Measured 2026-10-28T06:19Z

Twenty DOIs, three sources, one API call each

A benchmark of the free retraction-status lookup, not of a language model. Every figure below comes from sequential HTTP GETs against the live endpoint, recorded on 2026-10-28T06:19Z. The point is not a big number — it is that the answer is cheap, deterministic, and states where it has no evidence.

HTTP 200

20 / 20

no auth, no key, no retries

Median latency

1595 ms

mean 1965 ms, max 5318 ms

Retracted

three sources agreeing on each

Abstentions

stated reason, not a guess

Method

Corpus of questions. 16 retracted DOIs were drawn from the published retraction_status table (every DOI with a recorded retraction date), so the retracted set is not hand-picked from famous cases.
Controls. Three DOIs expected not to be retracted, plus one deliberately invalid DOI to exercise the abstention path.
One GET each, sequential, 350 ms apart, with a browser-style User-Agent, no retries counted as separate samples. Latency is wall-clock from request to parsed body, measured from a container reaching Cloudflare over the public internet.
Verdict derivation is mechanical: an abstention reason wins; otherwise any source asserting retraction gives “retracted”, all sources denying it give “not retracted”, and an empty status array gives “no rows”.

# reproduce: one GET per DOI, wall-clock ms, no retries
for doi in 10.1001/jama.293.14.1727 10.1038/nature14581 10.9999/not-in-snapshot; do
  curl -s -o /tmp/out.json -w "%{http_code} %{time_total}s $doi\n" \
    -A "Mozilla/5.0 (compatible; ashesborn-bench/1.0)" \
    "https://pay.ashesborn.cloud/api/v1/judge/paper?doi=$(python3 -c "import urllib.parse,sys;print(urllib.parse.quote(sys.argv[1],safe=''))" "$doi")"
  python3 -m json.tool < /tmp/out.json | head -20
  sleep 0.35
done

Every sample

DOI	Year	Verdict	Tiers	Retraction date	ms	Title
10.1001/jama.293.14.1727	2005	retracted	A+C	2018-09-19	2423	RETRACTED: Super Bowls: Serving Bowl Size and Food Consumption
10.1111/j.1365-2702.2011.03951.x	2012	retracted	A+C	2016-06-01	1447	Retracted: The impact of nurse short message services and telephone follow‐ups on diabetic
10.1002/14651858.cd003041.pub2	2003	retracted	A+C	2012-02-15	1526	Smoking cessation for the secondary prevention of coronary heart disease
10.4049/jimmunol.176.3.1534	2006	retracted	A+C	2015-06-15	1415	Pregnancy-Associated Exosomes and Their Modulation of T Cell Signaling
10.1111/j.1442-2042.2009.02429.x	2009	retracted	A+C	2010-10-01	3484	THIS ARTICLE HAS BEEN RETRACTED Periurethral injection of autologous adipose‐derived stem
10.1182/blood-2008-01-133710	2008	retracted	A+C	2012-02-16	3321	Osteolineage niche cells initiate hematopoietic stem cell mobilization
10.1182/blood.v98.9.2615	2001	retracted	A+C	2009-03-05	943	Purification and ex vivo expansion of postnatal human marrow mesodermal progenitor cells
10.1083/jcb.200501040	2005	retracted	A+C	2007-05-07	1509	RETRACTED: FOXO1a acts as a selective tumor suppressor in alveolar rhabdomyosarcoma
10.1056/nejmoa1101324	2011	retracted	A+C	2018-10-17	1468	RETRACTED: Evidence for Human Lung Stem Cells
10.1038/nature02889	2004	retracted	A+C	2006-06-29	1595	Induction of DNA methylation and gene silencing by short interfering RNAs in human cells
10.1111/j.1365-2044.1996.tb12557.x	1996	retracted	A+C	2021-01-14	1365	Retracted: The influence of volume therapy and pentoxifylline infusion on circulating adhe
10.1038/srep08209	2015	retracted	A+C	2026-02-23	1772	RETRACTED ARTICLE: Angiotensin-converting enzyme 2/angiotensin-(1–7)/Mas axis prevents lip
10.1016/j.molcel.2004.06.043	2004	retracted	A+C	2019-03-07	1702	RETRACTED: Methyl-CpG Binding Protein MBD1 Couples Histone H3 Methylation at Lysine 9 by S
10.3892/ijmm.2015.2172	2015	retracted	A+C	2023-02-20	1934	Hedgehog signaling is involved in the BMP9-induced osteogenic differentiation of mesenchym
10.1016/j.orgel.2005.03.010	2005	retracted	A+C	2007-02-22	1866	RETRACTED: Properties of organic light-emitting diodes by aluminum cathodes modification u
10.1007/s001340050511	1998	retracted	A+C	2013-11-23	2314	RETRACTED ARTICLE: Volume therapy in the critically ill: is there a difference?
10.1038/nature14581	2015	not retracted	A+C	—	5318	Glypican-1 identifies cancer exosomes and detects early pancreatic cancer
10.1126/science.aac4716	2015	no rows	—	—	1373	Estimating the reproducibility of psychological science
10.1056/NEJMoa2034577	2020	no rows	—	—	1570	Safety and Efficacy of the BNT162b2 mRNA Covid-19 Vaccine
10.9999/fake-not-a-real-doi-42	—	abstained	—	—	948

Limits, stated plainly

n = 20, one snapshot, one egress. This measures one endpoint from one machine at one hour. It does not estimate throughput under load and it is not a comparison with other services.
Two controls returned no status rows. The paper is in the graph, but no source asserts anything about its retraction status. Read that as “no evidence”, never as “clean”; the derivation returns a distinct verdict for exactly this reason.
“Retracted” here is a status, not a misconduct finding. Reason text is excluded from the dataset on purpose.
Latency is dominated by cold container wake-ups — the 5318 ms sample was the first request after idle. A warm median of ~1.4 s matches the server-timing header, so the engine itself answers in single-digit milliseconds.
The paid judgment is a different product. Propagation and evidence audit cost $0.02 and $0.10 per call and are not exercised by this benchmark.

Run it yourself

The interactive lookup on the datasets page hits the same endpoint with no key, so you can spot-check any DOI in this table — including the controls that return no rows.

Try a DOI How the graph is built