Token-Generation-Control-Deterministic-Logical-Timeline

Abstract

Text generation by large language models is essentially a probabilistic sampling process: the model computes a probability distribution over the next token based on the input context, and then samples a token from that distribution. This probabilistic design gives large models creativity but also introduces a fundamental engineering problem: the same input may produce different outputs at different times, in different environments, or on different hardware. For AI agents (automated trading, code generation, industrial control) that require stability and reproducibility, this non-determinism is fatal. This paper proposes a token generation control method based on a deterministic logical timeline. The method uses the logical state S(n) output by a Protocol Security State Machine (PSSM) as the initialization seed of a pseudo-random number generator (PRNG), replacing random sources based on physical time or hardware noise, thereby transforming the token sampling process from "physical non-determinism" to "logical determinism". The paper provides the complete token generation flow, two working modes (deterministic mode and divergent-but-reproducible mode), a cryptographically bound record format, and, building on the independent verification and replay mechanism disclosed in the applicant's prior patent, an auditing method for token generation records. The method does not change the model's native probability distribution (intelligence representation); it only controls the token sampling path, achieving cryptographic-level determinism across environments, hardware, and frameworks.