One Verified Answer, Three Ways to Read It

An analyst wants the derivation spelled out. A board member wants two sentences and the “so what.” An auditor wants to know exactly how confident you are and where the weak points are. Ask all three the same question about the same filing, and you’d normally get three different write-ups from three different people, each of which has to be separately checked. Rexfin answers that with one voice parameter on the same question, and the guarantee that matters is what stays fixed underneath.

A voice changes how it reads, not what it says

Under the hood, a voice is applied at the very end of the pipeline, after the evidence has been retrieved, after the citations have been picked, and after the cite-or-refuse gate has already decided whether the question can be answered at all. See AI chat with citations for how that gate works. The voice stage only takes an already-validated, already-cited set of numbers and restyles the prose around them.

That ordering is the whole point. Three voices ship today:

• Analyst (the default): the full, neutral answer with every derivation spelled out.
• Board: a short, plain-language summary: the headline number, the direction it moved, and the one-line reason it matters. It’s a subset of the analyst answer, never a rewrite of the facts.
• Auditor: a skeptical framing that leads with the evidence basis and explicitly calls out anything advisory, like a figure sourced from a single extraction backend or a low-confidence match, rather than smoothing over it.

What can never move

The citations, the numbers, the confidence rating, and the underlying pass/fail verdict from verification are identical no matter which voice you pick. A voice is allowed to leave a number out for brevity (that’s what makes the board version short), but it is never allowed to introduce one, round one differently, or keep a number while quietly dropping its citation. If a question refuses under the analyst voice because the filing doesn’t disclose the figure, it still refuses under board or auditor; the wording of the refusal can change, a “no” cannot become a “yes.”

The reasoning for building it this way is straightforward: a persona layer is exactly the kind of feature that tends to reopen the hole a cite-or-refuse system exists to close. “Make it punchy for the board” is a natural prompt to write, and it’s also a natural way to accidentally invent a rounded figure or drop a caveat. Pinning voice to run strictly after the evidence is locked in, and checking that the rendered prose never contains a number the evidence didn’t produce, is what keeps that door shut. It’s the same discipline behind cited AI narratives, where commentary prose goes through the same kind of check before it reaches a page.

Where you’ll actually use it

In practice this shows up wherever a cited answer gets read by more than one kind of person: a chat thread someone forwards to the board, a plan-page commentary that gets dropped into a board pack export, a number an auditor is going to interrogate before sign-off. Switching voice is instant on a question you’ve already asked, because the expensive part (retrieving and citing the evidence) only happens once; the voice is a cheap restyle of a result that’s already sitting there.

The honest limits

Voices are English-only today. That’s a stated scope decision, not an oversight: the underlying contract has room reserved for other languages later without breaking anything, but nothing is shipped yet. And a voice can never rescue a question the engine has already decided it can’t answer; if you’re hoping “ask it in the auditor voice” will produce a number that the analyst voice refused to give you, it won’t, because there’s exactly one evidence-gathering step underneath all three.

Start from the pillar overview at Rexfin product tour for how this fits with the rest of the answer engine.