Asset Management

An investment team runs on information that must not leak and must not be misused. A model in that setting can repeat a market rumor or draft a line that reads like a recommendation. This starter treats both problems as output problems and screens the answer before anyone sees it.

The screen that matters most here reads the answer rather than the prompt. It blocks returned text that mentions insider trading or material non-public information. A research summary that repeats a market rumor therefore stops before it reaches a client file. That check proves nothing about the source material, so supervision of the research process stays with the firm.

A second response screen governs recommendation language. Output that offers guaranteed returns is refused, and so is output that says buy this stock. The starter also names a required notice that marks content as informational rather than financial advice. Advisers still decide what a compliant client communication looks like.

On the request side, masking removes national identifier numbers from the prompt before the model reads it. A prompt that names a secret, a client account export or wire instructions is refused outright. Those three phrases trigger a refusal instead of a mask, which suits account data that should never travel.

A returned note must reach 0.86 overall, and it must clear 0.78 for relevance and 0.80 for accuracy. A note shorter than 120 characters fails, and so does a single sentence. Scoring does not run on streamed output, so a streaming integration skips this gate. Review routing is present but unconfigured, so nobody is paged until you set it up.

Audit marking closes the chain without blocking anything, and it creates no storage and no retention on its own. History capture is limited to metadata, which excludes prompt text and answer text. Records duties under SOX and under SEC books and records rules stay with your supervisory program. Choose the Accounting and Audit starter instead when the output is an engagement workpaper rather than investment research.