From Draft Selection to Evidence-Preserving Revision

Which intermediate drafts should be retained for audit and learning? This review answers by treating trajectory governance as a property of a sociotechnical workflow rather than a feature that can be read from average accuracy. The focal setting is multi-agent language reasoning, where speed and fluency can conceal semantic loss, correlated self-evaluation errors, or domain-specific failure. Evidence from the assigned publications is synthesized with foundational studies of calibration, distribution shift, causal structure, and responsible deployment. Four requirements follow: preserve the lineage of source texts, alternative drafts, reward signals, confidence estimates, domain corpora, and human judgments; measure stability across relevant perturbations; connect confidence to a specific action; and maintain a route for human challenge and correction. The framework distinguishes descriptive performance from decision utility and separates uncertainty about the world from uncertainty created by the model and its evaluator. It also shows why faster inference or richer reasoning is valuable only when it improves a defined decision under a transparent resource budget. The article is a literature review and research agenda, not a report of a newly completed trial.

1. Zhang, Yin, et al. "SAINF: Intrinsic Self-Correction for Robust Machine Translation with Large Language Models." *Frontiers of Computer Science* (2026).
2. Chen, Yiwei, et al. "One-Step Generative Distillation." *ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)* (2026).
3. Rafailov, Rafael, et al. "Direct Preference Optimization: Your Language Model Is Secretly a Reward Model." *Advances in Neural Information Processing Systems*, vol. 36, 2023.
4. Sutton, Richard S., and Andrew G. Barto. *Reinforcement Learning: An Introduction*. 2nd ed., MIT Press, 2018.
5. Tishby, Naftali, Fernando C. Pereira, and William Bialek. "The Information Bottleneck Method." *Proceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing*, 1999, pp. 368-377.
6. Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. "Distilling the Knowledge in a Neural Network." *NIPS Deep Learning and Representation Learning Workshop*, 2015.
7. Guo, Chuan, et al. "On Calibration of Modern Neural Networks." *Proceedings of the 34th International Conference on Machine Learning*, 2017, pp. 1321-1330.
8. Geifman, Yonatan, and Ran El-Yaniv. "Selective Classification for Deep Neural Networks." *Advances in Neural Information Processing Systems*, vol. 30, 2017.
9. Araci, Dogu. "FinBERT: Financial Sentiment Analysis with Pre-Trained Language Models." *arXiv preprint arXiv:1908.10063*, 2019.
10. Loughran, Tim, and Bill McDonald. "When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks." *Journal of Finance*, vol. 66, no. 1, 2011, pp. 35-65.
11. Gururangan, Suchin, et al. "Don't Stop Pretraining: Adapt Language Models to Domains and Tasks." *Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics*, 2020, pp. 8342-8360.
12. Vaswani, Ashish, et al. "Attention Is All You Need." *Advances in Neural Information Processing Systems*, vol. 30, 2017.