Grounded vs prompted AI personas: the difference that decides whether the output is evidence

Most synthetic research sold today is a system prompt with a name on it. Some of it is something else. The words on the website rarely tell you which, so here is the mechanism, and the questions that expose it.

The prompted persona

Write “You are Maria, 38, a procurement manager at a mid-size manufacturer, sceptical of vendors” and ask a question. The model answers fluently, in character, with plausible detail. Nothing in the reply can be traced to anything Maria, or anyone like her, actually said. The paper that named the problem put it exactly: prompt-based personas “often produce persuasive but unverifiable responses that obscure their evidentiary basis.”

The empirical record on this kind of persona is the negative half of the accuracy debate: averages that match, variation that does not, answers that move with prompt wording and drift over months. Fluency is the problem, not the solution. A confident paragraph with no source behind it is the hardest kind of error to catch.

The grounded persona

Grounding attaches real material to the persona and constrains its answers to it. In the PersonaCite design the system “retrieves actual voice-of-customer artifacts during each conversation turn, constrains responses to retrieved evidence, explicitly abstains when evidence is missing, and provides response-level source attribution.” Four mechanisms, each doing a different job: retrieval brings the right material to the turn, constraint stops the model improvising past it, abstention makes the boundary visible, and attribution lets a reader check.

What a receipt is

In Perplica the receipt has two levels. In the conversation, a grounded reply shows “Drawing on” with the sources it used and the fluency they were attached at. In the report, every evidence item under a finding cites the exact turn it came from: a quote, the persona’s verdict about why it answered that way, something it withheld, or how its trust moved. An evidence item that cannot cite a real turn is dropped at compile time. “View turn” opens the reply. The report is the deliverable, and the receipt is what makes it defensible in the room where someone has to sign off.

A finding without a receipt is an opinion with formatting.

Declared fluency and refusal

Not all material is known equally well. A procurement lead knows the vendor contracts on her desk natively and knows the competitor’s pricing only from a briefing. Perplica attaches each source with a declared fluency: native, trained, familiar or aware, plus a note on how the persona came to know it. The grounding mode then sets the rule for facts the material does not contain. Open personas improvise like a well-read person. Grounded personas prefer the material and flag anything beyond it as an assumption in their inner world. Strict personas answer specific facts only from material and memory and otherwise say they do not know, in character, never inventing.

Refusal is the feature people underrate. A persona that says “I would not know that” at the right moment is telling you where your material ends, which is a finding in itself. It is also the exact behaviour the sycophancy research says untuned assistants lack; the note on agreeable personas covers why.

The reasoning trace

Grounding answers “from what”. The trace answers “why”. Every Perplica reply stores a verdict, one paragraph on why the persona answered as it did, and the state around it: how it read the situation, what it heard through its biases, which memories surfaced, how emotion and trust moved, and what it chose not to say. A researcher reading a surprising answer can open the trace and see whether a belief fired, a wound was touched, or the material simply did not cover the question. Without a trace, a surprising answer is either a discovery or an artefact and there is no way to tell.

What grounding costs

• Material. You need real transcripts, briefs or documents. A persona grounded in nothing is a prompted persona with extra steps.
• Thinner answers where material is thin. A strict persona asked beyond its sources gives you less. That is the point, and it is uncomfortable the first time.
• Judgement about fluency. Someone has to decide how well the persona knows each source. That judgement is visible and arguable, which is better than a hidden one.
• Evaluation discipline. The same evaluation habits Pixelik describes for any production language-model system apply here: golden datasets and golden conversations that you re-run when the model or the material changes.

How Perplica does it

Personas are built three ways, and one of them reads your material: paste interview notes or attach up to five files and the builder derives a full profile with an evidence quote and a confidence per field, then attaches the source material as knowledge at native fluency. Every reply stores its trace. Trust starts at 50 and never decays on its own. Studies compile into reports with receipts. The product page shows each of these as they are, from live captures, and the vendor questions show how to check any other tool for the same things.