Retrieval and synthesis designed to keep claims separable from conclusions.
Claims before conclusions
A polished answer can conceal a weak chain of support. Evidence-first synthesis reverses the order: identify atomic claims, attach the strongest available support, record conflicts, and only then write a conclusion.
This produces a claim graph rather than a block of prose. The graph can be inspected, re-ranked, or regenerated without discarding the entire answer.
Source quality is contextual
Authority depends on the question. A standards body may be decisive for a protocol, while a primary benchmark repository may be more useful for a model comparison. Horizon records why a source is relevant instead of assigning one universal trust score.
Recency, independence, directness, and reproducibility are evaluated separately. A source can be current but indirect, or primary but incomplete. Those differences should survive into the final artifact.
Calibrated output
When evidence is thin, the system should narrow the claim. When sources disagree, the disagreement is itself a finding. Confidence is not a decorative percentage; it is a statement about what would need to change for the conclusion to change.
North evaluates evidence handling independently from prose quality so an elegant answer cannot receive credit for unsupported certainty.