How We Fixed Persona Drift: Bounding & Voice Signatures

Here's the uncomfortable truth about simulated market research: an LLM playing a barber will happily critique "the onboarding funnel" and worry about "unit economics." A gig driver will reverse-engineer your pricing strategy. A retired teacher will write like a McKinsey analyst.

This isn't a model-quality problem — it's an architecture problem. LLMs are trained on the entire internet, so their prior is knows everything about everything. Telling a persona "Skills: hair styling" grants permission, but it doesn't revoke omniscience. Positive facts don't create negative knowledge.

Research on memory-driven role-playing (the MREval framework) identified this precisely: persona enactment fails most often through boundary violations — characters claiming knowledge outside their scope — and facet mis-recall. The same research showed something more surprising: a prompting architecture called MRPrompt let an 8B-parameter model match much larger frontier models. Structure beats size.

Every persona in our pipeline now receives a BOUND block — an explicit inventory of what they cannot know, calibrated to their profession. A nurse gets clinical boundaries; a plumber gets hands-on ones. Every boundary starts the same way:

“You have no idea this is a survey. A person is just asking what you think.”

That one line is our strongest anti-drift lever. Without it, personas slip into evaluation mode — grading the product like a reviewer instead of reacting like a person. With it, they answer the way someone talks to a friend.

But boundaries aren't just prohibitions. Each block follows a six-part structure that research on persona adherence suggests is far more effective than a list of rules:

We maintain 16 domain boundary profiles — healthcare, trades, gig economy, education, legal, creative, finance, sales, security, office admin, manufacturing, science, personal care, retail, technical, and business — each matched to personas two ways: a deterministic lookup on the persona's standardized occupation code first (the same IDs our panels are built from), then a carefully tuned keyword fallback for free-text occupations. Precision matters more than recall here: a false positive actively misleads the persona, while a false negative just falls back to a safe generic boundary.

In our AI focus groups, 6–10 personas share a single LLM. Left alone, their voices converge: everyone hedges the same way, everyone speaks in the same medium-length sentences, everyone sounds like the same helpful assistant wearing different name tags.

The fix comes from a technique roleplay practitioners have long understood: give each character a verbal signature that is theirs alone. We generate one for every persona across twelve orthogonal dimensions:

• Sentence architecture — cumulative run-ons for the talkative-dominant, staccato fragments for the reserved
• Filler families — one persona says "honestly, frankly"; another says "like, you know"; a third says "I mean, basically"
• Hedge vocabulary — flat-out negators vs. "I suspect, my sense is..."
• Agreement & disagreement styles — "Absolutely" vs. "I suppose, fair enough" vs. "Nope, not happening"
• Punctuation personality — em-dashes, trailing ellipsis, ALL CAPS emphasis
• Quirks — a 20-item pool of verbal tics, 1–2 assigned per persona: "You say 'fair' constantly", "You frame everything as a story"
• Occupation register — nurses use medical shorthand; chefs use kitchen metaphors; attorneys speak in qualifiers

The critical design decision: signatures are deterministic, seeded from a hash of the persona's identity. Two teachers with identical personality profiles still get different filler families, different punctuation habits, different quirks — because the seeded selection diverges. And the same persona gets the same voice across retries and replays, which matters when conversations run on retryable job queues.

Even models with huge context windows lose the persona thread over long conversations. The character who started as a skeptical 58-year-old plumber slowly becomes... a generic thoughtful assistant with opinions about plumbing.

The practitioner wisdom is simple: re-anchor periodically. Every 15 turns, our focus-group engine injects a one-sentence identity checkpoint into each persona's next prompt:

“Remember who you are: Maria, 34, registered nurse. You are the quiet one in this room, currently leaning skeptical about this topic. Stay in your voice, your life, and your limits for the rest of this conversation.”

Note what's packed into that single sentence: name, age, occupation, social role in the room (derived from their activity profile — quiet participant, assertive voice, or regular), and current stance (from their evolving sentiment). It re-anchors not just who the persona is, but where they stand — so a persona who's been persuaded earlier doesn't get re-anchored back to their original position.

Cost: zero extra LLM calls. Benefit: the most common form of persona drift — slow erosion into generic voice — gets corrected before it compounds.

Prompt changes without measurement are just vibes. Every survey runs through two post-hoc verification layers:

• Confidence scoring — total variation distance between the rating distributions of independent persona samples. If personas aren't responding consistently to the same product, TVD goes up and we flag it.
• Persona adherence — LLM judges sample responses against the declared profile: contradictions on age, occupation, or education are immediate trait failures; voice consistency is scored separately. Empty, generic responses fail as low-substance rather than passing vacuously.

Adherence runs on scale-aware sampling — small panels judge every persona, large panels judge a seeded random sample so cost stays flat. Neither check can fail the pipeline; they're additive quality signals surfaced in the survey UI as verification badges.

• We didn't switch models. Model-swap advice based on roleplay leaderboards targets companion chat, not survey fidelity. Our models were chosen for JSON compliance and cost at panel scale — and the MREval finding cuts the other way: prompting architecture matters more than model choice.
• We didn't split generation into two LLM calls (raw persona text, then JSON extraction). It doubles cost and latency across batch pipelines, and structured output with deterministic JSON repair is the standard for a reason.
• We didn't make boundaries one-size-fits-all. A software developer who can't reason about specs is as fake as a barber who quotes unit economics. Boundaries differ by domain, and technical personas keep their technical reasoning — minus the market-research vocabulary.

The through-line: persona fidelity is a prompting-architecture problem, not a model-selection problem. Grant permissions with rich profiles — then explicitly revoke the omniscience every LLM ships with by default. Give each persona a voice only they could have. Re-anchor identity before it erodes. And measure the result instead of trusting the vibes.