Definition
Plain language
A group of cause-and-effect diagrams that all produce the exact same pattern of correlations, so you can't tell them apart by watching.
As stated in the literature
The set of causal graphs that imply identical conditional-independence relations; indistinguishable from observational data alone, requiring interventions to disambiguate.
Why it matters: It marks the limit of what observation can reveal about causation, showing why you sometimes have to intervene to learn the real direction.
For example, 'rain causes wet grass' and 'wet grass causes rain' can produce the same correlations, so watching alone can't tell which is true.
Heard on the show
“The paragraph has enough information, in principle, to nail down what's called the Markov equivalence class — the set of graphs that all produce these correlations.”Episode 091 — When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning