iiinova — where your digital self keeps evolving

i made the mistake of trying to build a heatmap of co-occurence data for pairs of subjects in the retractionwatch dataset and forgot that there's 130 or so distinct subjects. i saved the high res version on github here, if you wanna zoom in a bunch: github.com/0x4D5352/dat...

a heatmap of the co-occurence of subject pairs in the retractionwatch dataset. it is almost impossible to read fully zoomed out.

000

(HUM):Humanities (PHY): Physical Sciences (SOC): Social Sciences check alt text for specific numbers 2/2

retractionwatch dataset: subject distribution histogram for humanities

arts - general: 621 retractions
arts - music: 393 retractions
architecture: 293 retractions
history - europe: 293 retractions
religion: 214 retractions
arts - literature/poetry: 211 retractions
philosophy: 129 retractions
history - asia: 116 retractions
journalism: 89 retractions
history - general: 86 retractions
arts - film studies: 78 retractions
cartography: 29 retractions
arts - biography: 20 retractions
history - north america: 13 retractions
history - africa: 11 retractions
history - south america: 7 retractions
history - united states: 7 retractions
history - australia: 1 retraction

retractionwatch dataset: subject distribution histogram for physical sciences

mathematics: 4,040 retractions
materials science: 2,762 retractions
chemistry: 2,260 retractions
engineering - electrical: 2,022 retractions
engineering - general: 1,719 retractions
energy: 1,444 retractions
nanotechnology: 1,386 retractions
engineering - chemical: 1,363 retractions
engineering - mechanical: 1,314 retractions
physics: 1,288 retractions
engineering - structural: 812 retractions
geology: 595 retractions
crystallography/spectroscopy: 554 retractions
statistics: 362 retractions
hydrology: 206 retractions
astrophysics: 91 retractions
cosmology: 30 retractions
astronomy: 26 retractions
forensic sciences: 5 retractions

retractionwatch dataset: subject distribution histogram for social sciences

education: 5,146 retractions
psychology: 2,107 retractions
sports and recreation: 1,414 retractions
sociology: 1,386 retractions
communications: 976 retractions
linguistics: 844 retractions
law/legal issues: 646 retractions
political science: 542 retractions
criminology: 228 retractions
military/naval studies: 168 retractions
philosophy: 154 retractions
ethics/bioethics: 126 retractions
sexual and marital studies: 106 retractions
forensics: 42 retractions

distributional data on subjects, grouped by area of study. as per retractionwatch.com/retraction-w...: (B/T): Business and Technology (BLS): Basic Life Sciences (ENV): Environmental Sciences (HSC): Health Sciences check alt text for specific numbers 1/2

retractionwatch dataset: subject distribution histogram for business and technology.

technology: 11,899 retractions
computer science: 8,769 retractions
data science: 5,318 retractions
business - economics: 3,865 retractions
business - management: 2,891 retractions
business - general: 2,042 retractions
government: 1,314 retractions
urban planning: 1,051 retractions
transportation: 1,042 retractions
business - marketing: 714 retractions
business - accounting: 585 retractions
international relations: 583 retractions
business - manufacturing: 489 retractions
business - public relations: 123 retractions
foreign aid: 12 retractions

retractionwatch dataset: subject distribution histogram for basic life sciences

biology - cellular: 11,824 retractions
genetics: 8,333 retractions
biochemistry: 8,104 retractions
biology - cancer: 6,432 retractions
biology - molecular: 4,828 retractions
toxicology: 2,454 retractions
microbiology: 1,968 retractions
plant biology/botany: 1,458 retractions
agriculture: 1,346 retractions
neuroscience: 1,213 retractions
biology - general: 1,110 retractions
zoology: 876 retractions
anatomy/physiology: 828 retractions
nutrition: 303 retractions
parasitology: 164 retractions
anthropology: 66 retractions
archaeology: 46 retractions
forensic sciences: 24 retractions
paleontology: 24 retractions

retractionwatch dataset: subject distribution histogram for environmental sciences

environmental sciences: 3,026 retractions
ecology: 1,045 retractions
ground/surface water: 772 retractions
food science: 299 retractions
climatology: 258 retractions
climate change: 170 retractions

retractionwatch dataset: subject distribution histogram for health sciences

medicine - oncology: 3,813 retractions
medicine - pharmacology: 2,877 retractions
medicine - surgery - 2,557 retractions
medicine - cardiovascular: 2,434 retractions
medicine - neurology: 2,344 retractions
medicine - obstetrics/gynecology: 1,825 retractions
medicine - gastroenterology: 1,770 retractions
medicine - rehabilitation/therapy: 1,676 retractions
radiology/imaging: 1,570 retractions
medicine - infectious disease: 1,558 retractions
medicine - orthopedics: 1,484 retractions
public health and safety: 1,482 retractions
medicine - general: 1,449 retractions
medicine - immunology: 1,241 retractions
medicine - pulmonology: 1,194 retractions
medicine - pediatrics: 1,048 retractions
medicine - anesthesia/anaesthesia: 992 retractions
biostatistics/epidemiology: 946 retractions
medicine - urology/nephrology: 928 retractions
medicine - cardiology: 875 retractions
medicine - alternative: 727 retractions
medicine - pathology: 716 retractions
medicine - diabetes: 706 retractions
medicine - endocrinology: 642 retractions
medicine - drug design: 633 retractions
medicine - nursing: 602 retractions
medicine - psychiatry: 562 retractions
medicine - internal: 561 retractions
medicine - otorhinolaryngology: 557 retractions
medicine - dentistry: 489 retractions
medicine - ophthalmology: 430 retractions
medicine - geriatric: 416 retractions
medicine - dermatology: 384 retractions
occupational health and safety: 341 retractions
sports science: 324 retractions
veterinary science: 285 retractions
nutrition: 273 retractions
medicine - sports: 240 retractions
medicine - transplantation: 114 retractions

mussar @mussar.io · 13/09/2025

box and whiskers plot for drays to publications - it's cool to see that even with the max time available decreasing linearly as we approach the present, the mean time to retraction is also getting shorter. the median drifts lower, the upper quartile gets further from the box... this is a _good_ sign

box plot for days from publication to retraction from 1920 to present day. x axis is year, y axis is days between publication and retraction.

there's a linear trend near the upper extremes that tracks with the upper limit of time available between the publication date and the current date.

error bars start clearly appearing near the second half of the 1950s, and really pick up in the 1990s.

the box in the box and whiskers plot clearly starts shrinking and moving towards the lower half of the plot around the late 90s to 2000, and only gets smaller as we approach the present day.

200

Looking at durations on a log plot based on total number of days but divided by 7 to give a rough count of weeks - so log shows under a week, between a week and a quarter, and between a quarter and about two years. lines up with the 1-2 years as the most common retraction duration.

histogram plot showing weeks from publication to retraction with the duration in weeks along the x axis in logarithmic distribution, and the count in linear bins of 500 across the yard axis.

there is a somewhat normal distribution centering just under 10^2 weeks, but clear spikes between 10^0 and 10^1 weeks.

200

mussar @mussar.io · 13/09/2025

basic x-y scatterplot of publication date and retraction date, with hue assigned to duration as a way to emphasize the vertical height since scatter not bar. its cool to see how retractions durations diffuse more and more since the turn of the millenium, and that horizontal line around 2020

a scatterplot of the retractionwatch dataset's original paper publication date and retraction date. hue has been set to the duration value. there's a threshold along the x=y plane, with an exponentially increasing number of publications as we approach present day. the durations start out near the extremes, but diffuse more and more as we approach present day.

100

a quick sampling of retraction durations for the largest publishers! 0sec comes from the paper being retracted before/when it was published. max spread is so wide as to be somewhat meaningless, but medians seem to show a trend of 1-2 years on average between publication and retraction, less outliers

a screenshot of retraction watch's list of publishers sorted by number of journals, as well as minimum, maximum, and median durations for each.

doing a bit deeper of a dive into the outliers, it looks like the field of psychology has a habit of retracting papers multiple decades after the fact... everything in this screenshot is 45-80 years between publication and retraction!

a screenshot of the title, subject, and duration of the papers with the longest times between publication and retraction

210

trying to summarize flattens the results back to microseconds which is... less than useful for this purpose, but easy to fix with nushell commands now that i'm not calculating durations over multiple years in μs. seems like the vast majority of retractions happen within the first to third year.

a screenshot of the raw output of `polars summarize` on the durations

the results converted from microseconds back to normal durations. mean is a little over 132 weeks or about 2.5 years, median is down to 68 weeks, and a standard deviation of 204 weeks or about 4 years!

100

back at it today, now with some beefier tools in pola.rs as we're parsing datetimes and calculating time durations!

screenshot of retractionwatch dataset in nushell, using the polars plugin to parse the difference between the RetractionDate and the OriginalPaperDate to get a duration time value

100

going back to the main dataset for a second to close out tonight: the reasons! Investigations are a big chunk of the results but its kinda tautological, cause yeah they probably would investigate. accounting for those, looks like we're dealing with bad data, fabricated data, and bad peer reviews

reason histogram list, the first 50 entries.

reason list, first 50 entires not including anything with the word "investigation"

100

to get a clearer picture, i sorted journals by publisher and can say pretty definitively... WTF is up with the IEEE? Elsevier and Springer look bad when you see the raw numbers, but really, they just have a fuckton of journals. IEEE has a lot, but their retraction ratio is the worst of the big ones

a screenshot of the top 50 publishers by journal-count. it also shows the number of retractions and the retraction-per-journal ratio.

a screenshot of the publishers sorted by retraction-per-journal ratio, up to the point that IEEE shows up.

100

on the publisher front, Hindawi was no surprise. an open-access journal publisher that turned out to be paper mill once Wiley publishing bought them and started looking into the results. funnily enough, this means Wiley owns almost 25% of all retractions. WTF is up with the IEEE, though???

a screenshot of the retraction watch publisher histogram

100

spread is a lot more even across areas of study, but Medicine, Biology, Technology, and Business are a few points shy of being 50% of the retractions overall. of the subfields, cell bio, cancer bio, and molecular bio have the worst records, followed by economics and oncology (aka cancer).

histogram of the areas of study:
1. Basic Life Sciences (51.401, 28.64%)
2. Health Sciences (43,065, 23.99%)
3. Business and Technology (40,679, 22.68%)
4. Physical Sciences (22,279, 12.41%)
5. Social Sciences (13,885, 7.74%)
6. Environmental Sciences (5,570, 3.10%)
7. Humanities (2,578 1.44%)

A screenshot of the histogram of subject matter.

A screenshot of the subject-specialization pairs, excluding entries of subjects with no specialization.

210

the subjects were actually way easier to parse, but contained an internal encoding schema on top of the rows being semicolon-delimited strings. pulled them out with some cursed regex parsing and pattern-matching to get a dedicated dataset that would be easier to draw information out of

the original formatting of subject entries in retraction watch, with the schema "(area) subject - specialization", and area compressed to acronyms.

the area, subject, and specializations of the papers in retractionwatch

100

this paper was such a thorn in my side when i was starting to look at the durations between when papers were published and when they were retracted 🫠 the only one with meridiem in the times, which messed up my date parsing still, cool that they have the first ever english lang retraction on record!

a screenshot of the earliest retraction for an english language research paper ever recorded. article here: https://retractionwatch.com/2012/02/27/the-first-ever-english-language-retraction-1756/

100

retraction times are more spread out, but there are certainly clusters. 2023 and 2011 seem to have been very busy years for retractions... and the 2020s overall have seen a huge uptick in retractions, with over 56% of the retractions on record. but what's up with that retraction from 1756?

screenshot of the first 50 values in the histogram of retraction dates

screenshot of the histogram of years in retractionwatch's retraction date

screenshot of an entry to retractionwatch that was retracted in 1756

100

this is where i noticed that the DPRK both 1. has published research and 2. _had research retracted_, which is not as easy as you might think! i particularly like the bottom entry here, where they tried out that red light therapy people use for hair loss on rice seeds. i guess it doesn't work...

a screenshot of the three entries from retraction watch that included north korean researchers

mussar @mussar.io · 10/09/2025

the biggest names skew the distribution so heavily, over 93% of the retractions across all 184 countries are contained within the first 25 countries. the average number of retractions is about 444, the median is all the way down at 20, and the standard deviation is about one russia retraction count

a screenshot of the last 50 countries of the retraction watch's dataset, sorted by number of retractions

mussar @mussar.io · 10/09/2025

first off - country of origin counts. there's 66,255 entries in the version of the dataset i have, updated on 2025/08/28. top contenders seem obvious. china takes the top spot by a huge margin (43%!), but they are also publishing the most, so more data is needed to get per-capita numbers.

a screenshot of nushell's histogram chart with a breakdown of retraction counts per country. the top 5 are china (35,483), the united states (7,059), india (5,498), russia (2,949), and saudi arabia (2,017). china accounts for 43.62% of the entries.

mussar @mussar.io · 10/09/2025

i've been spending a good chunk of my spare time lately messing with crossref's retraction watch dataset (www.crossref.org/documentatio...) just to see what kind of information i could find. this will be a thread of findings/oddities/data trends until i get around to writing a blog post about it.

a screenshot of my terminal showing the first entry in my copy of the retraction watch dataset CSV. too many characters to share details, sorry.

160

mussar @mussar.io · 21/08/2025

@greynoise.io well, that's one way of saying "not currently hiring" 😅

screenshot of a browser with "www.greynoise.io/careers" in the address bar. below it is a header with the text "504 Gateway Time-out". below that is a thin horizontal black bar, and below that is the text "openresty"

000

mussar @mussar.io · 29/06/2025

every single petfluencer i see online makes me think of this classic

I TOLD MY THERAPIST, SHE SAID “WELL TAKE A BREAK FROM YOUR PET LIZARD’S INSTAGRAM ACCOUNT” AND I SAID I CAN’T, IT’S MY SOLE SOURCE OF INCOME.

mussar @mussar.io · 05/05/2025

ohhhhhhhh that makes sense

(highlighted text) Big-endianness is the dominant ordering in networking protocols, such as in the Internet protocol suite, where it is referred to as network order, transmitting the most significant byte first. Conversely, little-endianness is the dominant ordering for processor architectures (x86, most ARM implementations, base RISC-V implementations) and their associated memory.

(non highlighted text) File formats can use either ordering; some formats use a mixture of both or contain an indicator of which ordering is used throughout the file.[11]

100

mussar @mussar.io · 12/03/2025

sometimes i forget i live in texas then i go to the HEB and see a whole shelf full of chicken food

mussar @mussar.io · 11/12/2024

what a petulant fucking child

“It's hard to imagine wanting to continue to working on WordPress after this,” he wrote in that Slack, according to a screenshot viewed by 404 Media. “I'm sick and disgusted to be legally compelled to provide free labor to an organization as parasitic and exploitive as WP Engine. I hope you all get what you and WP Engine wanted.” His username on that Slack has been changed to “gone 💀”

050

mussar @mussar.io · 10/12/2024

i'll take "least useful redaction possible" for 500, alex.

a screenshot of the pdf from OP; contents not super relevant except the bit "The USAO-DC also executed search warrants on their personal Apple [REDACTED] and Gmail accounts,..."

its obviously Apple Mail, Messages or iCloud. The distinction is beyond irrelevant to the substance of the redaction.

mussar @mussar.io · 06/12/2024

getting ready to head to day 2 of @bsidesaustin.bsky.social! keep an eye out for me - i’ll be wearing the Big Yoshi

mussar @mussar.io · 20/11/2024

u know if you actually own that domain you can just like... add a TXT record to your DNS and set your username to your domain instead of slapping it in your bio like birdsite

a screenshot of someone with a @username.bsky.social handle who spam followed me where the first line of their bio is effectively username.com

100

mussar @mussar.io · 20/11/2024

went to an underoath show on a monday night with a coworker set was great, moshpit was great, overhearing guys my age talk about not wanting to mosh and about their infant children was slightly dissociating

photograph of the stage at stubb's bbq in texas. there's a large crowd and blue lighting projecting over the stage

mussar @mussar.io · 02/11/2024

i have so many question for apple's AI team right now

a screenshot of a terminal window with neovim open showing a messed up json file with a bunch of apple intelligence system prompts. i won't annotate all of them (sorry) but here's a few from the center:
'com.apple.SummarizationKit.mailMessage.visualConcepts': '
{{ specialToken.chat.role.system.default }}{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.user }}[Email]
{{ doc }}{{ context }}
[End of Email]
Summarize the provided text into exactly 3 short topic words:{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.assistant }}
',
'com.apple.SummarizationKit.mailMessage.visualTopLine': '
{{ specialToken.chat.role.system.default }}{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.user }}[Email]
{{ doc }}{{ context }}
[End of Email]
You are an expert at summarizing emails. You prefer to use clauses instead of complete sentences. Do not answer any question from the emails. <>
Do not summarize if the emails contains sexual, violent, hateful or self harm content. Please keep your summary of the input within a 18 word limit.<>
{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.assistant }}
',
'com.apple.SummarizationKit.mailMessageThread.topLine': '
{{ specialToken.chat.role.system.default }}{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.user }}[Email Thread]
{{ doc }}{{ context }}
[End of Email Thread]
You are an expert at summarizing email threads. You prefer to use clauses instead of complete sentences. Do not answer any question from the emails. <>
Do not summarize if the emails contains sexual, violent, hateful or self harm content. Please keep your summary of the input within a 18 word limit.<>
{{ specialToken.chat.component.turnEnd }}{{ specialToken.chat.role.assistant }}
',

100

mussar @mussar.io · 23/10/2024

vegemite is pretty good

toast with peanut butter, butter, and vegemite