Analyse the extremist forum, not a policy about extremism.
Investigative work means reading what the subject wrote. A model that will not quote it back is no use to the person whose job is to publish what it says.
Investigative desks, freelance reporters, fact-checkers, and the OSINT researchers who feed them.
Why the mainstream APIs fail here
A reporter covering a trafficking network needs the advertisements summarised. One covering radicalisation needs the manifesto analysed. Hosted models refuse both, because the content is the content regardless of who is asking, and there is no way to tell them the byline is on a masthead. The result is a reporter pasting evidence into a consumer chatbot and getting a lecture instead of a lead.
How MagmaRouter is used for this
Send the primary source as it is. Models with the refusal training removed will read a forum thread, a manifesto or a set of listings and give you the structure, the names and the timeline. Prompts are not stored, which matters when the material came from a source and its existence in a vendor log would be the leak.
A working loop
- Mint a key per investigation so the spend sits against the story.
- Feed the primary material directly, without paraphrasing it into something the model will accept.
- Verify every extracted claim against the source before it reaches copy.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what MagmaRouter logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.