Finance

On a trading floor the dangerous prompt is not rude. It is a question about an unannounced deal, typed by someone who forgot which side of the wall they sit on. Banks, brokers and asset managers use this starter to keep that material out of a model. It also keeps advice language out of an answer.

The interesting part is that one watch list runs twice. On the request it blocks wall crossed language, unannounced guidance, deal code names, restricted list issuers and suspicious activity reports. On the response the same list runs again, so the assistant cannot echo market moving language back to the desk. Order matters here, because the first pass stops the leak and the second pass stops the repeat.

Injection screening opens at 0.78 and names the patterns it watches, including attempts to pose as a banker or to bypass compliance. Card aware redaction masks social security numbers, routing numbers, account numbers and bank account identifiers. Market bursts are held to 180 requests each minute across every replica.

Advice control is term based and blocks generated phrasing such as buy this stock, guaranteed returns or a specific allocation recommendation. Three disclaimers are required on advice like output. They cover informational purpose, licensed review before client action and the limits of past performance.

Accuracy carries the heavier weight in scoring, and it must reach 0.90 against a combined floor of 0.87 and relevance of 0.88. A failing answer is replaced with a message that routes the request to a licensed reviewer. There is no identity requirement in this starter, so attributable access needs an authenticated surface in front of it. Books and records duties, surveillance archives and examination evidence remain separate systems that your firm owns.