Pick a conclusion deterministically from the relation graph when the
model abstains. A candidate is a normal-typed claim that is the
target of at least one support relation. Pure sinks (a target that
supports nothing else) are preferred — the textbook "therefore X"
terminus; when none exists (e.g. a support cycle) any supported normal
claim is eligible. The winner has the highest in-degree, ties broken
by document order (earliest appearance in the claim-type list) so the
pick is stable across runs. Returns null when no candidate exists —
there is no argument to draw a conclusion from.
Exported so a cheaper ingestion pipeline that produces conclusion
candidates in one combined LLM call can reuse the identical
resolution when adapting its output into the conclusion-selection
slot — keeping the resolved conclusionMiniId contract uniform
across pipelines.
Pick a conclusion deterministically from the relation graph when the model abstains. A candidate is a
normal-typed claim that is the target of at least one support relation. Pure sinks (a target that supports nothing else) are preferred — the textbook "therefore X" terminus; when none exists (e.g. a support cycle) any supported normal claim is eligible. The winner has the highest in-degree, ties broken by document order (earliest appearance in the claim-type list) so the pick is stable across runs. Returns null when no candidate exists — there is no argument to draw a conclusion from.Exported so a cheaper ingestion pipeline that produces conclusion candidates in one combined LLM call can reuse the identical resolution when adapting its output into the conclusion-selection slot — keeping the resolved
conclusionMiniIdcontract uniform across pipelines.