Prism reads what your data contributes,
the coverage behind the number, the score, the rules of your jurisdiction, and the verdict on the deal you propose.
Observations arrive bucketed and anonymised. Prices, deal ids, counterparty names and account identifiers are refused at ingest, and each accepted batch is linked to the one before it. The audit log refuses every update and delete at the database itself.
Prism assembles the signals it can observe on a name and records the source that produced each value. A name with too little behind it comes back as a refusal that lists what is missing.
A 0 to 100 figure lands in one of four bands, and the confidence band widens as the data thins: five points either side at full coverage, twelve at partial. The model version and the weight table sit on the score itself.
A threshold rule returns the values it used and a readable line of the comparison, so a compliance officer can point at the figure that moved. When the signal behind a threshold is missing, the rule comes back skipped with the reason named.
The check computes exposure against your approved limit, reads screening state and settlement behaviour, and runs a published cascade in which the hardest verdict wins. Insufficient data comes back as a verdict of its own, with the missing input named.
Settlement reliability, pair by pair.
Each name is rated per currency pair and per time of day, weighted toward recent behaviour. Below twenty observations the rating is marked cold start.
