VCT: A Verifiable Transcript System for Large Language Model Conversations

Large Language Model (LLM) interaction records are increasingly being used as digital evidence for forensics and compliance. However, traditional linear tamper-evident logs cannot capture the complex nonlinear conversation evolution, such as prompt-editing, response regeneration, session deletion, multi-device concurrency, and selective sharing. This study proposes the Verifiable Conversation Transcript (VCT), which models these operations as authenticated account-level state transitions across a three-layer cryptographic structure: Q&A pairs form branch hash chains and branch tails aggregate into session-level Merkle roots (MTRcon). These session-level roots are further aggregated into the account-level Merkle root (MTRacc), which is jointly authenticated by both user and server signatures. The VCT also introduces a serialized deletion protocol, a deterministic merge protocol that preserves concurrent non-deletion updates, a gossip-based fork detection mechanism that produces non-repudiable forensic evidence, and a verifiable sharing protocol that constructs an independent dual-signed share hash chain across disclosed nodes. Based on standard cryptographic assumptions, the VCT guarantees integrity, consistency, share verifiability, and non-repudiation of account-level conversation records. A Python prototype achieves sub-millisecond mean latency under minimal state with short synthetic payloads. With 21 KB per-node payloads, the estimated metadata-to-content ratio is approximately 0.9% across the tested account scales. These results support the computational feasibility of the prototype under the tested conditions and provide practical guidance for the engineering design and optimization of verifiable LLM conversation systems.