Conversational repair, the process by which interlocutors detect and correct misunderstandings, is central to human language use but rarely studied in the internal representations of large language models (LLMs). We ask: when a model commits to an incorrect interpretation of an ambiguous utterance and the user corrects it, where and how does the corrected interpretation emerge across layers? We construct a controlled diagnostic dataset of 291 three-turn dialogues, pairing each repair run (an ambiguous user turn, a model commitment to a default reading, and a user correction) with a surface-matched control that performs no repair, isolating the correction’s effect. Across six instruction-tuned models from three families, we find that (i) repair revises meaning in a narrow band of mid-to-late layers, where activation patching causally shifts output to the corrected reading; (ii) when repair fails behaviourally, the corrected reading is causally absent rather than weak; (iii) attention, not the feed-forward sublayer, carries the revision; and (iv) revision is stronger when the misreading is attributed to the assistant rather than the user, indicating more than generic in-context override. These results characterize conversational repair as a localisable, model-dependent computation and provide a reusable diagnostic for meaning revision in interaction.

Multi-agent LLM coordination typically reduces to voting or arbitration over self-reported labels and reasoning that is uncoupled from agent behavior and fails predictably under structured bias. A shortcut-reliant agent is confident about its shortcut, and its predictions survive any rule that consults only self-reports. We propose VOUCH, a coordination protocol in which every agent claim is coupled to a deterministic test the protocol runs itself. An agent commits to a counterfactual prediction about an intervention’s outcome; the protocol applies the intervention and verifies the contract only if the actual outcome matches. The natural single-side instantiation is unsound as a self-coherent agent satisfies its own commitment by construction. We resolve this with a dual-side construction where the proposer additionally commits to a second agent’s post-intervention prediction, a cross-side claim that self-coherence cannot fulfill. Across multimodal classification (Hateful Memes, CrisisMMD, MM-IMDb), code (MBPP, HumanEval), and math (GSM8K, MATH, SVAMP), the dual-side fix reduces wrong-verified events by 46%–85% per domain while never falling more than 1 point below the Ensemble baseline on accuracy. VOUCH is therefore designed not to outperform ensembling, but to preserve task accuracy while making coordination decisions testable and auditable.