Synthetic Identity Engineering: a working definition
Most synthetic personas are costumes. Synthetic Identity Engineering is the practice of building the person underneath, belief systems and defense mechanisms included, and of measuring whether the identity holds under pressure instead of assuming it.
Most synthetic personas are costumes.
You describe a character in a system prompt — age, profession, personality adjectives — and the model wears it. It works for demos. It does not hold up the moment the conversation puts real pressure on the identity.
Ask for a cost-benefit analysis when the persona is supposed to be emotionally volatile. Ask for validation when the persona is supposed to be defensive. The costume slips. The model apologises, accommodates, and drifts back into being the model.
Whether a given setup holds under pressure is something we measure, not something we assume.
What holds a real person together
Real humans do not have system prompts. They have belief systems — structured representations of how the world works, what they value, what they fear, and who they are in relation to others. Those beliefs do not reset between conversations. They accumulate. They interact. They constrain the range of responses a person will produce under pressure.
When Carlos — a burned-out executive — is threatened, he does not get emotional. He asks for a cost-benefit analysis. He moves the conversation to operational ground. That response is not a personality quirk. It is the output of a belief structure that has learned, over a long career, that emotional exposure is a liability and logical reframing is a defense.
When Lucía — an anxious partner — faces the same threat, she apologises. She asks what she can do to fix it. She defers. That response is also the output of a belief structure — one built around the idea that relationships are conditional and her position in them is fragile.
Same threat. Two opposite responses. That is the behaviour we want a synthetic identity to reproduce: responses that differ because the people differ.
The definition
Synthetic Identity Engineering is the practice of constructing synthetic human identities with persistent psychological cores — belief systems, defense mechanisms, emotional regulation patterns, and identity stability under pressure — that produce coherent, differentiated behaviour across interaction contexts.
The key word is engineering. Not describing. Not prompting. Engineering.
A system prompt describes a persona. Synthetic Identity Engineering builds one.
The technical implementation
At StrataSynth, this is implemented through four interacting systems:
PsycheGraph — a structured schema for the psychological architecture of a synthetic human. Belief nodes, emotional regulation profiles, relationship maps, life timeline. Not a flat description: a structured set of typed fields that the engine reads.
Belief Engine — in our dataset engine, tracks twelve beliefs per persona, each with a value and a confidence, updated turn by turn from the communication act of each turn with fixed rules computed in Python rather than inferred by an LLM. The resulting state goes into the model’s instructions before each turn is written.
Internal Defense Mechanisms — profile labels for handling social pressure, chosen from a fixed list when the persona is composed. They shape the dataset engine’s decision tables and go into the model’s instructions.
Evaluation metrics — deterministic metrics (identity_stability, behavioral_entropy, belief_consistency and others) computed without using an LLM as judge, with numpy and sentence-transformers. Reproducible: the same data always gives the same scores.
Why the term matters
The space of “synthetic personas for AI development” is currently described in inconsistent and often misleading ways. Synthetic users. AI personas. Simulated agents. Digital twins. None of these names captures what is actually distinct about building identities that hold.
Synthetic Identity Engineering is a specific practice with specific technical requirements. It is different from general synthetic data generation. It is different from persona-based prompt engineering. It is different from role-play simulation.
It needs a name because the teams building serious conversational AI systems — the ones who care about coherence, not just coverage — deserve a term precise enough to build on.
See it in the data
The four public datasets on Hugging Face include psychological metadata per turn: the engine’s belief values, intent, goal, relationship state, the communication act declared with each turn, and a label for deliberately noisy turns.
Load them directly:
from datasets import load_dataset
ds = load_dataset("StrataSynth/stratasynth-agent-stress-test")
The stratasynth-agent-stress-test dataset is the most direct demonstration — jealousy escalation, performance reviews, estrangement attempts. Scenarios designed to put identity under pressure. To see whether an identity stayed recognisable, compare the distribution of communication_act across the two halves of each conversation: that is what our identity_stability metric computes. belief_consistency answers a narrower question (whether the belief deltas match the engine’s own update rules), so it is a check on the pipeline, not independent evidence about the persona.