Can Cultural Conditioning Survive Negotiation Pressure?
We made the same two synthetic negotiators close the same B2B deal 100 times — 50 in Great Britain, 50 in the United States — changing nothing but country conditioning. Some cultural signals weakened under pressure. Others got stronger. That asymmetry is the finding.
Most synthetic datasets that claim to capture culture change everything at once: country, personality, age, industry, scenario. Then they attribute whatever differences appear to “culture.”
We wanted to know what happens when you change only the country.
So we built a controlled experiment: the same two synthetic identities — a 47-year-old VP of Procurement (buyer) and a 36-year-old SaaS startup founder (vendor) — negotiate the same enterprise procurement deal 100 times. Fifty conversations are generated with British cultural conditioning, fifty with American. Same psychometric vectors. Same demographics. Same narrative arcs. Same seeds. Same planned communication acts.
The result is now public: stratasynth-cross-cultural-negotiation on Hugging Face.
This article is the story behind it: why we built it, what we were trying to measure, what surprised us, and what we still don’t know.
Why we built it
Earlier this year, an external research team back-tested our synthetic personas against real human survey data across countries. The result was uncomfortable: our US and GB personas were psychometrically almost interchangeable. The flagship cross-country contrast in interpersonal trust — well documented in human data — came out completely flat.
The diagnosis was structural. Personality norms for two Anglophone countries are nearly identical, and the attitudes that do differ between countries — trust in institutions, relationship to authority, comfort with directness — are not personality traits at all. They are cultural and institutional context. Our personas simply had nowhere to store them.
So we built a cultural conditioning layer with one strict rule: no survey data allowed. The conditioning derives from sources independent of any attitudinal ground truth — Hofstede’s six cultural dimensions and the World Bank’s Worldwide Governance Indicators. If country-level differences emerge from those priors and happen to match patterns observed in human data, that is out-of-source generalization, not circular injection.
Then came the question this benchmark exists to answer. A counterpart researcher working on negotiation simulation made a prediction before seeing any of our results:
Culture appears when the system has to choose between conflicting norms.
In other words: anyone can sprinkle “perhaps” into British sentences. The real test is what happens under pressure, when politeness competes with the need to push back.
That is a measurable claim. We decided to measure it.
What we were trying to measure
The design goal was to eliminate every confound we could think of:
| Held constant | How |
|---|---|
| Personality | Explicit psychometric seeds (Big Five + Schwartz values + attachment), identical in both halves |
| Demographics | Explicitly fixed: buyer 47F VP of Procurement, vendor 36M SaaS founder |
| Scenario | Enterprise procurement negotiation, restricted to professional narrative arcs |
| Conversation structure | Identical arc distribution (17/17/16), identical turn counts (1,172 per country) |
| Behavioral plan | The decision engine selects each turn’s communication act before text generation, from the same seed |
| Varied | Country conditioning: GB vs US |
One detail matters here: in our pipeline, intent, goal, and communication act are computed deterministically from the persona’s belief state before any language model runs. With shared seeds, both halves of the corpus follow the same behavioral plan. Culture can only show up in how that plan is rendered into language — which is exactly where we wanted to isolate it.
We also defined in advance what failure would look like. If a reader could identify the country in thirty seconds, we would be looking at stereotypes. Zero detectable difference would mean the conditioning does nothing. The honest target was the uncomfortable middle: real, systematic, subtle.
What surprised us
Surprise 1: our scenario descriptions never reached the language model
The first pilot run produced professionally plausible conversations — about server logs and project deadlines. Not procurement. Not negotiation. Only ~2% of turns contained any negotiation vocabulary.
Tracing the pipeline revealed an embarrassing, instructive gap: the scenario’s description existed in our catalog, was validated by our API, was visible everywhere — except in the prompt that generates conversation turns. Topics had been emerging from personality and emotional arcs alone. Our family and relationship datasets never exposed this, because relational domains do emerge from psychology. A B2B procurement negotiation does not.
One small, explicitly scoped fix later (domain grounding for this scenario only, byte-identical output for every existing scenario), negotiation vocabulary went from ~2% to ~30% of turns, with buyer and vendor roles correctly assigned. The lesson generalizes: in multi-stage generation pipelines, check what actually arrives at the prompt. Metadata that every other layer can see may be invisible to the one layer that matters.
Surprise 2: the signal lives in style, not content
With the topic pinned by design, both halves of the corpus discuss the same things — SLAs, pricing tiers, compliance, references. Topic distributions converge. What diverges is how the same negotiator behaves:
| Metric | GB/US ratio |
|---|---|
| Hedging (“perhaps”, “I wonder”, “a bit”…) | 1.38x |
| Hedged disagreement | 1.79x |
| Authority references (board, legal, committee…) | 1.5–2.1x |
| Words per turn | 1.07x |
And crucially, no clichés: zero instances of stereotypical British or American expressions across 2,344 turns. An external reviewer’s first impression was that the conversations looked “surprisingly homogeneous” — which we now consider a feature. If culture were visible in thirty seconds, it would be costume, not conditioning.
Surprise 3: pressure does not erase the signal — it transforms it
Every turn carries its arc-planned tension value, identical across countries by shared seed. That allows a clean comparison: how do the same cultural markers behave at low, medium, and high pressure?
| Marker (GB/US ratio) | Low tension | Mid | High |
|---|---|---|---|
| Hedging | 1.73 | 1.34 | 0.66 |
| Direct disagreement | 0.84 | 1.96 | 1.91 |
| Authority appeal | 2.12 | 1.54 | 1.56 |
| Explicit apology | 0.94 | 0.47 | 0.52 |
Read that first row again. At low tension, the GB half hedges 73% more than the US half. At high tension, it hedges 34% less. The classic British politeness shield is real in this corpus — and it drops when the negotiation gets hard, giving way to direct disagreement (1.91x) and appeals to institutional authority (1.56x). The US half does something different: it holds its register and repairs friction with explicit apologies, at roughly twice the British rate under pressure.
Some cultural signals weaken under pressure. Others strengthen. The contrast between the two halves is not a constant offset — it is a pressure-dependent profile.
Which is, as far as we can tell, exactly what the pre-registered intuition predicted: culture is not a vocabulary. It is which norms win when norms conflict.
What we still don’t know
We want to be precise about the limits, because the design is the product here.
Two synthetic individuals are not a culture. The fixed-pair design is what makes the experiment clean — and what limits generalization. We measured how country conditioning shifts the behavior of one buyer-vendor pair, not how British or American professionals negotiate. The dataset card says this explicitly, and we mean it.
Our markers are lexical proxies. Hedging, directness, and authority appeals are measured with deterministic regular expressions over the text — reproducible by anyone from the published columns, but shallow compared to full pragmatic annotation. Better instruments may find structure we missed, or weaken findings we report.
The pressure analysis rests on ~190 high-tension turns per country. The monotonic trends across bands are more trustworthy than any single endpoint, and the low-count rows (apologies at high tension) should be read as directional.
Blind validation is pending. The right test of “real but not caricatured” is whether independent evaluators — human experts and models from other families — can classify country from text at meaningfully above chance but well below certainty. Our own expectation is 60–75% accuracy. A blind version of the corpus exists for exactly this protocol, and we would rather publish the number than predict it.
One scenario, one dyad, two countries. Whether the pressure-dependent profile replicates across other scenarios, other persona pairs, and the other countries in our conditioning layer is the obvious next experiment. If “hedging converges, authority persists” turns out to be a general property of the conditioning — or breaks somewhere informative — either result teaches us something.
The point
Synthetic data is usually evaluated by asking “does it look real?”. We think the more useful question is “can it isolate a variable that real data cannot?”.
No human dataset can give you the same two negotiators, with the same personalities and the same plan, running the same negotiation in two countries. A synthetic pipeline can — if the architecture separates cognition from rendering, and if you are willing to publish the controls along with the conversations.
The dataset, the methodology, the headline numbers, and the full list of limitations are on Hugging Face. The markers are reproducible from the published columns in a few lines of pandas. If you find something we missed — in either direction — we want to know.