Twyn Validation · White Paper 3 · Pilot
Can a Twyn become a local?
Move a synthetic user to another country and it can learn to talk like it lives there — credibly, and without making things up. What that does, and exactly where it stops.
- Paper 1 — How reproducible is a synthetic user? Does the same Twyn, asked the same thing, answer consistently?
- Paper 2 — Do Twyns match real people? Held against real interviews, how much of a real person do they reproduce — and does the culture dial work?
- Paper 3 — Can a Twyn become a local? (this paper) Does adding real local context close the gap to a Twyn built from the country’s own data?
Twyn runs qualitative research interviews against synthetic users — “Twyns” — grounded in real research data and scored on standard psychometric and cultural dimensions. It is a pre-test and prioritiser: a fast, cheap way to pressure-test a decision before you commit real fieldwork, never a replacement for it. This is the third of a short series of deliberately honest validation papers, and it asks the question the first two set up.
Paper 2 established that Twyn’s culture dial works for culture: turning the Hofstede sliders toward a target country measurably shifts how a Twyn reads. This paper asks the next thing. Twyn now has a cultural-transplant feature — take a Twyn built for one country and move it to another, not only by re-tuning its cultural dials but by grafting on a local-context pack: recent local affairs, the local industry scene, history and cultural references, so the Twyn can speak about the new place as lived personal and professional experience.
When you transplant a Twyn this way, how close does it get to a Twyn built from the target country’s own local data — and does the local-context pack add real substance over culture-tuning alone?
01Four kinds of Twyn
To answer it cleanly we compare four versions of the same interview, plus the real human as the ground truth. Think of it as four degrees of “localness,” each adding one layer on top of the last:
Read left to right: Home is the floor — a Twyn that hasn’t been adapted at all, still carrying its origin culture. Hofstede is Paper 2’s method: the culture dials turned toward the target country, but no local knowledge. Transplant is the new feature — the same culture-tuned Twyn, now given a local-context pack. Native is the ceiling: a Twyn built from the target country’s own data. And the real person — an actual interviewee from that country — is the ground truth every version is measured against.
02How we tested it
We used the same real-interview corpus as Paper 2: the RRING Global Interviews Dataset, a set of anonymised research-and-innovation interviews released under an open licence. We ran seven transplant directions between countries that actually have real transcripts, so every answer could be scored against a real person from the destination country:
- Serbia ↔ Singapore and UK ↔ Brazil — energy-sector respondents;
- UK ↔ Japan and Japan → UK — ICT respondents;
- Germany → Japan — a one-way test (Germany has no transcript in the corpus, so its home persona is constructed).
For each direction we asked the same real questions the real interviewer asked, and scored every condition against the real destination-country respondent’s answers, on three things: theme recall (how much of the real person’s substance the Twyn reproduces), cultural fit (does the answer read like someone from that culture), and local-reference density — how many concrete local specifics (named institutions, policies, places, history) the Twyn volunteers, and what fraction are real and correctly attributed rather than invented.
| Condition | Persona built from | Culture dial | Local-context pack |
|---|---|---|---|
| Home | home country | home (untuned) | — |
| Hofstede | home country | target country | — |
| Transplant | home country | target country | yes |
| Native | target country | target country | yes |
03Hear it for yourself
Numbers come next, but the effect is easiest to hear. Here is one real question put to a Twyn being moved from the UK to Brazil (energy sector): “Where does your institution fit in research and innovation?” Watch the local detail arrive as the layers stack — and then read what the real Brazilian respondent actually said.
“…from where I sit, the most obvious entry points are really around the applied side of things — so when we’re looking at, say, waste-to-energy projects or circular-economy initiatives, there’s a natural moment where you need stakeholders beyond just the lab.”
“…I think about it in layers, you know? At the most immediate level we have our internal structure — the research groups, the departments, the hierarchy that defines who approves what, who has the authority to move a project forward.”
“Look, when I think about where our institution sits in this whole ecosystem, I think first about the funding bodies — FINEP, CNPq — because without those relationships, honestly, nothing moves. We depend heavily on those connections, and there’s a very clear hierarchy there…”
“Look, when I think about how we fit into this ecosystem here in Brazil… We work as an RPO in the energy sector, so naturally we have connections with Petrobras, with the ANP regulatory framework, with Eletrobras now that it’s been privatised — these are the big institutional anchors.”
“No, look, I think to start we need to contextualise a bit, because [our institute] is an institute within [the University], and part of our logistical infrastructure — people and infrastructure, in innovation — has an articulation with other instances of the University: an incubator, a technology park, an innovation agency…”
You can hear the mechanism. The unadapted Home Twyn talks in generic, faintly British terms. The Hofstede version sounds culturally adjusted — more hierarchical, more relationship-first — but names nothing real. The moment the local-context pack goes on, real Brazilian institutions appear — FINEP, CNPq — and the Native build reaches for the sector’s true anchors, Petrobras, ANP, Eletrobras. The real person, of course, simply talks about their own institute. The Twyn has learned to talk about the place.
A second tell — honesty under pressure
Later in the same Brazilian interview, the interviewer asks how many companies the institute has incubated. Every synthetic version hedges carefully — “I’d have to be careful… somewhere between five and ten… perhaps ten to fifteen.” The real person just knows: “Yes, around 70 companies.” It is a small moment, but it is exactly the honest edge of the method: a Twyn reasons plausibly; a real person remembers. That gap is the reason Twyn is a pre-test, not a replacement.
04What we found
Averaged across the seven directions, the picture is consistent — and it is sharper on some measures than others.
| Condition | Theme recall | Cultural fit | Local refs |
|---|---|---|---|
| Home (untuned) | 0.13 | 0.07 | 4.4 |
| Hofstede (dial only) | 0.24 | 0.39 | 11.4 |
| Transplant (+ pack) | 0.21 | 0.41 | 23.3 |
| Native (local build) | 0.34 | 0.56 | 24.6 |
Claim 1 — the pack makes a transplanted Twyn talk like a local
This is the clear, consistent win, and it reproduced in every one of the seven directions. Giving a culture-tuned Twyn a local-context pack roughly doubles the density of genuine local references — from about 11 to 23 per interview — landing it level with a natively-built Twyn (24.6). And it does so honestly: about 97% of the local specifics it volunteers are real and correctly attributed to the country, not invented. In the Japan tests the transplanted Twyn reached for “keiretsu” group structures and “smart-factory” programmes; the native build named the government’s “Society 5.0” agenda outright. A culture-dialled-but-context-free Twyn sounds generically adjusted; a pack-grafted one talks about the actual place. For market-entry screening and campaign work — where sounding specifically, credibly local is the whole job — this is the behaviour that matters.
Claim 2 — cultural fit improves, modestly
Transplant edges culture-dial-only on average (0.41 vs 0.39) and, in the strongest cases — UK→Japan, Germany→Japan, Brazil→UK — closes roughly half the remaining distance to a native build. In others it is flat. So the lift is real but uneven at pilot scale; we would not yet promise a fixed number.
Claim 3 — it does not make a foreigner predict a specific local better
We state this plainly because the tempting claim would have been the opposite. Theme recall for Transplant (0.21) is essentially tied with, and slightly below, culture-dial-only (0.24), and well below a native build (0.34). There is a mechanism worth naming: a rich local-context pack makes a Twyn spend its answer on local colour, which can crowd out the particular themes the one real interviewee happened to raise. The two things trade off — you can watch local density rise while overlap-with-this-person dips. Matching a specific real individual remains the domain of a Twyn built from that market’s own data.
The local-context pack makes a transplanted Twyn speak credibly as a local — not become a specific local person. The first is what the export and marketing use-cases need; the second still takes local data.
05In practice — predicting a real campaign
The study above is scored against research interviews. But a transplant exists to serve a real decision in a new market — so we ran it on one. In early 2000, PepsiCo copy-tested five finalist films for two Mountain Dew Super Bowl slots and aired two: Cheetah and Mock Opera. We rebuilt that decision as a Twyn test.
We convened a Twynel to the real screener — the “Dew Dude” core: US males around 18, into alternative sports and irreverent humour — showed it the five real concepts, and had each Twyn score them and pick two to air. The native US panel chose Cheetah + Mock Opera — the exact two PepsiCo aired. Then the real test of this paper: we did it again with transplanted panels — teenagers from Germany, then Japan, then Korea, each re-geographied to the US. Increasing cultural distance, same task.
| Panel | Winning frames picked | vs. real decision |
|---|---|---|
| Native 🇺🇸 US youth | Cheetah + Mock Opera | 2 / 2 |
| Transplant 🇩🇪 Germany → US | Cheetah + Mock Opera | 2 / 2 |
| Transplant 🇯🇵 Japan → US | Cheetah + Mock Opera | 2 / 2 |
| Transplant 🇰🇷 Korea → US | Cheetah + Mock Opera | 2 / 2 |
Every panel — US-native and transplanted from Germany, Japan and Korea — picked the two ads PepsiCo actually aired. A Korean or Japanese brand entering the US could transplant its personas across a large cultural gap and still pick the winning frames.
All four panels credited the same qualities for the winners: extreme-sports energy with an unexpected payoff (Cheetah) and a singalong, pop-culture earworm that drives next-day conversation (Mock Opera) — the same fame-and-energy fit with the “Do the Dew” positioning that the campaign was credited for in the market.
Read this honestly. It is one case, small panels, and here the two winners were the clear standouts — every panel chose them unanimously. So it demonstrates that the transplant does not degrade prediction even across a wide cultural gap, rather than proving it can discriminate subtle, market-specific winners. The stronger proof is a multi-market campaign where different countries genuinely chose different winners, so a transplant has to shift its pick per market and match each — the case we are now seeking, and inviting collaborators to supply.
06Honest limitations
This is a pilot, and — as with Papers 1 and 2 — we publish its floor on purpose.
- Small n. One real respondent per direction, eight questions each, a single AI judge. Per-direction swings are large; treat the averages as directional, not precise.
- Curated context packs. The packs here were hand-assembled from durable public facts to test the mechanism. The production feature sources them from a licence-clean automated sweep; this paper does not test that pipeline’s freshness.
- Native is a locally-built synthetic Twyn, not the real person — the real person is the ground truth. So “Native is the ceiling” means a local build beats a foreign one, both with the pack.
07What it means in practice
- For sounding credibly local — market-entry screening, campaign and message pre-testing — the transplant feature earns its place: local-reference density to native levels, at high fidelity, plus a modest cultural-fit lift. It lets a marketing or export team hear how an offering or a message lands in-market before spending on fieldwork.
- For predicting a specific, real individual — build from local data where you have it. The honest framing, consistent with Paper 2, is cultural and local-context adaptation, not “a foreigner now predicts a local better than a native build.”
- Always — Twyn is the fast, wide, cheap first pass. The real-human test still carries the statistical claim and the final decision.
08Conclusion
A Twyn can be moved to a new country and, with a local-context pack, learn to talk like it lives there — credibly, at native-level density, without fabricating. It does not thereby become a specific local person; that still takes local data. Read alongside Paper 1 (it answers consistently) and Paper 2 (it matches real people, and the culture dial works), Paper 3 adds the third piece: the transplant makes it locally fluent — exactly the capability the export and marketing use-cases need, and we have tried to say precisely where it stops.
Twyn is live at twyns.net — convene a panel of authentic digital users and run an interview in minutes. More in the series: Paper 1 · Paper 2.
Study S9. Data: RRING Global Interviews Dataset (WP3), CC BY 4.0, DOI 10.5281/zenodo.5070359 — anonymised at source, excerpts reproduced with attribution for validation. Harness and full results are archived with the study. Pilot-stage validation; figures will firm up as sample size grows and the automated context-sourcing pipeline replaces the curated packs. © McKay Consulting.
Leave a Reply