decoder_v3 wheel. It runs with no external Python, pip, or internet connection: on first launch the bundled decoder wheel is activated into a per-user managed site automatically. No online update checks are performed — the app runs entirely from what ships in the box.

Honest posture (from the upstream QECTOR Decoder v3 manual): every LER/throughput/latency figure is hardware-, driver-, seed-, and workload-dependent simulation — regenerate before quoting. PyMatching is the speed leader on standard surface-code MWPM; QECTOR's exact BlossomDecoder reaches its logical error rate but is not faster. Every decoder always satisfies H·c == s (mod 2). This is a research/evaluation platform, not a real-time or fault-tolerant hardware decoder.

Download & Quick Start

👉 Download the latest release from this repository's Releases page.

1. Download QectorWorkbench-Portable.exe.
2. Double-click to launch. No installation, no admin rights, no internet connection required.

For the headless MCP server:

QectorWorkbench-Portable.exe --mcp     # 56-tool stdio MCP server (no display needed)

Runtime data (logs, exported documents) is written to your per-user data directory (%LOCALAPPDATA%\QectorWorkbench; override with QECTOR_DATA_DIR).

System Requirements

Requirement	Detail
OS	Windows 10 (64-bit) or later
Disk space	Sufficient for the bundled Python runtime, scientific stack, and decoder wheel (portable, single executable)
Internet	Not required — the backend is provisioned entirely from the bundled wheel on first launch
GPU (optional)	CUDA-capable NVIDIA GPU for accelerated batch decode; CPU-only operation is fully supported

Features

Code Explorer

Build codes from 10 families with configurable distance/parameter. Live properties display including qubits, checks, and code rate, plus a clean bipartite Tanner-graph view.

• Graphlike / topological: repetition, ring, rotated_surface, unrotated_surface, toric, heavy_hex, and hypergraph_product (a CSS code built from a repetition-code seed — graphlike, so all matching decoders apply).
• qLDPC: bicycle (parameter = circulant size) and bivariate_bicycle (the IBM BB code family — e.g. the [[72,12,6]] "gross" code — selected from verified presets). qLDPC codes are non-graphlike: use bp_osd, blossom, hybrid, or auto_router (the Diagnostics tab's resilient decode auto-falls-back to a compatible decoder).
• Color codes: color_code (triangular & 2D), decoded with the native colour-code decoder.

Decoder Lab

Test any of 16 decoders with configurable error rate and seed. View error, syndrome, and correction arrays. Every decoder is verified to satisfy H·c == s (mod 2). A Clear Decoder Cache button frees cached decoder instances between experiments.

Decoder	Notes
union_find / fast_union_find	Fast approximate decode; higher LER than exact MWPM (throughput lever).
blossom	Weight-optimal exact MWPM; matches PyMatching's LER but is not faster.
sparse_blossom	Region-growing, near-optimal (experimental); not exact.
sparse_blossom_radix_neighbors	Radix-neighbor variant of the region-growing decoder.
bp_osd	Belief propagation + ordered statistics for LDPC / quantum-LDPC codes; tunable via bp_method (exact|min_sum) and osd_order (0|1|2).
auto	Self-selecting decoder: picks the best available backend per problem size.
hybrid	Combines a fast heuristic pass with exact matching; trainable weights.
lookup_table	Precomputed syndrome→correction table with O(1) lookup (refused above 20 checks).
predecoded	Resolves easy/low-weight syndromes in a fast pre-decoding pass before matching.
auto_router	Policy decoder: inspects the code and dispatches the best concrete decoder — matching for graphlike codes, bp_osd for qLDPC. Universally compatible, including on bivariate_bicycle.
hybrid_cascade	Union-Find pre-filter with Blossom/BP-OSD escalation; exposes live cascade statistics. Graphlike codes only.
gnn_belief_matching	GNN-predicted per-qubit weights guide a weighted matching decode, with a faithfulness fallback to plain MWPM. Graphlike codes.
belief_matching	Sum-product BP posteriors reweight an exact Blossom matching step; faithfulness-checked with MWPM fallback.
two_stage	Two-stage decode pipeline (fast stage + exact escalation).
ambiguity_cluster	Ambiguity-clustering decoder for degenerate syndrome structure.
colour_code	Native color-code decoder (restricted to true color-code topologies).

Benchmark Suite

Run configurable benchmarks with throughput/latency metrics. Export results to JSON.

Batch & Streaming

Batch decode multiple error samples with success rate. Sliding-window streaming session controls.

Hardware Dashboard

Auto-detect CUDA, OpenCL, and CPU backends. System info with CPU/RAM utilization and hardware-optimized decoder recommendations. Note: the standard backend ships a CUDA path but no OpenCL kernels, so OpenCL reports unavailable unless the backend was built from source with the opencl feature.

Diagnostics (Self / Auto-Debug)

• Run Self-Diagnostics — environment/decoder/hardware self-test with per-check pass/warn/fail status.
• Probe Decoders — reports which decoders produce a valid (syndrome-verified) correction for the current code.
• Resilient Decode — decodes with an automatic multi-decoder fallback chain (verifying H·c == s at each step) and shows the full attempt trace.

Documentation Studio

Export professional documentation in Markdown, HTML, JSON, LaTeX, PDF (matplotlib multi-page) and SVG (standalone Tanner graph) with full provenance metadata.

MCP Server

56-tool Model Context Protocol server (stdio JSON-RPC 2.0) for programmatic access — including self_diagnostics, probe_decoders, resilient_decode, version_info, diagnostic_decode, native_recommend, native_streaming, list_codes, compat_report, hybrid-cascade stats, neural pre-decoder training, and decoder-option-aware decode tools. Every tool is wired to the real decoder backend, not a mock or stub.

Offline Versioning

• The window title and status bar show the workbench version (v0.5.2) and the installed decoder backend version (qector-decoder-v3 v0.7.0).
• No PyPI queries, no auto-updater: the shipped bundle is the single source of truth.

License

Workbench: source-available — see EULA (free use including commercial, with QECTOR watermark retention per EULA §2).

Backend qector-decoder-v3: separately licensed by Guillaume Lessard / iD01t Productions — free for personal/academic/non-commercial research; commercial use requires a paid commercial license (www.qector.store, admin@qector.store). Honor the backend license for any commercial deployment.

Citation

@software{lessard2026qectorworkbench,
  author  = {Guillaume Lessard},
  title   = {{QECTOR Decoder Workbench}: Windows Desktop Application for Quantum Error Correction Analysis},
  year    = {2026},
  version = {0.5.2},
  url     = {https://github.com/qectorlab/qector-decoder-workbench-windows},
  orcid   = {0009-0000-3465-3753}
}

Related

• PyPI Package: qector-decoder-v3 0.7.0
• Core Library: GuillaumeLessard/qector-decoder
• Commercial Licensing: qector.store

Author: Guillaume Lessard / iD01t Productions · ORCID 0009-0000-3465-3753 · www.qector.store