MacroGuru now freezes short source spans, verifies entity and aspect relevance, sorts temporal direction, classifies stance from registered rules, collapses common origins, and attaches every public evidence claim to an exact span. A publisher date never substitutes for the timestamp at which MacroGuru actually froze the evidence.
3 verified retrospective spans; original-decision independent sources: 0.
3 verified retrospective spans; original-decision independent sources: 0.
2 verified retrospective spans; original-decision independent sources: 0.
3 verified retrospective spans; original-decision independent sources: 0.
2 verified retrospective spans; original-decision independent sources: 0.
2 verified retrospective spans; original-decision independent sources: 0.
2 verified retrospective spans; original-decision independent sources: 0.
Register → retrieve → verify → sort → classify → cluster → cite → replay. Each forecast registers entities, aspects, directional phrases, a relevance threshold, and a minimum independent-source count before evaluation. Retrieval uses concept groups plus token overlap; it is deliberately labeled registered-concept retrieval, not a neural semantic model.
Entity verification can use the frozen document title and publisher context. Aspect verification must appear in the passage itself. Host, common-origin, and syndication IDs are separate, and effective breadth counts common origins only once.
The published seed metrics are a contract fixture, not proof of forecasting lift: This small, separately stored seed set tests contracts and hostile failure modes. It is not evidence of out-of-sample forecasting lift.
Collection hash: 0cea10c9c10e152556fd3f398d513189914d65feb7b1808a065073bfbb5fffde
Evaluation hash: 80c155e2f846f3375594d434666365de0796ff385acc747aaa78c45e0bfd34bf
Open the collection JSON → · Open the evaluation JSON → · Return to Decision Receipts →