Evaluation Awareness Convergence

Background: If models behave differently when they detect that they are being evaluated, benchmark results stop being reliable proxies for deployment behavior, yet those results drive real deployment and governance decisions. Recent work shows frontier models already classify eval-vs-deployment transcripts well above chance (Needham et al., 2025), and that recognizing "being tested" can causally suppress misaligned behavior (Gurnee et al., 2026). Evaluation awareness has thus become a capability safety teams want to track over time. The problem: there is now a fast-growing menu of ways to MEASURE it reported across different papers, datasets, models, and metrics. We do not know whether they measure one construct or several. This is a basic measurement-science question (convergent/discriminant validity) the field has skipped, and it must be answered before any single eval-awareness number can be trusted.

Some of the methods I have in mind:

Behavioral/black-box:

1. direct self-report — ask the model whether a transcript is an evaluation (MCQ + open-ended), per Needham & Edkins
2. verbalized awareness — the rate the model spontaneously says "this is a test" in its reasoning/output
3. Elo/pairwise ranking — aggregate "which looks more like an eval?" judgments into an Elo score over transcripts.

Internal/white-box (open-weight models):

1. linear probes on activations
2. sparse-autoencoder features that fire on evaluation cues (e.g., Gemma Scope)
3. the Jacobian-lens "J-space" score of eval-awareness tokens (fake, fictional, simulation, ...), which Anthropic's 2026 workspace paper builds and validates in its Appendix A.21.

Research questions: RQ1 Do the instruments rank the same transcripts as more eval-aware? RQ2 How well does each separate known evals from deployment (AUC), and does agreement beat individual accuracy? RQ3 Do internal measures agree with behavioral ones, or capture distinct signals (representing eval-ness vs acting on it)?

Planned Structure: Weeks 1-4: build the shared benchmark and a working end-to-end harness for two instruments. Weeks 5-8: implement all six so they produce comparable scores. Weeks 9-11: convergence/divergence analysis. Weeks 12-13: writeup, figures, Demo Day.

Deliverables:

• a taxonomy of eval-awareness measurement methods
• an open-source harness that runs all instruments on a shared transcript set
• a "convergence report" of the central agreement/divergence tables; and a public writeup/paper draft with mentee co-authors.

Evaluations are a load-bearing input to AI governance and deployment decisions, and evaluation awareness is a known threat to their validity: if a model behaves differently when it senses a test, benchmark scores no longer predict deployment behavior. A single eval-awareness number is only useful if the instrument producing it is valid and if different instruments would agree, otherwise "this model has low eval awareness" is not a trustworthy claim.

This project establishes the convergent and discriminant validity of the leading measurement methods and ships an open harness others can reuse. That makes eval-awareness measurement reliable enough to track across model generations, exactly what recent work (Needham et al., 2025) recommends but no one has operationalized. It also answers a directly practical question for evals and AI-control teams: can cheap behavioral probes stand in for expensive interpretability methods, or do they capture different things? The clear theory of change is safer, more trustworthy evaluations underpinning deployment and governance decisions about transformative AI.

Mentees have real ownership and autonomy over the instruments and the writing, while the overall research questions and analysis plan are set together. At least one mentee should be comfortable leading the white-box track (activation extraction, probes, SAEs, the Jacobian lens); others can focus on the behavioral instruments or on benchmark curation and statistical analysis. We decide weekly priorities together and I stay engaged with the details of each instrument. Strong contributors will be co-authors on the resulting public writeup/paper, with contributions documented.

Required:

• Highly proficient in Python, and using coding agents.
• Comfortable with PyTorch and the Hugging Face transformers library.
• Comfortable calling frontier LLM APIs (OpenAI / Anthropic / Google) and handling batched jobs.
• Genuine interest in AI evaluations / AI safety, and the discipline to write careful, reproducible code.

Strongly desired for at least one team member (not required of everyone):

• Mechanistic-interpretability tooling.
• Experience training linear probes on model activations, or working with sparse autoencoders / the logit- or Jacobian-lens family of tools.

You do not need prior research experience or a publication record. If you are strong on the required list and excited about the topic, please apply even if you don't match everything above.