Part 1 of 4 · 1 min read
Overview
An agent that quotes the wrong refund window or invents a policy is worse than one that says nothing. Interviewers ask how you detect and reduce hallucination; the strong answer separates the two detectors, names what each misses, and moves to design-level prevention.
About the example, and what is real
Real: the claim splitter, the overlap, number and negation checks, the evaluation on a fixed claim set, and the consistency experiment in code/mini-faithful, with 12 tests. Assumed: the claim set (22 items written by me, in six kinds), the context (six policy sentences), and the self-consistency model (a model "knows" 70 percent of facts and otherwise guesses uniformly among six wrong answers). Not real: no model answers anything. The faithfulness scores come from rules I tuned on these same claims, so they do not predict performance on real text.