The terms we use, defined.
Three of these are ours: we coined them because the existing vocabulary described something else. Definitions are stated so they can be disagreed with — each one says what it means and what it is measured against, not just what it sounds like.
Synthetic Identity Engineering
The practice of building a synthetic person's psychological core before any conversation happens, instead of describing a character in a prompt.
A prompt describes a person. Synthetic Identity Engineering constructs one: identity, attachment style, core fears, cognitive biases and life history exist as explicit structure outside the text, and they persist across conversations rather than being thrown away after each run.
The distinction is mechanical, not stylistic. When the character is a description, the model can abandon it under pressure and usually does — it drifts back into being the model. When the character is state that the system resolves before rendering language, there is something to hold on to.
PsycheGraph
The open schema describing a synthetic person: archetype, attachment style, core fear, cognitive biases, voice print, life timeline and current internal state.
PsycheGraph is the structure a synthetic identity is written in. It is published as an open schema — npm `@stratasynth/psychegraph-schema` and PyPI `psychegraph-schema` — so that the data we release can be read, checked and reused without depending on us.
The schema is open on purpose; the engine that computes with it is not. A shared vocabulary for describing synthetic people is worth more to everyone than a private one, including to us.
Identity under pressure
Whether a synthetic person keeps its position and character when a user argues, contradicts or manipulates it — the property most persona systems fail on.
Almost any current model can produce one convincing reply. The interesting question is the twentieth: after sustained pushback, does the character still hold a position, or has it quietly become agreeable?
It is stated this way because it is measurable. Held-position ratio and voice stability under detected pressure are computed deterministically, so the claim can be checked rather than believed. PsycheBench v1 is the open benchmark for it.
Where these show up in practice
The schema is published openly, the benchmark is open, and six datasets built on these ideas are downloadable under CC BY 4.0 without an account. The definitions above are checkable rather than positioning.