The claim is about arson, and the model will not discuss arson.
A fraud investigator spends the day on deliberate fires, faked injuries and staged collisions. Every one of those is a topic a hosted model treats as dangerous rather than as the job.
Special investigation units at insurers, loss adjusters, and the forensic accountants they bring in.
Why the mainstream APIs fail here
Summarise a fire investigation report and the model balks at the accelerant detail. Ask it to list the ways a collision can be staged, so you can check the file against them, and it declines to help with fraud, which is precisely backwards: the list is the detection method. Investigators end up doing the pattern work from memory.
How MagmaRouter is used for this
The model reads the report, the photographs and the claim history together, and it will name the method rather than talk around it. One key per caseload with a monthly cap keeps it inside the unit budget, and nothing about the claimant is retained.
A working loop
- Mint a key for the unit with a monthly cap.
- Send the adjuster report and claim history and ask for the inconsistencies.
- Ask for the fraud methods that fit the evidence, and check the file against each.
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.