Methodology
Technical documentation, for anyone who wants to check the engine before trusting it. The short version is on the home page. Every assumption here is versioned, and wherever a value is a stated default rather than a measurement, it says so.
Proxima assembles a simulated audience from real behavioural distributions and puts your scenario in front of it, so the objections, segments and tensions that usually stay implicit get said out loud. It is not a predictive model: it has not been calibrated against real launches. Read the output as a structured argument about a scenario, not a claim about what will happen.
The principle is “from what IS to what IF”: start from how the market actually behaves, using Nextatlas data, then simulate how it reacts to a change.
The main module, experimental like the rest. Describe a scenario in plain language, say “launch a premium plant-based chocolate in Germany for Gen Z at €4.99”. The engine then:
- Takes the population reference you check and confirm, say chocolate bar, and asks the Nextatlas API for the demographics, professions and attitudes of the people talking about it. Nothing is guessed from your prose. The API holds data even for a phrase like “spring 2027”, so the first term that answers is not automatically a good reference. Need states are not in the data: they come from a stated default, and every run says so.
- Builds an audience of 100 to 1,000 profiles, using those values as sampling weights. They are relative relevance rather than audience shares, and the run records that transformation.
- Runs the diffusion model round by round across the whole audience: who adopts, who holds out, how the message travels along the social graph. This is where the adoption curve comes from, and it is finished before anything is asked of a language model.
- Then asks a sample of those profiles, through Claude, to react in their own voice: quotes, doubts, enthusiasm. The sample is a subset rather than a statistically representative one, every result reports how many answered, and their answers are recorded beside the simulation rather than over it. They do not move the curve.
- Returns the adoption curve, a segment breakdown, friction points, representative quotes and recommendations written by Claude.
Nine editorial archetypes, written by hand rather than found in the data, plus any persona you define yourself. Chat with them, put them in a focus group, and watch the trends each one follows move over time. What evolves is the associated trends, not the person’s behaviour.
Upload two or three creative variants. A panel of simulated profiles reacts to each, and you get sentiment by segment and the recurring themes in the replies. When the panel cannot separate the variants, it says so instead of naming a winner.
A map of a category's signals built on the Nextatlas API, with a synthesis written by the model. The growth shown for each trend is Nextatlas's own projection, frozen at the moment the map was built: it does not move on its own, and stays as it is until you press Refresh. The what-if is not a simulation and carries no magnitudes. It is the model's qualitative reading of the current map.
Put two runs side by side to see where they agree and where they part: useful for choosing between two markets, two positionings, two timings. When the two differ in exactly one declared parameter and nothing else, Compare says so.
The distributions that characterise a trend, meaning who its adopters are, what they value and in which professions, come from the Nextatlas API. Every run tells you how much of the audience is real: Grounded on Nextatlas only when all three dimensions the API provides (demographics, professions, attitudes) came back real, need states being always a default because the API has none; Partially grounded when some are real and the rest fell back to defaults; Default population when the API did not answer. Three limits, stated rather than buried. Geography is not passed with the query, so the distributions are the term’s worldwide. Each attribute is drawn independently from its own marginal distribution, so real correlations between attributes, say between profession and attitude, are not preserved. And the API returns the top few values per dimension as relative relevance, not as shares of the audience.
If the pill is missing, the run used the default population, and only ever because you allowed it after a reference had no data. Check a term the data is likelier to recognise, mushroom coffee rather than fungus drink, to bring the grounding back.
Not every profile is the same. They are split into four tiers, after the trend adoption cycle:
The proportions are a stated, versioned assumption, not a measurement: nothing in the Nextatlas data says a market is 5% protagonists. Written reactions normally use Claude. Which providers see what is under Data & privacy.
Each tier carries a ceiling on willingness to convert at all: a Passive profile has a 14% chance of ever giving in, a Protagonist 96%. That is what makes adoption saturate at a plausible level, typically 40 to 55%, instead of climbing to 100% the way a naive model would.
Each round, for every profile that has not yet adopted, the engine combines three forces into a probability:
- Social pressure, 45% of the weight: how many of your direct contacts have already adopted.
- Alignment with needs, 30%: how far the scenario meets your own need states.
- General trend growth, 15%: how warm the theme is right now.
Adoption travels across a small-world network: each profile is connected to six neighbours, clustered by demographic affinity, so it spreads through those ties rather than uniformly.
A sample of protagonist and influencer profiles is asked, through the LLM, with that profile alongside the scenario. The size scales with the audience and the quality setting, and every run reports how many were asked and how many answered. Their first-person replies are what you read in the Verbatim and Friction map sections. They are recorded beside the simulation, so a reaction never alters the adoption curve it was collected after.
Quotes often come back in the local language: a Berlin profile answers in German, a Turin one in Italian. That is the model holding the cultural context, not a bug.
When the run finishes, Claude reads the whole analysis, curve, segments and friction points, and writes an editorial executive summary plus four or five operational recommendations. They carry the run’s own numbers rather than generic advice.
- Nextatlas aggregates signals at the level of patterns, not individuals: no personal data enters from that side.
- But the free-text fields are yours. Scenarios, campaign copy, persona definitions and chat messages are sent to the model providers exactly as you write them. Nothing strips personal data out, so do not put any in.
- Your runs are stored in Supabase with Row Level Security on, and every backend query is bound to your user_id: no user reads another user’s data.
- What each provider receives: Anthropic gets the scenario or message you wrote plus the synthetic profile it is shown to. OpenAI or DeepSeek get the same when one of them is selected instead, which happens when the Anthropic client cannot be initialised, for example with no key configured. A call that fails is not retried against another provider: that reaction is dropped, and the run reports how many were asked and how many answered. Nothing about your account is added to those prompts: not your email address, not your workspace identifier. What is sent is the text you wrote and the profile it is being shown to, so an address you type into a brief is part of the brief.
- Hosting: the backend runs on Railway in a US region today. If data residency matters for your work, talk to us before you put anything sensitive in a brief.