Risk Evaluation · Method

Monitoring Monitorability

NUWA research note · independent review pending

First published: 2025-12-20Source checked: 2026-10-25

Original source ↗Cite this record ↓Deception →

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

NUWA interpretation: monitorability is a two-sided agent–monitor measurement, conditional on evaluation archetype and scope.

Research question

How can monitorability be measured across different kinds of observable behavior?

Design and assumptions

The paper defines intervention, process and outcome-property evaluations, multiple monitor scopes and dataset-level g-mean² metrics.

Source findings

CoT monitoring outperformed action-only monitoring in practical tested settings.

Longer reasoning was generally more monitorable; studied frontier-scale RL did not materially reduce monitorability.

NUWA interpretation

NUWA interpretation: monitorability is a two-sided agent–monitor measurement, conditional on evaluation archetype and scope.

Limits and contrary evidence

Intervention estimates can be undefined or unstable when effects are nonpositive or tiny; present training-scale results do not guarantee future preservation.

Research and decision relevance

Model recommendation: cite archetype, scope, inclusion criteria and uncertainty with any aggregate monitorability score.

Precise source locations

Monitoring Monitorability · AbstractSelected original resource https://arxiv.org/html/2512.18311; sha256:3c73b648d18ee87eca3d3b306dedc3746665058c627189715b0662e4985d31bd; archived source capture; exact capture time not independently established
Monitoring Monitorability · §2 Evaluating MonitorabilitySelected original resource https://arxiv.org/html/2512.18311; sha256:3c73b648d18ee87eca3d3b306dedc3746665058c627189715b0662e4985d31bd; archived source capture; exact capture time not independently established
Monitoring Monitorability · §3.1.3 Evaluation Design GuidanceSelected original resource https://arxiv.org/html/2512.18311; sha256:3c73b648d18ee87eca3d3b306dedc3746665058c627189715b0662e4985d31bd; archived source capture; exact capture time not independently established
Monitoring Monitorability · §9 Discussion and ConclusionSelected original resource https://arxiv.org/html/2512.18311; sha256:3c73b648d18ee87eca3d3b306dedc3746665058c627189715b0662e4985d31bd; archived source capture; exact capture time not independently established

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Institutions and roles

Research by OpenAI

Authors (12)

Melody Y. Guan · Miles Wang · Micah Carroll · Zehao Dou · Annie Y. Wei · Marcus Williams · Benjamin Arnav · Joost Huizinga · Ian Kivlichan · Mia Glaese · Jakub Pachocki · Bowen Baker

Related work

Extends · Beyond g-mean²: complementary metrics for monitorability

Beyond g-mean²: complementary metrics for monitorability

Revisits the original monitorability metric for complementary decision objectives.

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{monitoring-monitorability,
  title = {Monitoring Monitorability},
  url = {https://arxiv.org/abs/2512.18311},
  author = {Melody Y. Guan and Miles Wang and Micah Carroll and Zehao Dou and Annie Y. Wei and Marcus Williams and Benjamin Arnav and Joost Huizinga and Ian Kivlichan and Mia Glaese and Jakub Pachocki and Bowen Baker},
  year = {2025},
  date = {2025-12-20},
  version = {arXiv:2512.18311v1},
  note = {arXiv:2512.18311v1}
}

NUWA record · BibTeX

@misc{monitoring-monitorability-nuwa,
  title = {Monitoring Monitorability — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/monitoring-monitorability/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--monitoring-monitorability.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--monitoring-monitorability.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Miles Wang, Micah Carroll, et al.. Monitoring Monitorability. 2025-12-20. v1 [arXiv]; Paper author block and correspondence at openai.com; Abstract; §2 Evaluating Monitorability
[2] Melody Guan, Yuxiang Xie, et al.. Beyond g-mean²: complementary metrics for monitorability. 2026-10-20 ; Research background and discussion in the original publication