Conversational Path Dependence in Ambiguous LLM Judgment

Large language model decision support unfolds across turns, so a final context may contain both externally supplied facts and the assistant's earlier interpretations. Prior work shows that multi- turn presentation, prior model outputs, failed attempts, and visible initial choices can alter later behavior; conversational path dependence itself is not claimed as new. We present a controlled three-phase diagnostic decomposition of same-fact conversational path dependence in ambiguous LLM judgment. Across synthetic interpersonal cases, Phase 1A held the final external record constant while varying fact order and prior assistant commitment. Its largest matched effect occurred in one apparently ambiguous case: relative to the order-matched control, commit-to-A shifted the final category from B to genuine indeterminacy. Phase 1B crossed three ex ante evidence-diagnosticity levels with two intended polarities and estimated prior- commitment, order, and salience sensitivity separately. In the tested case family, pooled absolute commitment sensitivity was 0.0280, 0.0335, and 0.1150 at low, medium, and high ambiguity; order and salience did not follow the same ordering. The largest effect was a High-A commit-to-B contrast of -0.342 in directional belief, but it was strongly asymmetric, and High-A failed its neutral-condition polarity check. Phase 1C tested the stronger hypothesis of recursive self- reinforcement over three no-new-evidence continuation turns. No-new-evidence and self- summary trajectories showed zero mean directional drift, zero mean confidence growth, and no conclusion or recommendation changes. Fresh evidence produced systematic movement, confirming that the pipeline could register longitudinal updating when the external record changed. Together, the results show immediate prior-commitment sensitivity under some tested conditions but no recursive amplification under the Phase 1C protocol.