Summing Evidence Weights Shifts Your Decision Threshold

A paper names a bug baked into every score-summing detector: the threshold slides as you add sources, per arXiv 2608.14509.

Any system that scores risk by summing weighted signals and checking the total against a cutoff carries the same flaw. The drift grows as source reliability varies, and no fixed threshold works across every mix of sources.

The paper separates two jobs most systems fuse: reading a single source, and combining several readings into one call. It proposes a four-field evidence format as the boundary between those steps.

When sources carry different reliability, a majority-vote rule and a posterior-probability rule rank the same evidence in different orders. No single threshold reconciles the two, the paper finds.

The fix doesn't need a rebuild. Pool calibrated log-likelihood ratios instead of summing raw weights, and the threshold stops sliding as sources get added. The same weakness sits inside score-summing triage engines and additive multi-signal detectors.

If your scoring system sums weights and thresholds the total, look at when that threshold was tuned. It was probably tuned for fewer sources than you're running now.