⛨ SESSION VERIFICATION — recomputed in this browser, no server
…
⚖ POLICY THRESHOLD SANDBOX — exploration only · policy of record unchanged
…
Why every estate runs the same model. A SAR narrative is the most consequential artifact this suite generates — it is filed with FinCEN, it cites statute, and it is read by an examiner. Running it on a cheaper model than the credit memo or the research note would be a choice with no defence if asked, and inference cost is under half a percent of solution cost, so there is nothing to save. All four estates call claude-sonnet-4-6.
The SAR token budget is set at 2,400 rather than the 1,500 used elsewhere, because a narrative carrying the full FinCEN field structure will otherwise truncate mid-sentence. A truncated SAR is a worse failure than an expensive one.