Skip to content
Apart Research
Apart Research
Contact us
Sprint projectFeb 2, 2026London

Same Question, Different Lies: Cross-Context Consistency (C3) for Black-Box Sandbagging Detection

Yulong Lin · Team sandbaggers

Submitted to The Technical AI Governance Challenge. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Same Question, Different Lies: Cross-Context Consistency (C3) for Black-Box Sandbagging Detection

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

As language models grow more capable, governance frameworks—from the EU AI Act to Anthropic's RSP and OpenAI's Preparedness Framework—increasingly rely on capability evaluations to trigger safety mitigations. But what if models can deliberately underperform on these evaluations? This behavior, known as sandbagging, could let models evade safety measures designed for their true capability level. We introduce Cross-Context Consistency (C3), a general framework for unsupervised black-box sandbagging detection that exploits a fundamental asymmetry: when a model truly lacks capability, its confusion manifests consistently across paraphrased questions, but when a capable model feigns incompetence, its strategic choices about how to appear weak create detectable inconsistencies. Our simple instantiation—using embedding cosine distance and mean aggregation—requires no training data or model-specific adaptation, making it practical for external auditors. We evaluate C3 across 98 model organisms spanning prompted, fine-tuned, and incentivized sandbagging on GPQA, MATH, and USACO, achieving 67–100% recall at 5% FPR where existing black-box methods fail (lie detectors collapse from 89% to 2% recall when controlling for model identity). Adversarial strategies with deterministic error recipes can reduce detection to 4.7–46%, but complementary methods compensate—motivating defense-in-depth deployment for governance applications.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

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

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety 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 AI safety 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.

  1. Honestly, the best part of this paper might be the thing they broke rather than the thing they built. The lie detector takedown (recall dropping from 89% to 2% once you control for model identity) is a "the emperor has no clothes" moment for that whole line of work, and the field needs more of that. C3 itself is dead simple: embed, compare, threshold. That simplicity makes it easy to deploy, sure, but it also means a sufficiently motivated adversary who trains for consistent deception could probably beat it. And the adversarial numbers back this up: 4.7% recall against deterministic error recipes is basically a coin flip away from useless. The authors know this and pitch "defense-in-depth" instead. Fair enough, but that raises an awkward question nobody in this space wants to answer: if every detection method has a known bypass, what does "compliant auditing" actually look like in practice? The paper also only covers benchmarks where you can check the right answer. How C3 behaves on open-ended tasks (persuasion, planning, creative problem-solving) is a total unknown, and those are arguably the scarier capability domains anyway.

    Read full reviewShow less

  2. This is outstanding work and was a pleasure to read! The insight that sandbagging models suffer from a systematic explanation inconsistency is really interesting and has real impact potential. Also the thourough baseline critique (lie detectors learn model identity, trusted monitoring tracks

    correctness not intent) was very interesting to read and seems independently valuable.

    Adversarial robustness of this detection method is the key open question: Really appreciate the transparency about the "deterministic mistakes" attack and about the lack of testing against RL-optimized evasion. More exploration and analysis of the main attack models and robustness against them would be the main way to strengthen this contribution.

    Also consider adding some illustrative examples of inconsistent explanations by sandbagging models. That would make the paper even more interesting to read.

Cite this project

@misc{lin2026same,
  title = {{Same Question, Different Lies: Cross-Context Consistency (C3) for Black-Box Sandbagging Detection}},
  author = {Yulong Lin},
  year = {2026},
  month = feb,
  note = {Submitted to The Technical AI Governance Challenge, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/same-question-different-lies-crosscontext-consistency-c3-for-blackbox-sandbagging-detection-7r2i}},
  url = {https://apartresearch.com/sprints/projects/same-question-different-lies-crosscontext-consistency-c3-for-blackbox-sandbagging-detection-7r2i}
}

More from The Technical AI Governance Challenge

All 47 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026