Pearl's causal hierarchy separates observational, interventional, and counterfactual questions, but the separation is usually stated qualitatively. Seyed Morteza Emadi asks a quantitative version: once we know the answers to lower-rung queries, how many additional bits are required to specify higher-rung answers? The paper formalizes this as query-class description length, using Kolmogorov complexity of the answers a structural causal model gives to a class of queries, with system size fixed.

The main construction produces acyclic structural causal models on n binary variables whose common observational distribution has constant description length, yet single-variable intervention answers can require Θ(n²) additional bits. For models with independent noises, fixed finite gate-schema and noise libraries, and at most d parents per variable, an upper bound of O(nd log(en/d) + n log n) shows the quadratic construction is order-optimal in the dense regime; a rooted-tree construction achieves order-optimality for bounded indegree.

The quadratic separation persists even when descriptions are only required to be ε-accurate in total variation for every fixed ε < 1/4. Moving to the next rung, knowing every hard-intervention distribution can still leave a Θ(n) counterfactual description gap, and a masked-parity construction pushes this to Θ(n²). The paper proves an ambiguity-to-bits theorem and a Shannon analogue, showing that in these constructions the gaps equal the logarithm of residual higher-rung ambiguity up to lower-order terms.