Collapsed Model Disagreement
Pre-aggregated ensemble scores hide which models agreed, disagreed, and why
Disagreement between models is epistemic signal — collapsing it before human review destroys information the reviewer needs to escalate.
Two models scoring 0.91 and 0.73 on the same case are not the same as two models both near 0.82 after averaging — yet a pre-aggregated ensemble score of 0.82 looks identical in both situations. Collapsed model disagreement is the failure mode where individual model outputs, confidences, and disagreement flags are merged into a single number (or single recommendation) before the human at the commitment boundary sees them.
Strengthening the Commitment Boundary Mechanism 4: disagreement is information about case difficulty and ambiguity — exactly what a human reviewer needs to know that escalation is warranted.
In software systems
The fraud API returns { score: 0.82, modelVersion: "ensemble-v2" }. Dashboards show one gauge. The commit event stores one field. Operations tunes one threshold. Post-incident review cannot reconstruct which sub-model drove the bind or whether models conflicted on edge features.
Counter direction
Disclosed model disagreement — each model’s output and confidence in the record; disagreementFlag and escalation when spread exceeds policy. Same pattern as showing multiple independent LLM outputs separately in LLMs at the Commitment Boundary.
Related: compliance theatre when “multi-model” is claimed but only ensemble output is stored.