Most "AI agents" are a language model behind a persona prompt. They are fluent, cheap, and unfalsifiable — you cannot say how close their behaviour is to any real population, because nothing was measured. Anity builds the other kind: digital avatars derived from consented cohort data, calibrated against held-out observations, and versioned so that any result can be reproduced exactly.
A generic agent role-plays a demographic description. An Anity agent is conditioned on a consented corpus from a real cohort — longitudinal survey waves, interview transcripts, diary entries, sensor logs — with identifiers stripped at the ingestion boundary. Its responses trace back to observed evidence rather than to the model's prior about who that person should be.
Before an agent is used in a study, it is scored against responses that were withheld from fitting. We report alignment deltas per construct, so you can see which measures the agent reproduces within tolerance and which it does not. An agent that fails calibration on a construct is not silently used for it.
Every run records its model version, data slice, protocol version, and random seed. Re-executing the same run identifier reproduces the same transcript. Results you publish can be re-derived by a reviewer from the exported bundle rather than re-generated approximately.
Agents are instruments with a domain of validity. Each release ships with the constructs it was calibrated on, the population it was derived from, and the conditions under which its output should not be treated as evidence. Extrapolation beyond that envelope is a documented limitation, not a hidden one.
Test question wording, ordering effects, and survey length across thousands of agent participants before recruiting a single human, then take a shorter, cleaner instrument to the field.
Run counterfactual conditions and simulate effect sizes on a cohort-derived population to size a study realistically instead of guessing from published averages.
Conduct exploratory interviews with agents to surface themes, probe follow-up structures, and train interviewers, keeping human sessions for confirmation rather than discovery.
Place calibrated agents in multi-agent environments, markets, or interaction protocols where the research question is about behaviour under conditions that are impractical to stage with people.
A calibrated agent is not a replacement for human participants, and we do not present it as one. It is a pre-registration and design instrument: it narrows the space of studies worth running, and it makes the assumptions in a design explicit and testable. Confirmatory claims about people still require people. The full set of standards we hold ourselves to — provenance, falsifiability, validity boundaries, consent governance — is published in the research manifesto.
