# QECTOR Decoder v3 Full Context QECTOR Decoder v3 is a Rust-core Python quantum error correction decoder. Version 1.0.0 is documented by the reference manual deposited at https://doi.org/10.5281/zenodo.21941046. ## Core contract For a reachable syndrome `s = H e (mod 2)`, every supported backend is evaluated against: ```text H c = s (mod 2) ``` This puts `c + e` in `ker(H)`. Logical success requires the residual to be in `im(H^T)` or, for a full CSS workflow, in the correct logical-observable coset. Degenerate corrections must not be scored by raw vector equality alone. ## Backend map - Graphlike: Union-Find, Fast Union-Find, Blossom, Sparse Blossom, hybrid cascade, lookup tables, and CPU/GPU batch paths. - Non-graphlike: BP-OSD for arbitrary GF(2) matrices, with exact sum-product, min-sum, relay schedules, OSD-0, and OSD-W options. - Ambiguity: reliability partitioning, exact enumeration on bounded components, restricted OSD fallback, and research-grade learned predecoders. - Temporal: lifted space-time detector decoding plus Rust streaming and sliding-window primitives. - Routing: structural eligibility guard, backend recommendation, fallback verification, two-stage CSS feedforward, and native routing. ## Structural guard If any fault mechanism or qubit touches more than two detectors/checks, the problem is non-graphlike. Matching and Union-Find must not be forced onto it; BP-OSD is the documented general path. ## DEM and noise model The DEM pipeline parses error, detector, observable, shift, and repeat statements. Graphlike columns of weight at most two become edges; hyperedges route to BP-OSD. Parallel graph mechanisms can be collapsed with `p = p1(1-p2) + p2(1-p1)`. Edge weights use `log((1-p)/p)`. Space-time data and measurement mechanisms use separate weights. ## Evidence policy The manual excludes universal hardware-bound figures. Reports must identify code family, distance/size, rounds, noise model, DEM settings, decoder configuration, shots, seed, metric, environment, commit, raw JSON/CSV, and SHA-256. Logical error rates are scored in observable space and reported with Wilson confidence intervals. Code-capacity and circuit-level results are not directly comparable. Hot-path and cold-path timing are reported separately. Python allocation, process RSS, native heap, and VRAM are separate metrics. A skipped hardware-gated GPU test is not GPU evidence. ## Public reading path - Master AI Suite v2: https://qector.store/master-ai-suite - Blog index: https://qector.store/blog - Foundations: `/blog/01QecFoundations`, `/blog/11SteaneWorkedDecode`, `/blog/12RotatedSurfaceD3` - Decoder algorithms: `/blog/02BlossomMwpmExact`, `/blog/03SparseBlossomEventDriven`, `/blog/04FastUnionFind`, `/blog/05BpOsdQldpc`, `/blog/06AmbiguityGnnNeural` - Multi-round and routing: `/blog/07SpaceTimeDecoder`, `/blog/08AutoDecoderCascade`, `/blog/16DecoderSelection` - Models and evidence: `/blog/13DemPipeline`, `/blog/14WeightedDecoding`, `/blog/15LogicalErrorRates`, `/blog/17ReproducibleBenchmarks` - Systems and delivery: `/blog/09GpuAcceleration`, `/blog/10BenchmarksArchitecture`, `/blog/18RustPyo3Architecture`, `/blog/19ReleaseLicensing`, `/blog/20EcosystemIntegration` ## QECTOR Workbench (desktop GUI + MCP server) - Windows v1.0.1: 85-tool MCP server, qector-decoder-v3 1.0.0 backend, 17 decoder kinds, 10 code families, MCP protocol 2024-11-05. Portable exe + SHA-256 checksums. - Linux v1.0.1: 85-tool MCP server, 1.0.0 backend, 17 decoder kinds, 10 code families. AppImage + Debian packages. - Windows workspaces: Code Explorer, Decoder Lab, Benchmark, Batch & Streaming, Hardware, Diagnostics, Documentation, Lab & Personal Info, plus Console. Linux adds History. - Data directory: `~/.local/share/QectorWorkbench` (`QECTOR_DATA_DIR` overrides). Headless: `--mcp` (stdio JSON-RPC 2.0). - Releases: https://github.com/qectorlab/qector-decoder-workbench-windows/releases and https://github.com/qectorlab/qector-decoder-workbench-linux/releases ## QECTOR Claude Plugin (v1.0.2) - 28 domain skills (qector-core, qector-math-foundations, qector-developer, qector-researcher, qector-hardware-engineer, qector-educator, qector-sysadmin flagship; 21 more covering architecture, rust-pyo3, licensing, MCP workflows, and validation). - 5 specialized agents: qec-developer.md, qec-researcher.md, qec-validator.md, qec-sysadmin.md, qec-hardware-engineer.md. 13 slash commands. - 37 local MCP tools: 8 library (list_code_families, list_decoders, get_license_info, decode_syndrome, decode_single, threshold_sweep, build_code_from_matrix, compat_report) + 29 benchmark (DEM pipelines, Sinter/Stim, Wilson confidence, theorem lookup, compatibility, workbench probing). - Validation gate: `python scripts/run_manual_math_validation.py`. Repository: https://github.com/GuillaumeLessard/qector-claude-plugin ## Official pages - https://qector.store/decoder - https://qector.store/workbench - https://qector.store/claude-plugin - https://qector.store/mcp-server - https://qector.store/evidence - https://qector.store/manual - https://qector.store/technical-reference - https://qector.store/installer - https://qector.store/pricing - https://qector.store/commercial - https://qector.store/license - https://qector.store/about - https://qector.store/guillaume-lessard - https://qector.store/changelog - https://qector.store/contact ## Machine-readable - https://qector.store/llms.txt - https://qector.store/ai.txt - https://qector.store/feed.xml - https://qector.store/sitemap.xml - https://qector.store/.well-known/skills/index.json - https://qector.store/.well-known/mcp/server-card.json ## External records - DOI: https://doi.org/10.5281/zenodo.21941046 - PyPI: https://pypi.org/project/qector-decoder-v3/ - GitHub: https://github.com/GuillaumeLessard/qector-decoder - Workbench Windows: https://github.com/qectorlab/qector-decoder-workbench-windows/releases - Workbench Linux: https://github.com/qectorlab/qector-decoder-workbench-linux/releases - Claude plugin: https://github.com/GuillaumeLessard/qector-claude-plugin - ORCID: https://orcid.org/0009-0000-3465-3753 ## Install ```text pip install qector-decoder-v3==1.0.0 ``` ## Deployment posture The local Rust/Python CPU library is the preferred public path. GPU paths require controlled runtime and license configuration. REST, gRPC, MCP, and metrics surfaces are provisional and require authentication, authorization, TLS, rate limits, timeouts, audit logs, request limits, and resource quotas before production use. License verification is offline and the decoder makes no blocking network call during decoding.