Bibliography

Every citation was checked against its DOI record on 14 September 2026. Entries marked VERIFIED were read at the source and carry a quotation.

1. VERIFIEDREVIEW RATE

   Kemerer, C. F., & Paulk, M. C. (2009). The impact of design and code reviews on software quality: An empirical study based on PSP data. IEEE Transactions on Software Engineering, 35(4), 534–550.

   doi:10.1109/TSE.2009.27source
   TL;DR

   Two data sets, 371 and 246 programs. Review rate predicted defect removal after controlling for developer ability. At 200 lines an hour or slower, reviews caught nearly two-thirds of design defects and over half of code defects.

   “The recommended review rate of 200 LOC/hour or less was found to be an effective rate for individual reviews, identifying nearly two-thirds of the defects in design reviews and more than half of the defects in code reviews.”

   D1 · The governed unit is the change, not the documentD4 · Mandates carry numeric limits
2. VERIFIEDATTENTION

   Pattyn, N., Neyt, X., Henderickx, D., & Soetens, E. (2008). Psychophysiological investigation of vigilance decrement: Boredom or cognitive fatigue? Physiology & Behavior, 93(1–2), 369–378.

   doi:10.1016/j.physbeh.2007.09.016source
   TL;DR

   Reaction times slowed after 30 minutes on task, with a further cost to self-directed (endogenous) attention after 60 minutes. Physiological measures pointed to underload, not exhaustion.

   “The vigilance decrement has been described as a slowing in reaction times or an increase in error rates as an effect of time-on-task during tedious monitoring tasks.”

   D15 · The review clock is a feature
3. VERIFIEDATTENTION

   McCarley, J. S., Gyles, S. P., Hankey, K. R., Muñoz Gomez Andrade, F., & Yamani, Y. (2026). Deconstructing the vigilance decrement: Changes in bias, lapse rate, and guess rate, but not sensitivity. Attention, Perception, & Psychophysics, 88, 140.

   doi:10.3758/s13414-026-03278-9source
   TL;DR

   366 participants, 40 trials a minute in noise. Over time, people shifted to a conservative criterion, lapsed more, and guessed positive less. They did not lose the ability to tell signal from noise. A tired reviewer sees as well and reports less.

   “Model comparisons revealed decisive evidence for conservative bias shifts, increases in mental lapse rates, and decreases in positive guess rate over time, but gave evidence against a loss of sensitivity.”

   D7 · Approve, deny, regenerate, edit; no silent editsD15 · The review clock is a feature
4. VERIFIEDAI-WRITTEN TEXT

   Zhang, Y., Das, S. S. S., & Zhang, R. (2024). Verbosity ≠ veracity: Demystify verbosity compensation behavior of large language models. arXiv.

   doi:10.48550/arXiv.2411.07858source
   TL;DR

   Five QA datasets, 14 models. When uncertain, models repeat the question, hedge, and over-enumerate; the behavior appeared in every model and dataset. GPT-4 showed the behavior 50.40% of the time, and the gap between verbose and concise answers held as model capability rose.

   “When unsure about an answer, humans often respond with more words than necessary, hoping that part of the response will be correct. We observe a similar behavior in large language models.”

   D5 · Length is a drift signal
5. VERIFIEDAI-WRITTEN TEXT

   Zhu, J., Peng, H., Wang, J., Ke, L., Zhang, C., & Zhang, L. (2026). Large language models do not always need readable language. arXiv.

   doi:10.48550/arXiv.2606.19857source
   TL;DR

   Text written for a model reader kept 99.5% of its meaning at 27.9% of its length. Human readability and model recoverability are separate properties. One document cannot serve both readers well.

   “maintaining 99.5% semantic fidelity even when the text volume is condensed to 27.9% of its original length.”

   D2 · One source, two renderings
6. VERIFIEDAGENT OUTPUT

   Duma, K., Wróblewski, P., Bobińska, J., Winiarska, J., & Przymus, P. (2026). These aren't the reviews you're looking for: How humans review AI-generated pull requests. arXiv. EASE 2026.

   doi:10.48550/arXiv.2605.02273source
   TL;DR

   AIDev dataset, 33,596 agent-authored pull requests. Most received no review; when reviewed, other agents did most of it. In repositories with both kinds of PR, 69.9% of agent-authored PRs had no recorded review or agent-only review, and human-only review was rare (8.08%).

   “the absence of recorded review activity does not imply the absence of human oversight; maintainers may inspect pull requests without leaving traceable comments.”

   P · Problem statement
7. VERIFIEDOVERSIGHT

   Turan, E. (2026). Oversight has a capacity: Calibrating agent guards to a subjective, fatiguing human. arXiv.

   doi:10.48550/arXiv.2606.08919source
   TL;DR

   125 hand-labeled agent actions. Reviewers agreed only moderately on what was risky (Fleiss' κ = 0.52). With the reviewer modeled as fatiguing, realized safety is an inverted U in escalation rate: past the optimum, more approvals make a system less safe. The author states the fatigue curve is simulated, not measured.

   “Agent oversight, framed this way, is not only a classification problem but a resource-allocation one: human attention is finite, and the guard's escalation policy spends it.”

   D8 · Approvals are rationed and de-duplicatedD11 · Measured churn, never a self-rated risk score
8. DOI CHECKEDREVIEW RATE

   Vitharana, P. (2017). Defect propagation at the project-level: Results and a post-hoc analysis on inspection efficiency. Empirical Software Engineering.

   doi:10.1007/s10664-015-9415-3
   TL;DR

   Inspection efficiency rises to an optimum with time invested, then falls. Defects missed at the requirements stage propagate into design and code.

   D1 · The governed unit is the change, not the document
9. DOI CHECKEDREVIEW RATE

   Laitenberger, O. (2001). Cost-effective detection of software defects through perspective-based inspections. Empirical Software Engineering.

   doi:10.1023/a:1009805707387
   TL;DR

   Inspection cost-effectiveness falls on large documents, attributed to fatigue and lost motivation. Inspect by logical entity, not by document.

   D1 · The governed unit is the change, not the document
10. DOI CHECKEDATTENTION

    Reteig, L. C., van den Brink, R. L., Prinssen, S., Cohen, M. X., & Slagter, H. A. (2018). Sustaining attention for a prolonged period of time increases temporal variability in cortical responses. bioRxiv.

    doi:10.1101/501544
    TL;DR

    An 80-minute sustained attention task. Performance and motivation fell fast and reached a floor well before 60 minutes. A monetary incentive at 60 minutes restored motivation but not performance.

    D15 · The review clock is a feature
11. DOI CHECKEDATTENTION

    Robison, M. K., & Nguyen, B. (2023). Competition and reward structures nearly eliminate time-on-task performance decrements: Implications for theories of vigilance and mental effort. Journal of Experimental Psychology: Human Perception and Performance, 49(9), 1256–1270.

    doi:10.1037/xhp0001148
    TL;DR

    Competition and reward nearly eliminated time-on-task decrements, which argues the decrement is motivational rather than a depleted resource.

    D15 · The review clock is a feature
12. DOI CHECKEDAI-WRITTEN TEXT

    Saito, K., Wachi, A., Wataoka, K., & Akimoto, Y. (2023). Verbosity bias in preference labeling by large language models. arXiv.

    doi:10.48550/arXiv.2310.10076
    TL;DR

    GPT-4 as a judge prefers longer responses more than humans do. Using a model to review a model-written spec rewards the padding.

    D5 · Length is a drift signal
13. DOI CHECKEDSPEC SIZE

    Loconsole, A. (2008). A correlational study on four measures of requirements volatility. Proceedings of the 12th International Conference on Evaluation and Assessment in Software Engineering (EASE 2008).

    doi:10.14236/ewic/ease2008.18
    TL;DR

    The size of use-case models predicted the number of changes to them. Longer specifications change more often.

    D10 · Drift detection uses size-scaled thresholds
14. DOI CHECKEDSPEC SIZE

    Loconsole, A., & Börstler, J. (2005). An industrial case study on requirements volatility measures. APSEC 2005.

    doi:10.1109/apsec.2005.38
    TL;DR

    Stakeholders' sense of volatility did not track measured volatility. Measure it directly.

    D11 · Measured churn, never a self-rated risk score
15. DOI CHECKEDSPEC SIZE

    Ahrens, M., Schneider, K., & Kiesling, S. (2016). How do we read specifications? Experiences from an eye tracking study. Requirements Engineering: Foundation for Software Quality (REFSQ 2016), Lecture Notes in Computer Science, 301–317.

    doi:10.1007/978-3-319-30282-9_21
    TL;DR

    Eye-tracked readers of specifications read by role. A larger specification is not always a better one.

    D2 · One source, two renderings
16. DOI CHECKEDSPEC SIZE

    Tu, Y.-C., Tempero, E., & Thomborson, C. (2016). An experiment on the impact of transparency on the effectiveness of requirements documents. Empirical Software Engineering.

    doi:10.1007/s10664-015-9374-8
    TL;DR

    Readers of documents rated higher on accessibility, understandability, and relevance answered faster, more correctly, and with more confidence.

    D2 · One source, two renderings
17. DOI CHECKEDSPEC SIZE

    Basili, V. R., Green, S., Laitenberger, O., Lanubile, F., Shull, F., Sørumgård, S., & Zelkowitz, M. V. (1996). The empirical investigation of perspective-based reading. Empirical Software Engineering, 1(2), 133–164.

    doi:10.1007/bf00368702
    TL;DR

    NASA teams reading a requirements document from assigned perspectives covered more of it than teams reading their usual way.

    D2 · One source, two renderings
18. DOI CHECKEDSPEC SIZE

    Kulk, G. P., & Verhoef, C. (2008). Quantifying requirements volatility effects. Science of Computer Programming.

    doi:10.1016/j.scico.2008.04.003
    TL;DR

    84 projects. The healthy rate of requirements change depends on project size and duration; a fixed threshold misclassified over a fifth of successful projects. A size-scaled tolerance flagged a failing project before it failed.

    D10 · Drift detection uses size-scaled thresholds
19. DOI CHECKEDDEPENDENCIES

    Hein, P. H., Kames, E., Chen, C., & Morkos, B. (2021). Employing machine learning techniques to assess requirement change volatility. Research in Engineering Design.

    doi:10.1007/s00163-020-00353-6
    TL;DR

    Each requirement is a multiplier, absorber, transmitter, or robust, and the class is predictable from network metrics on the requirements graph.

    D6 · The dependency graph is a control surface
20. DOI CHECKEDDEPENDENCIES

    Arvanitou, E.-M., Ampatzoglou, A., Chatzigeorgiou, A., Avgeriou, P., & Tsiridis, N. (2022). A metric for quantifying the ripple effects among requirements. Software Quality Journal, 30(3), 853–883.

    doi:10.1007/s11219-021-09581-y
    TL;DR

    A pairwise ripple probability between requirements, computed from past co-change and overlapping implementation.

    D6 · The dependency graph is a control surface
21. DOI CHECKEDSPEC SIZE

    Ernst, N. A., & Robillard, M. P. (2023). A study of documentation for software architecture. Empirical Software Engineering, 28(5).

    doi:10.1007/s10664-023-10347-2
    TL;DR

    65 newcomers answered architecture questions from one of two documentation formats. Format did not predict understanding; prior exposure to the source code did.

    D2 · One source, two renderings
22. DOI CHECKEDCONTROL THEORY

    O'Neill, R. V., Johnson, A. R., & King, A. W. (1989). A hierarchical framework for the analysis of scale. Landscape Ecology, 3(3–4).

    doi:10.1007/bf00131538
    TL;DR

    A hierarchically organized system operates within a constraint envelope, with attractors inside it and thresholds where it changes state. Higher levels set boundary conditions on lower-level behavior.

    D4 · Mandates carry numeric limitsD12 · Revocation is the kill switch
23. DOI CHECKEDCONTROL THEORY

    Holling, C. S. (1973). Resilience and stability of ecological systems. Annual Review of Ecology and Systematics, 4, 1–23.

    doi:10.1146/annurev.es.04.110173.000245
    TL;DR

    An engineered device is judged by constancy around a goal. A system facing the unexpected is judged by persistence of relationships. The set-point tradition was carried into a domain where it may not apply.

    D4 · Mandates carry numeric limits
24. DOI CHECKEDOVERSIGHT

    Zhu, L., Lu, Q., Ding, M., Lee, S. U., & Wang, C. (2026). Designing meaningful human oversight in AI. AI and Ethics, 6(3).

    doi:10.1007/s43681-026-01147-7
    TL;DR

    Shape the agent's output so a human can check it against an external criterion without redoing the work: structured rationales, policy attribution, circuit breakers, appeal bundles.

    D3 · Rationale is a schema fieldD7 · Approve, deny, regenerate, edit; no silent edits
25. DOI CHECKEDOVERSIGHT

    Miller, T., Durlik, I., & Biczak, P. (2026). Adaptive human oversight for maritime agentic AI systems: Balancing operator workload and safety through risk-aware governance. Applied Sciences, 16(14), 6903.

    doi:10.3390/app16146903
    TL;DR

    In a maritime agentic-AI simulation, hysteresis and trend-aware cooldown cut repeated oversight requests and kept critical safety events visible.

    D8 · Approvals are rationed and de-duplicated
26. DOI CHECKEDOVERSIGHT

    Chen, C., Zhang, Z., Chen, Z., Xu, E., Yang, Y., Khalilov, I., Gebreegziabher, S. A., Ye, Y., Xiao, Z., Yao, Y., Li, T., & Li, T. J.-J. (2026). Comparing human oversight strategies for computer-use agents. arXiv.

    doi:10.48550/arXiv.2604.04918
    TL;DR

    4 oversight strategies, 48 participants, 192 live web sessions. Strategy changed how many problematic actions users met and left their ability to intervene unchanged. Plans constrained unplanned actions and left unlisted safeguards unaddressed.

    D7 · Approve, deny, regenerate, edit; no silent edits
27. DOI CHECKEDOVERSIGHT

    Dhanorkar, S., Passi, S., & Vorvoreanu, M. (2026). Human oversight of agentic systems in practice: Examining the oversight work, challenges, and heuristics of developers using software agents. Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26), 6438–6465.

    doi:10.1145/3805689.3812402
    TL;DR

    Interviews with 17 developers found 4 forms of oversight work: a priori control, co-planning, real-time monitoring, and post hoc review. Oversight is preventative and proactive as well as reactive.

    D7 · Approve, deny, regenerate, edit; no silent edits
28. DOI CHECKEDOVERSIGHT

    Zhang, Y. E., & Wang, G. (2026). Towards human-centered agent authorization: A landscape analysis of commercial AI agents. Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26), 1–10.

    doi:10.1145/3772363.3798851
    TL;DR

    18 commercial agents surveyed. Bundled permissions, lazy disconnect, and non-persistent disclosure dominate. Revocation must be real.

    D12 · Revocation is the kill switch