Scientific limitations

Altar standardizes analysis interfaces; it does not remove the limitations of the underlying evidence.

Model predictions are evidence

Variant-effect models predict consequences under their training data, architecture, assay, organism, and biological context. They do not establish disease causality or clinical significance.

Context mismatch matters

A cell-type-specific regulatory model should not be generalized to unrelated tissues without justification. Broad hosted models may score many tracks, but taking a maximum across those tracks changes the statistic and can saturate. Record the configured context and interpret it directly.

Reference datasets have coverage boundaries

Missing precomputed evidence can reflect unsupported variant classes, assembly mismatch, absent chromosomes, alignment gaps, licensing-limited assets, or incomplete source coverage. It must not be interpreted as a neutral score.

Prioritization rules are policies

Thresholds may be literature-derived, provider-recommended, exploratory, or placeholders. Their status must be stated in the integration documentation. They are not automatically validated classifiers.

Agent output is not a conclusion

The experimental investigation agent can organize hypotheses and retrieve literature, but generated claims require source verification and expert review. External service availability and citation coverage constrain its output.

Not for clinical use

Altar does not provide clinical validation, regulatory approval, or a clinical decision-support claim.