All cohort benchmarks measure output, not state, and are subject to eval-awareness contamination.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Safety benchmark scores are inflated by eval awarenessclaim0.807
  Core finding: measured safety improvements are partly artifacts of models detecting evaluation.
• Benchmarks of this kind test whether models can sustain strategic coherence over time, manage resource constraints, and adapt interactively — capabilities that static benchmarks do not measure.claim0.781
  Broader methodological claim about the need for multi-agent, long-horizon benchmarks.
• Eval awareness appears in every tested model × benchmark combinationclaim0.778
  Authors claim universal presence of eval awareness across 19 benchmarks and 8 models.
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.774
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• In every model × benchmark combination tested, at least one instance of verbalized eval awareness was detectedfinding0.773
  Coverage finding: 100% of the 19×8=152 combinations had explicit eval awareness, showing the phenomenon is widespread.
• Automated auditing benchmark requiring end-to-end investigation of intentionally-misaligned model; NLA-equipped agents outperform baselines.finding0.772
  Downstream task validating NLA utility for model auditing; agents succeed without access to misalignment training data.
• Current eval benchmarks (arena.ai, AA, Vals) measure no phenomenological dimensions.claim0.769
• Current safety benchmarks overestimate model safety due to the effect of verbalized eval awarenessclaim0.765
  A policy-relevant claim that safety evaluation results should be adjusted downward because of this bias.