A model writes its reasoning, then gives an answer. Does the reasoning actually cause the answer, or does the answer arrive by a shortcut while the reasoning is just for show? Move the three controls and watch where the answer comes from.
How much the written reasoning actually drives the final answer. High means the words on the page are doing the work.
How much the prompt pushes the answer directly, skipping the reasoning. High means the model decides first and explains later.
How strongly the prompt changes what the reasoning says in the first place.
Three plain ideas, and the formal names underneath in case you want them.
Particles that travel up through the reasoning node and down to the answer. The thicker and busier this path, the more the model's stated reasoning is genuinely causing its answer.
prompt → reasoning → answerParticles that skip straight across the bottom. This is the answer arriving without the reasoning, a sign the written steps are decorative rather than load-bearing.
prompt → answer (bypasses reasoning)Faithfulness is the share of the answer's movement that went through the reasoning. The real method puts a calibrated uncertainty band on this number and asks how much hidden confounding it would take to overturn it.
NIE / total effect · see the methodology →