From Prompt-Level Goals to Prefix-Conditioned Execution: A Position Paper on Incremental Task Execution and Cumulative Audit

@misc{rm3692026prompt,
  author       = {{RM369}},
  title        = {{From Prompt-Level Goals to Prefix-Conditioned Execution: A Position Paper on Incremental Task Execution and Cumulative Audit}},
  year         = {2026},
  month        = oct,
  date         = {2026-10-25},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  url          = {https://lifeofresearch.com/preprints/137},
  abstract     = {This position paper proposes Prefix-Conditioned Execution and Cumulative Audit (PECA), an approach to improving AI instruction following by incrementally processing prompts, executing meaningful microtasks, and repeatedly auditing the accumulated result against the instructions understood so far. Rather than treating a prompt exclusively as a single end goal, the proposed architecture maintains a revisable working state, updates its interpretation as new text arrives, and corrects earlier work when later context changes the meaning of previous instructions. The paper describes the proposed control loop, identifies potential benefits and failure modes, and outlines a controlled evaluation comparing incremental execution with whole-prompt generation followed by final review. The approach is a hypothesis, not a demonstrated performance improvement; empirical testing is needed to determine its effects on instruction coverage, constraint violations, regressions, computational cost, and reward hacking.},
  keywords     = {AI agents, instruction following, reward hacking, process verification, iterative refinement, agent architectures, AI evaluation},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}