Methodology
Predictions are generated before the resolution date using ARMIS causal graph and signal aggregation.
No post-hoc modification. All timestamps are UTC and immutable once committed to git.
Accuracy measured against ground truth at resolution. Brier score measures calibration (0 = perfect, 1 = worst).
Pending predictions remain open until their target date passes and a resolution is recorded.