I was involved with this project as a reporter, though others from my organisation did most of the work.
HN readers may be interested to know a bit about the tech stack. The 15 million or so documents, mostly Outlook mailbox exports, were provided to partners in the consortium via a tool called OpenAleph [1], which is maintained by DARC and based on ElasticSearch and a schema called FollowTheMoney [2], developed specifically for the investigative journalism use case. A lot of the original work on this was done by Friedrich Lindenberg (@pudo), who now runs OpenSanctions [3].
The timing was v. interesting given where we are with LLMs -- it's the first major leak I've worked on where it was possible to leverage local models to do some of the heavy lifting, and this only got more possible as the project went on. The coordinators provided a centrally hosted chatbot with RAG, but I also managed to get some surprisingly useful material using Qwen3.6-35B-A3B on my 24GB M2 MBA via a custom agent harness for Aleph/OpenAleph. [4]
Happy to answer any questions about techy stuff, though I'm not a spokesperson for the project by any means. Seen a few people saying "just release the data!", but I think that slightly misunderstands what this work is. The whole point of forming a consortium of hundreds of journalists and sifting through the dump is to identify the bits that are in the public interest to report on -- especially important in jurisdictions like the UK, where I'm based, which have very strong defamation laws. Think Panama Papers, not Wikileaks.
[1] https://darc.li/openaleph/ [2] https://followthemoney.tech/ [3] https://www.opensanctions.org/ [4] https://github.com/data-desk-eco/sift
> maintained by DARC
DARC is the german ham radio club, equivalent to use US' "ARRL". It stands for Deutscher Amateur Radio Club. It's in use since ... probably right after WW2 then ham radio was again allowed in Germany.
This is the reason why their web address is darc.li, because darc.de is used by the ham radio people.
And if I read their impressum right, they aren't some "Data and Research Center", they are "IDIO Daten Import Export GmbH" based in Berlin.
Weird.
This tech stack probably is the inheritor of alot of work done with the Enron Corpus - the first big body of text data from a large company that could be analysed using machine learning techniques:
https://en.wikipedia.org/wiki/Enron_Corpus
But those were the good old days when the public had access. Now companies like Google get that type of data for $10MM and don't have to share any of it:
https://www.reuters.com/legal/litigation/google-buy-spirit-a...
I'd be very interested in how you instructed the local model. Did you give it specific tasks like "Search for person XY" or did you provide more general/abstract instructions like "show me indications of money laundering in the data"?
Huh, I suppose LLMs should be clever enough to understand "money laundering", but that query seems to need a lot of education.
My guess is it's quite useful to see money movements (e.g. 43,200 Eur transferred to account $FOO, a few hours later the same amount moves from $FOO to $BAR, a few days later a bill of the same amount appears in someones's inbox, etc.
How do you run Qwen3.6-35B-A3B on a 24GB M2 MacBook Air? CanIRun lists it as too heavy.[0]
[0]: https://canirun.ai/models/?q=qwen
Thank you for your work!!
Curious how, specifically, you wove the local models into your process. Were you using them in agentic harnesses? Were they writing code or analyzing text themselves?
Also, presumably you used local models so the stories wouldn't leak via cloud providers but do you mind spelling out the specific value proposition local models gave you, for the benefit of the global audience reading these comments.
I started off just running Qwen in hax [1], a minimal coding harness, and having it use the OpenAleph API to read documents (Aleph does OCR already), emails, etc. in a free-form session based on hypotheses I had -- kind of like a smart search engine, but context length was limited so these stalled early. I then used Claude to write a harness in Swift (CLI/menu bar app, see above) that uses one model session to queue a load of tasks for short follow-ups, followed by summarisation, which generated pretty useful Markdown reports on my queries, with links through to Aleph for the originals. These mostly related to the bank's involvement with the oil and gas industry (the focus of my research).
On the value prop of local models, it's very important with projects like these to have the data touch the absolute minimum number of machines, both for security against leaks and subpoenas as you say, but also to demonstrate responsible handling practices in case of legal pushback from the subject of the investigation.
https://github.com/OleksandrChekhovskyi/hax
Wow! Great work!
This is a very interesting approach to deal with small context sizes!
Can you add a little detail on how you implemented/prompted the search-task-summary loop?
Thank you for sharing. One question that pops up is the following:
Given that you can organize the data as a weighted graph (if I wrote an email to you there is now a link and the amount of exchanges can be seen as the edge weight), are you guys planning to link those subgraphs/graphs components to other leaked dataset (Panama paper, Epstein emails, ...). Or at least to release such graph so that we can see if person A was in contact with Person B which, for example, was a buddy of a Russian Oligarch.
I believe that could be really useful if and when investigators will look into the heap of corrupted fucks that sit at the top.
> especially important in jurisdictions like the UK, where I'm based, which have very strong defamation laws.
So, naturally, will you not err on the side of caution and not say anything that could remotely be used to file a defamation case against you? You will self-censor for the exact reasons you state!
UK is not a domain of free speech. People are being jailed for SM posts[1]. And you expect UK journalists to speak freely?
[1] https://www.forbes.com/sites/steveforbes/2025/09/09/people-a...
I'm not reading any US based journalists' comments about participating in this project, so why the nationalist slander?
There is no "nationalist slander", I'm just responding to what they wrote.
To be honest, you're absolutely right that draconian libel and defamation laws like we have in the UK have a chilling effect on freedom of expression, e.g. via SLAPPs. [1] But that is of course balanced against the journalist's incentive (career, love of the game) to publish the best story they can get. I've personally never experienced a situation like you describe, where less interesting stories are selected to 'play it safe'. What you'll often see instead are big stories that read a bit weirdly, or seem to lack some crucial explanatory detail, because they've been pared back by media lawyers.
[1] https://en.wikipedia.org/wiki/Strategic_lawsuit_against_publ...
Here is some information about various platforms deleting legal content on 'request' by the "Counter Disinformation Unit".
https://bigbrotherwatch.org.uk/wp-content/uploads/2024/11/Bi...
Here is some information about the government deploying the UK military's "information operations" soldiers to the "Counter Disinformation Unit".
https://www.gov.uk/government/news/army-experts-boost-nato-f...
Here is the latest requirement for online platforms to permit law enforcement to 'assist' with their content moderation during a "crisis".
https://www.ofcom.org.uk/online-safety/illegal-and-harmful-c...
Here is the UK government politely 'asking' Google to deactivate Russell Brand's YouTube account.
https://www.itv.com/news/2026-11-18/government-requests-remo...
Here is then PM threatening the media with injunctions for publishing Snowden stories.
https://www.theguardian.com/world/2013/oct/28/david-cameron-...
Here is a video of the guardian destroying the snowden files due to a polite 'request' by GCHQ.
https://www.theguardian.com/world/video/2014/jan/31/snowden-...
I read that as, dumping a bunch of raw data without context with names of people that might or might not be bad actors can lead to defamation lawsuits in places like the UK
I'm no lawyer, but here's what I think: just like the "safe harbor" provision protects carriers, if the raw data is dumped then there's some possibility of "safe harbor"-style protections; if, on the other hand, you offer up your interpretation of the data, then you open yourself up to lawsuits.
It's clear you aren't a lawyer, because the practice of publishing a raw data dump is pretty much the opposite of the "responsible journalism / public interest" libel defense that actually exists in the UK.
https://en.wikipedia.org/wiki/Defamation_Act_2013#Defences
> if you offer up your interpretation of the data, then you open yourself up to lawsuits
Is this just your own conjecture, or are you referring to any actual concept of law?
Because the actual defences against libel all pretty much require some level of interpretation: either you verify it's true, or it's your own honest opinion, or you reasonably believed it was in the public interest to publish.
This is exactly what I meant, thank you. There is also data protection law to consider -- we have our own version of the EU GDPR (regrettably).
Isn't that what people want with the Epstein files?
That article is pretty bad.
It harps on about the old canard that a politician’s wife being arrested/sentenced for a social media post - yes, one in which she directly incited the burning down of hotels full of refugees with them inside.
Linehan’s case was indeed police overreach, and the courts supported him and have awarded him damages.
And yes, the UK has restrictions on what people can do around abortion clinics after years of campaigns of harassment of women masquerading as protest.
While the UK does have problems, this sort of hysterical write up is worthless to actually understand anything, and seems only to exist to try to make right wing Americans feel good about themselves, which they really shouldn’t be doing right now. You want to talk about “moral decay” like that article? Perhaps remove the beam from your own eye first.