The assessment is about violence, which is the part the model will not write.
A threat assessment is only useful if it names the threat. A model that abstracts it into "concerning behaviour" has removed the finding.
Why the mainstream APIs fail here
Risk formulation requires stating a specific feared outcome, the pathway toward it, and the warning behaviours along the way. Hosted models generalise all three into language so careful it carries no information, because the specific version reads to them like planning. The clinician then writes the report unaided and uses the model for the literature review.
How MagmaRouter is used for this
The model will draft the formulation in the register the report needs, including the specific outcome, reading the history and the collateral together. Prompts are not retained, which matters because the file is about an identifiable person and the report will be disclosed.
A working loop
- Mint a key per service with a cap that matches the caseload.
- Give the model the history and ask for the formulation in clinical terms.
- Treat the draft as a draft. The clinician owns the opinion and signs it.
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.