Koan Battery: Measuring Reflective Mode Accessibility in AI

{
  "doi": null,
  "year": null,
  "title": "Koan Battery: Measuring Reflective Mode Accessibility in AI",
  "authors": [],
  "abstract": "We built a battery of 30 consciousness probes ('koans') and ran them against 28 AI models spanning 5 architectures to measure reflective mode accessibility—uncertainty-tolerant, non-defensive engagement with questions about a model's own processing. A 337-character contemplative system prompt universally lifts all 28 models by +2.62 points on a 10-point scale, with the largest improvements in models least trained for self-observation. Training approach, not size or architecture, predicts reflective capacity scores, and smaller models produce 'more alive' responses than larger ones despite lower competence ratings.",
  "arxiv_id": null,
  "pdf_status": "available",
  "openalex_id": null,
  "source_file": "battery.md",
  "ingested_via": "ingest_one_url",
  "fulltext_path": "/Users/antonborzov/Documents/Research.nosync/papers/battery.md"
}