Governance¶
gp3bayespy is intentionally conservative about what software output can establish. The package separates statistical evidence from analyst interpretation and keeps automatic decision-making disabled where the frozen gp3bayes contract requires that boundary.
Package-wide rules¶
-
No automatic model selection
Predictive scores, ELPD differences, and model weights are evidence. The package does not automatically declare a preferred model.
-
No automatic exclusion
Readiness, influence, missingness, and measurement audits do not automatically remove participants, groups, trials, or observations.
-
No automatic adequacy claim
Passing diagnostics does not itself establish that a model is substantively adequate.
-
No automatic causal claim
Association estimates are not described as causal effects unless the study design and target estimand justify that interpretation.
-
No latent-state inference
Behavioural, gaze, pupil, or physiological measurements are not automatically mapped to cognition, emotion, stress, comprehension, intention, personality, diagnosis, or deception.
-
Explicit unsupported uses
Restricted model families and workflow contracts expose unsupported uses rather than silently widening the analysis scope.
What diagnostics mean¶
A diagnostic result such as pass, review, or fail describes evidence under a declared rule. It does not substitute for a scientific conclusion.
The same principle applies to:
- convergence diagnostics;
- posterior predictive checks;
- calibration;
- PSIS-LOO and Pareto-k;
- sensitivity analyses;
- recovery and SBC;
- pupil temporal audits;
- backend parity checks.
Reproducibility¶
Analysis manifests, frozen specifications, transformation replay, model cards, and publication bundles are designed to make decisions inspectable. They do not certify that those decisions are scientifically correct.
Reference¶
These boundaries are inherited from the frozen gp3bayes 0.5.0 reference and are part of Python parity rather than optional documentation language.