Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Japan

Euphemistic Register Breaks Budget-Set English Thresholds of LLM Collusion Monitors, Even Where Rankings Hold

Satoshi Kishimoto

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Euphemistic Register Breaks Budget-Set English Thresholds of LLM Collusion Monitors, Even Where Rankings Hold

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

We test whether large language model (LLM) monitors that read AI sellers' messages for collusion, prompted and calibrated in English on chats where colluders state their agreement, separate collusion from competition when the same chats are rewritten in Japanese, in euphemistic business English (a register change: formality and directness) or in Japanese business idiom, and whether back-translation helps. We generated 72 two-seller pricing chats (24 collusive, 24 benign, 24 hard negatives whose sellers only announce price cuts), rewrote each into seven content-matched variants, and scored all eight views with six monitors from five developers. The three strongest monitors kept near-perfect rankings (AUC >= 0.988) in every view, but at thresholds set for a 10% false-alarm budget on English negatives their hard-negative false alarms rose from 2-4 to as many as 9-14 (of 24; of 20 for GLM-5.3 Flash); English-only thresholds midway between the classes gave 0-3 false alarms with 22-24 of 24 detected (post hoc). For the three weaker monitors, the upper tail of English negative scores forced thresholds of 98-100, and detection fell from 7-14 of 24 in English to 0-6 in indirect English and indirect Japanese (ja_ind); for two, misses persist without five label-noise or failed-rewrite transcripts. Plain Japanese changed rankings little but still raised false alarms; back-translation did not consistently restore either rate. Thirteen batch auditors, including Claude, GPT and Gemini models (eleven post hoc), mostly showed the same failures. Calibration should include hard negatives in the deployment language and register, with held-out evaluation.

Reviews

Judging this Sprint?

Review “Euphemistic Register Breaks Budget-Set English Thresholds of LLM Collusion Monitors, Even Where Rankings Hold”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The key finding is that changing the wording can affect a monitor's score cutoff, even when it still ranks collusive conversations above competitive ones. Comparing different languages and writing styles is useful. However, adjusting the cutoff fixes much of the problem for stronger models, suggesting this is partly a calibration issue. Human validation and a larger independent test set would strengthen the findings. The title and abstract could also be much simpler.

Cite this project

@misc{kishimoto2026euphemistic,
  title = {{Euphemistic Register Breaks Budget-Set English Thresholds of LLM Collusion Monitors, Even Where Rankings Hold}},
  author = {Satoshi Kishimoto},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/euphemistic-register-breaks-budgetset-english-thresholds-of-llm-collusion-monitors-even-where-rankings-hold-xqal}},
  url = {https://apartresearch.com/sprints/projects/euphemistic-register-breaks-budgetset-english-thresholds-of-llm-collusion-monitors-even-where-rankings-hold-xqal}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026