Digital avatars lower the marginal cost of generating human-shaped responses. That efficiency creates a methodological obligation: to distinguish synthetic elicitation from observational inference, and to treat avatars as calibrated instruments whose outputs are only valid within specified operating conditions. This document defines the engineering, evaluation, and governance standards that govern how we build, validate, and release digital subjects for research.
We treat each avatar as a conditional distribution over behavioral responses, conditioned on a documented individual- or cohort-level signal. It is an abstraction layer — a computational proxy for elicited behavior under specified task conditions. Avatars are characterized before deployment: input coverage, response space, population strata, and known representational limits are published alongside any result.
A deployed avatar makes an explicit, testable claim: given a defined population, task, and protocol version, generated responses will fall within a pre-registered tolerance of held-out observations from the same stratum. We construct the evaluation protocol — metrics, hold-out set, and acceptance criteria — before inspecting model outputs. Claims that cannot be operationalized as a comparison task are not claims we make.
Each avatar's inputs carry lineage metadata: data source, collection mechanism, temporal scope, consent boundary, preprocessing steps, and any synthetic augmentation. Inputs without recoverable provenance are excluded from production builds. Exports include the same lineage envelope, enabling independent reconstruction of the modeling chain by downstream reviewers.
Model weights, inference seeds, protocol versions, and runtime dependencies are pinned. A study executed at one date must reproduce identically on the same configuration; deviations are treated as reportable events. We do not rely on nondeterministic interfaces or undocumented hyperparameters when reporting research outcomes. Demonstrations that cannot be re-run are excluded from evidentiary use.
Fluent or plausible-looking output is insufficient. We optimize for distributional fidelity: marginal and conditional response statistics, heterogeneity across subgroups, non-response and refusal patterns, and temporal drift. Calibration is evaluated against held-out empirical quantiles, not against researcher intuition or narrative coherence.
Every avatar ships with a documented failure-mode taxonomy: under-represented populations, question types that induce systematic distortion, known coverage error, and observed drift over time. We report these limits first. Silent degradation — output that appears stable while validity decays — is classified as a critical defect and triggers build retirement.
The individuals whose signals ground an avatar can inspect its scope, impose constraints, and revoke authorization. Revocation propagates through the system: the avatar is retired, derived artifacts are flagged, and dependent studies are notified. Consent is managed as a stateful relationship with auditable transitions, not as a static authorization record.
We do not construct avatars from individuals who have not provided informed, revocable consent for the specified research purpose. We do not deploy digital subjects in contexts where human participants are feasible and ethically required. We do not represent synthetic responses as observational data, and we do not release a cohort whose calibration has not been evaluated against held-out empirical observations.
Used within these constraints, digital avatars serve as pre-registration and design-validation instruments: they enable stress-testing of survey instruments, estimation of statistical power, and screening of study designs before the expenditure of finite human attention. Beyond these uses, they risk manufacturing confidence from unvalidated generative artifacts. We decline such uses rather than dilute the standard.
