BENCHMARK · 快照 ri-20261025-1526
这份判断到底准不准
把「带证据的判断」当成产品，第一件事是把它的错误率公开。下面是 28 个案例的真实运行结果，不是挑选过的演示；错的、弃权的、够不着的都留在表里。
01 方法
标签不是我们打的，是从库里按构造规则算出来的：
- A — 论文本身有 A 级（登记在 DOI 系统的）撤稿通知，判定应为 CONTESTED。
- B — 论文的参考文献里有撤稿论文，判定应为 CONTESTED 或弃权；判成 SUPPORTED 就是假阴性。
- C — 快照内两跳都碰不到撤稿论文的对照。判成 SUPPORTED 才算正确。
每个案例都是对生产端点 retraction_propagation 的真实调用（2026-10-28，快照 ri-20261025-1526），延迟是客户端墙钟时间，包含公网往返。
02 案例（点开看图谱）
| 组 | DOI | 判定 | Conf. | 覆盖 | 撤稿根 | ms |
|---|---|---|---|---|---|---|
| DEPTH0 | 10.1016/j.saa.2015.02.011 | CONTESTED | 0.50 | 0.00 | 1 | 1817 |
RETRACTED: Green synthesis of zinc oxide nanoparticles using Moringa oleifera leaf extract and evaluation of its antimicrobial activity (2015) · DOI 10.1016/j.saa.2015.02.011 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1817 ms | ||||||
| DEPTH0 | 10.1083/jcb.200501040 | CONTESTED | 0.50 | 0.00 | 1 | 2050 |
RETRACTED: FOXO1a acts as a selective tumor suppressor in alveolar rhabdomyosarcoma (2005) · DOI 10.1083/jcb.200501040 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2050 ms | ||||||
| DEPTH0 | 10.1182/blood-2008-01-133710 | CONTESTED | 0.50 | 0.00 | 1 | 2019 |
Osteolineage niche cells initiate hematopoietic stem cell mobilization (2008) · DOI 10.1182/blood-2008-01-133710 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2019 ms | ||||||
| DEPTH0 | 10.1016/s0092-8674(03)00436-7 | CONTESTED | 0.50 | 0.00 | 1 | 2024 |
RETRACTED: The Chromatin-Remodeling Complex WINAC Targets a Nuclear Receptor to Promoters and Is Impaired in Williams Syndrome (2003) · DOI 10.1016/s0092-8674(03)00436-7 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2024 ms | ||||||
| DEPTH0 | 10.1016/j.jacc.2011.12.043 | CONTESTED | 0.50 | 0.00 | 1 | 1921 |
RETRACTED: Plasma PCSK9 Levels and Clinical Outcomes in the TNT (Treating to New Targets) Trial (2012) · DOI 10.1016/j.jacc.2011.12.043 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1921 ms | ||||||
| DEPTH0 | 10.2337/db13-1526 | CONTESTED | 0.50 | 0.00 | 1 | 2289 |
Replication of Obesity and Associated Signaling Pathways Through Transfer of Microbiota From Obese-Prone Rats (2014) · DOI 10.2337/db13-1526 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2289 ms | ||||||
| DEPTH0 | 10.1074/jbc.m611728200 | CONTESTED | 0.50 | 0.00 | 1 | 2125 |
Arabidopsis Phosphatidylinositol Phosphate Kinase 1 Binds F-actin and Recruits Phosphatidylinositol 4-Kinase β1 to the Actin Cytoskeleton (2007) · DOI 10.1074/jbc.m611728200 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2125 ms | ||||||
| DEPTH0 | 10.1073/pnas.0610590104 | CONTESTED | 0.50 | 0.00 | 1 | 1781 |
RETRACTED: Metabolic regulation of SIRT1 transcription via a HIC1:CtBP corepressor complex (2007) · DOI 10.1073/pnas.0610590104 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1781 ms | ||||||
| DEPTH0 | 10.1038/nature26142 | CONTESTED | 0.50 | 0.00 | 1 | 2026 |
RETRACTED ARTICLE: Quantized Majorana conductance (2018) · DOI 10.1038/nature26142 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2026 ms | ||||||
| DEPTH0 | 10.1038/s41417-022-00501-1 | CONTESTED | 0.50 | 0.00 | 1 | 1736 |
RETRACTED ARTICLE: Nrf2 overexpression increases the resistance of acute myeloid leukemia to cytarabine by inhibiting replication factor C4 (2022) · DOI 10.1038/s41417-022-00501-1 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1736 ms | ||||||
| DEPTH0 | 10.1016/j.jcrysgro.2005.05.065 | CONTESTED | 0.50 | 0.00 | 1 | 2060 |
RETRACTED: Nucleation and growth behavior of well-aligned ZnO nanorods on organic substrates in aqueous solutions (2005) · DOI 10.1016/j.jcrysgro.2005.05.065 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2060 ms | ||||||
| DEPTH0 | 10.1001/jama.293.9.1082 | CONTESTED | 0.50 | 0.00 | 1 | 24704 |
Effect of folate and mecobalamin on hip fractures in patients with stroke: a randomized controlled trial. (2005) · DOI 10.1001/jama.293.9.1082 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 24704 ms | ||||||
| DEPTH0 | 10.1128/jvi.76.13.6815-6824.2002 | CONTESTED | 0.50 | 0.00 | 1 | 1657 |
P0 of Beet Western Yellows Virus Is a Suppressor of Posttranscriptional Gene Silencing (2002) · DOI 10.1128/jvi.76.13.6815-6824.2002 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1657 ms | ||||||
| DEPTH0 | 10.1038/ncb942 | CONTESTED | 0.50 | 0.00 | 1 | 2197 |
Cytokines suppress adipogenesis and PPAR-γ function through the TAK1/TAB1/NIK cascade (2003) · DOI 10.1038/ncb942 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 2197 ms | ||||||
| DEPTH0 | 10.1038/35099527 | CONTESTED | 0.50 | 0.00 | 1 | 1922 |
Magnetic carbon (2001) · DOI 10.1038/35099527 verdict CONTESTED · confidence 0.50 · coverage 0.00 (0/0 references resolved) · 1922 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000001052 | ABSTAINED | 0.50 | 1.00 | 10 | 2542 |
Heart Disease and Stroke Statistics—2022 Update: A Report From the American Heart Association (2022) · DOI 10.1161/cir.0000000000001052 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (3191/3195 references resolved) · abstained: single_aggregator_source · 2542 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000000558 | ABSTAINED | 0.50 | 1.00 | 2 | 2327 |
Heart Disease and Stroke Statistics—2018 Update: A Report From the American Heart Association (2018) · DOI 10.1161/cir.0000000000000558 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (2260/2269 references resolved) · abstained: single_aggregator_source · 2327 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000000485 | ABSTAINED | 0.50 | 1.00 | 3 | 2430 |
Heart Disease and Stroke Statistics—2017 Update: A Report From the American Heart Association (2017) · DOI 10.1161/cir.0000000000000485 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (2094/2101 references resolved) · abstained: single_aggregator_source · 2430 ms | ||||||
| DEPTH1 | 10.4158/ep161365.gl | CONTESTED | 0.99 | 0.99 | 3 | 2368 |
American Association of Clinical Endocrinologists and American College of Endocrinology Comprehensive Clinical Practice Guidelines For Medical Care of Patients with Obesity (2016) · DOI 10.4158/ep161365.gl verdict CONTESTED · confidence 0.99 · coverage 0.99 (1801/1818 references resolved) · 2368 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000000350 | ABSTAINED | 0.50 | 1.00 | 3 | 2244 |
Heart Disease and Stroke Statistics—2016 Update (2015) · DOI 10.1161/cir.0000000000000350 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (1743/1748 references resolved) · abstained: single_aggregator_source · 2244 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000000152 | CONTESTED | 1.00 | 1.00 | 2 | 24640 |
Heart Disease and Stroke Statistics—2015 Update (2014) · DOI 10.1161/cir.0000000000000152 verdict CONTESTED · confidence 1.00 · coverage 1.00 (1521/1525 references resolved) · 24640 ms | ||||||
| DEPTH1 | 10.1161/cir.0000000000000152 | CONTESTED | 1.00 | 1.00 | 2 | 2987 |
Heart disease and stroke statistics--2015 update: a report from the American Heart Association. (2015) · DOI 10.1161/cir.0000000000000152 verdict CONTESTED · confidence 1.00 · coverage 1.00 (1521/1525 references resolved) · 2987 ms | ||||||
| DEPTH1 | 10.1161/01.cir.0000441139.02102.80 | ABSTAINED | 0.50 | 0.99 | 2 | 1979 |
Heart Disease and Stroke Statistics—2014 Update (2013) · DOI 10.1161/01.cir.0000441139.02102.80 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 0.99 (1452/1460 references resolved) · abstained: single_aggregator_source · 1979 ms | ||||||
| DEPTH1 | 10.1097/01.hjh.0000431740.32696.cc | CONTESTED | 1.00 | 0.99 | 4 | 2226 |
2013 ESH/ESC Guidelines for the management of arterial hypertension (2013) · DOI 10.1097/01.hjh.0000431740.32696.cc verdict CONTESTED · confidence 1.00 · coverage 0.99 (1292/1299 references resolved) · 2226 ms | ||||||
| DEPTH1 | 10.1161/str.0000000000000046 | ABSTAINED | 0.50 | 1.00 | 4 | 2530 |
Guidelines for the Primary Prevention of Stroke (2014) · DOI 10.1161/str.0000000000000046 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (1221/1227 references resolved) · abstained: single_aggregator_source · 2530 ms | ||||||
| DEPTH1 | 10.36660/abc.20201238 | ABSTAINED | 0.48 | 0.97 | 8 | 2370 |
Diretrizes Brasileiras de Hipertensão Arterial – 2020 (2021) · DOI 10.36660/abc.20201238 verdict INSUFFICIENT_EVIDENCE · confidence 0.48 · coverage 0.97 (1110/1144 references resolved) · abstained: single_aggregator_source · 2370 ms | ||||||
| DEPTH1 | 10.1161/str.0000000000000024 | ABSTAINED | 0.50 | 1.00 | 4 | 2192 |
Guidelines for the Prevention of Stroke in Patients With Stroke and Transient Ischemic Attack (2014) · DOI 10.1161/str.0000000000000024 verdict INSUFFICIENT_EVIDENCE · confidence 0.50 · coverage 1.00 (1091/1095 references resolved) · abstained: single_aggregator_source · 2192 ms | ||||||
| CLEAN | 10.3748/wjg.11.7461 | SUPPORTED | 1.00 | 1.00 | 0 | 1595 |
Fibroblast growth factor-4 and hepatocyte growth factor induce differentiation of human umbilical cord blood-derived mesenchymal stem cells into hepatocytes (2005) · DOI 10.3748/wjg.11.7461 verdict SUPPORTED · confidence 1.00 · coverage 1.00 (13/13 references resolved) · 1595 ms | ||||||
03 它做不到什么
没有干净的对照组。这个快照是「撤稿邻域」抽样——从撤稿论文出发抓它们的引用文献，所以几乎每一篇都在污染范围内。28 个案例里只有 1 个真正的对照。因此这份基准能证明「抓得到」和「不乱判」，不能证明在干净文献上的精确率。补对照组需要重新采集（约 1 天的抓取额度），这是下一版的事。
弃权不是失败。12 个引用撤稿论文的案例里有 8 个弃权，原因全是 single_aggregator_source：只有 C 级（聚合器）声称撤稿，没有 A 级通知。我们选择不下结论，而不是拿聚合器当事实。
覆盖度是硬上限。平均引用覆盖率 0.46，低于 0.5 一律弃权。想让结论更硬，先补数据，不是调模型。
04 自己复现
判断层是确定性的图查询，同一个 DOI 在同一快照上永远返回同一个结果。
匿名 MCP（无需密钥，20 次/分钟/IP）：
curl -s -X POST https://mcp.ashesborn.cloud/mcp \
-H 'content-type: application/json' \
-H 'accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"bench","version":"1"}}}'
免费状态查询（无需密钥）： ashesborn.cloud/datasets ·
GET https://pay.ashesborn.cloud/api/v1/judge/paper?doi=<doi>
Raw run output — 28 records, verdicts, roots, coverage and latency as measured.
完整判断（含逐跳证据路径）按次计费 $0.02 / $0.10，无需注册： POST https://pay.ashesborn.cloud/api/v1/judge