开放量子纠错（QEC）测试方案 – 贪婪算法 vs. GE算法，均匀分布 vs. 聚类分布（k=4）

开放量子纠错（QEC）测试方案 – 贪婪算法 vs. GE算法，均匀分布 vs. 聚类分布（k=4）
Open QEC harness – greedy vs. GE, uniform vs. clustered k=4

原始链接: https://github.com/mrblakessinger-rgb/qec-evaluation-suite

本项目提供了一套纯数字化的诊断工具，用于评估有限规模的量子纠错（QEC），重点涵盖电路级表面码、Floquet 码以及全息张量网络模型（HaPPY）。本工具集侧重于遥测分析，而非生产级译码器或渐近阈值研究。 **主要组成部分包括：** * **分层评估：** 包括针对表面码和 Floquet 码的“冒烟测试”、几何实验室指标，以及针对 HaPPY 码贪婪恢复的特定消融研究。 * **性能指标：** 该套件量化了译码器性能，特别证明了贪婪恢复并非广义擦除（GE）恢复的子集，且该差距随深度增加而扩大。此外，它还评估了译码对均匀分布与聚类错误模式的抗性。 * **BP 微型实验室：** 包含针对置信传播（BP）和有序统计译码（OSD）的专业指标，区分了唯一可恢复性（GE）、码字匹配（BP）和伴随式有效补全（OSD）。该代码库提供可复现的 JSON 输出，优先使用预计算数据而非重复进行繁重的计算。为确保指标不被误用为通用的几何断言，项目在 `QEC_CLAIMS.md` 中强调了严谨的技术说明。本套件为开源项目，旨在支持严谨的小规模 QEC 诊断研究。

Hacker News 最新 | 往日 | 评论 | 提问 | 展示 | 招聘 | 提交登录 Open QEC harness – greedy vs. GE, uniform vs. clustered k=4 (github.com/mrblakessinger-rgb) Pachanoi 发布于 2 小时前，4 积分 | 隐藏 | 往日 | 收藏 | 讨论帮助指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系搜索：

原文

Finite-size diagnostic tooling for circuit-level and holographic-style quantum error correction.

Digital evaluation / telemetry only — not hardware, not a production decoder, not an asymptotic threshold paper.

Layer C framing: tensor-network model of bulk-boundary encoding with greedy recovery (HaPPY); optional matched-sample GF(2) GE ablation.

Claims detail: QEC_CLAIMS.md · Prefer JSON under public/data/ if anything disagrees with this README.

Layer	What it is	Primary artifacts
A Surface	Stim rotated surface; graphlike DEM; MWPM smoke	`public/data/layer_a_surface_smoke.json`
A Floquet	Period-3 honeycomb; schedule-valid; not graphlike MWPM peer of surface	`public/data/layer_a_floquet_smoke.json`
B Geometry	Educational graph invariants — not circuit-level QEC	`public/data/geometry_lab_metrics.json`
C HaPPY	Depth-1 (n=10, k=6) & capped depth-2 (n=20, k=16); greedy + matched GE	`public/data/happy_*.json`

Headline results (reproducible JSON)

Matched greedy vs GE (same erasure masks)

Depth	p_erase	Greedy central	GE central
1	0.3	0.664	0.304
2	0.3	0.59	0.0575

Non-dominance: GE is not a recovery superset of greedy on this instance. Gap widens at depth-2 under the linear face-parameter model.

Sources: happy_ge_ablation_metrics.json, happy_ge_ablation_metrics_depth2.json (seed=42).

Pattern stress × GE (depth-1, k-matched)

mode	k	Greedy central	GE central
uniform	4	~0.27	~0.02
clustered	4	1.0	0.0

k=4 stays special for greedy (clustered holds, uniform drops). Under GE both modes sit near the floor. Decoder × pattern, not a universal geometry claim.

Source: happy_ge_pattern_stress.json (400 trials/cell, seed=42).

git clone https://github.com/mrblakessinger-rgb/qec-evaluation-suite.git
cd qec-evaluation-suite
python -m qec_engine.layer_a_smoke
python -m qec_engine.repo_rollup

Prefer frozen JSON under public/data/ over re-running heavy sweeps when comparing claims. Full honesty language and non-claims: QEC_CLAIMS.md.

BP micro-lab (Phase E0/E1) — metric meanings

Classical Hamming [7,4] min-sum BP micro-lab (public/data/bp_micro_metrics.json). Not HaPPY.

Field	Meaning
`ge_unique_recoverability`	Unique erasure recovery: `rank(H_E) = \|E\|` over GF(2)
`bp_codeword_match`	BP hard decision equals the true all-zero codeword (and syndrome zero)
`bp_osd0_syndrome_valid`	OSD-0 found a syndrome-valid completion
`bp_osd1_syndrome_valid`	OSD-1 (order-1 free-bit search) found a syndrome-valid completion
`bp_osd1_codeword_match`	OSD-1 best candidate equals the all-zero true word

These metrics are not interchangeable. GE is uniqueness; BP is channel codeword match; OSD is list completion (not ML, not a threshold decoder).