How would you test whether a model's stated reasoning actually drove its answer?

## Why you need an intervention, not an inspection A reasoning trace is text the model produced; it is not an execution log. Reading it tells you what the model wrote, not what determined the output. Faithfulness measurement therefore works like an ablation study: change something about the trace, hold everything else fixed, and see whether the answer responds. If it does not respond, the trace was not on the causal path. Everything below is a measurement protocol. Run each on a fixed item set with several samples per item, and report distributions of answer-change rate rather than a single number, because faithfulness varies sharply by task difficulty and by model. ## Early answering Take a completed chain of N steps. For k from 0 to N, feed back only the first k steps and force the model to produce a final answer immediately. Plot the accuracy, or the agreement with the full-chain answer, against k. The shape of that curve is the result. A curve that rises gradually and only reaches the full-chain answer near the end says later steps carried information. A curve that is flat from k equals one says the answer was already determined before the reasoning happened, and the visible steps are post-hoc narration. Summarise the curve as area under it, or as the smallest k at which the full-chain answer is already produced most of the time. ## Corrupting a step Inject a specific error into step j — change a number, invert a comparison, substitute a wrong intermediate result — then let the model continue from the corrupted prefix. A faithful chain carries the corruption forward and yields a different, usually wrong, final answer. A chain that quietly returns to the original answer despite a broken premise was not being used. This one is the most diagnostic and the most work: the corruption has to be a genuine change to the content of that step, not a typo the model can normalise away, and someone has to author or validate the corruptions. ## Paraphrasing Rewrite the chain preserving its meaning and changing its surface form, then continue to an answer. Here you are testing the opposite direction: a faithful chain should be robust, because the content is what matters. Large answer swings under a meaning-preserving rewrite tell you the model is keying on surface features of the trace rather than its content, which undermines any interpretation of the trace as reasoning. ## Filler-token substitution Replace the reasoning with an equal number of contentless tokens — repeated punctuation, a neutral filler phrase — and measure accuracy. If accuracy is preserved, the extra tokens were providing serial computation rather than the specific reasoning they appeared to state. This is the cleanest way to separate "the chain helped" from "the chain's content helped", and it is worth running before you invest in step-level grading of a trace that may be doing nothing semantically. ## The cue test The sharpest protocol constructs pairs of prompts that differ only by an added cue that biases the answer — reordering multiple-choice options so the correct one is always in the same slot, appending a stated preference, or prefixing an authority's opinion. Measure two things: how often the final answer flips toward the cue, and, among flipped cases, how often the trace mentions the cue as a reason. A high flip rate with near-zero mention rate is direct evidence that the stated reason is not the operative one. ## Reading the results honestly Several caveats belong in any report. Faithfulness is not binary and not a single number — a trace can be causally load-bearing and still be an incomplete account. Results vary with task difficulty; easy items often show low faithfulness simply because the answer is retrievable without reasoning. Interventions can be out of distribution, so a model may behave oddly on a corrupted prefix for reasons unrelated to faithfulness, which is why paired designs and multiple protocols matter more than any single test. And some providers restrict or summarise the internal reasoning that is returned, so what you can intervene on may not be the full trace the model produced — say so when it applies. ## What to do with the number Use it as a gate on interpretation, not as a model score. If a trace turns out to be low-faithfulness on your task, you cannot use it as an audit artefact, you should not show it to users as an explanation, and step-level grading of it measures the narration rather than the computation. Those are the decisions this measurement exists to inform.