<result> <column_name>AVERAGE TIME ON PAGE (SECONDS)</column_name> <family>additive</family> <subtype>intensive_additive</subtype> <confidence>0.78</confidence> <requires_user_confirmation>false</requires_user_confirmation> <sum_meaningful>false</sum_meaningful> <signed>false</signed> <reducers> <primary_reducer>mean</primary_reducer> <valid_reducers>mean, min, max, median, distribution</valid_reducers> <excluded_reducers>sum (summing per-entity averages gives a meaningless level), product (not a ratio)</excluded_reducers> <average_semantics>arithmetic_mean</average_semantics> </reducers> <inner_outer_measure> <inner_measure_reducer>min</inner_measure_reducer> <outer_measure_reducer>max</outer_measure_reducer> </inner_outer_measure> <multiplicative_partner_derivable>false</multiplicative_partner_derivable> <reasoning> The table is STATE with no sequence axis, so multiplicative is implausible and there is no cumsum or cumprod reconstruction to test. The column is a per-entity average duration in seconds. It has no ratio naming, the mean (55) and median (20) are far from 1.0, and it contains zeros and large values up to 2794, so it is not a ratio. "Duration" would normally suggest real_additive, but the "AVERAGE" qualifier makes it an intensive level: summing averages is not meaningful, and a mean is the natural rollup. The high log_symmetry_gain only reflects right skew in a positive duration distribution and does not outweigh the other evidence. Some uncertainty remains because a weighted mean (weighted by page views) would be more exact than a plain mean. </reasoning> <table_type_conflict>false</table_type_conflict> </result>