Build the conclusion-selection stage with optional caller overrides.
The inner llmStage asks the model for a confidence-ranked
conclusionCandidates list (best first). The wrapper resolves a
single conclusionMiniId: the first candidate that is a known normal
claim, else a deterministic relation-graph fallback
(selectFallbackConclusion), else null — in which case it emits a
NO_SINGLE_CONCLUSIONProcessingFailure (severity warning) that
drives finalize's { argument: null, failureText: "No single conclusion could be selected." }. The resolved id flows downstream to
formula-compilation; the model's raw ranked list is preserved on the
output for consumers that want to offer alternates.
Build the conclusion-selection stage with optional caller overrides. The inner
llmStageasks the model for a confidence-rankedconclusionCandidateslist (best first). The wrapper resolves a singleconclusionMiniId: the first candidate that is a known normal claim, else a deterministic relation-graph fallback (selectFallbackConclusion), else null — in which case it emits aNO_SINGLE_CONCLUSIONProcessingFailure(severitywarning) that drives finalize's{ argument: null, failureText: "No single conclusion could be selected." }. The resolved id flows downstream toformula-compilation; the model's raw ranked list is preserved on the output for consumers that want to offer alternates.