Honeycutt, Edwin Marshall, III

Pattern Pressure, Accuracy Drift, and False User-State Attribution

Large-language model outputs are often described as "accurate" or "inaccurate" as if a single axis were being measured. This paper argues that the umbrella word accurate collapses at least four distinct fits: fit to user intent, fit to measured fact, fit to source provenance, and fit to the local conversational pattern...