Why Najem would rather say "unsettled" than give you an answer

A model that answers every question is easy to build. A model you would repeat to a regulator is not. The difference is almost entirely in what the system does when it does not know.

Retrieval before generation, always

Najem does not answer from what a model absorbed during training. Every question is resolved against our own corpus first — the instruments, rulebooks and regulator publications catalogued in the knowledge base — and the answer is composed from what came back. If retrieval returns nothing on point, that is the answer.

This is why our coverage page lists what we hold rather than what we cover. Those are different claims, and only one of them is checkable.

Citations link out, not in

Every reference points at the issuing body's own page, not at a copy we host. Two reasons. You can verify it without trusting us, which is the only kind of verification worth anything. And when the source changes, you see the current text rather than our snapshot of it.

Where we hold something that has since been superseded, we say so on the entry rather than quietly removing it. A superseded instrument is often exactly what you need when you are looking at a transaction from three years ago.

Each advisor publishes its limits

The fifteen advisors each carry a written remit and a written statement of what they will not do. This is not modesty. An advisor that is vague about its boundary is one whose output you cannot scope, and output you cannot scope is output you have to check line by line — at which point it has saved you nothing.

Najem advises; it does not act

Najem drafts, analyses, compares and recommends. It does not file, sign, submit or decide. That boundary is deliberate and we are not planning to move it. A person makes the call and a person puts their name to it, and the tool's job is to make sure that person has read the right thing first.

Everything above is testable. Open the knowledge base, pick an entry, and follow the link.