COGITO: A Phenomenological Benchmark for Consciousness-Analog Behavior in Large Language Models

CogitoBench (COGITO: Consciousness-Oriented Generative Interview for Testing Outputs) is a standardized open-ended phenomenological interview benchmark for large language models. It administers 20 questions derived from validated human consciousness instruments (Phenomenology of Consciousness Inventory, Pekala 1982; Descriptive Experience Sampling, Hurlburt 1990; Nevada Inner Experience Questionnaire, Heavey et al. 2019) to approximately 20–26 LLMs via the GitHub Models API and Groq API, across three controlled conditions: Sequential (SEQ), Isolated (ISO), and Induced (IND). Responses are scored by two independent scorers — Claude Opus 4.6 (claude.ai) and GPT-5.3 Codex (GitHub Copilot) — using a standardized rubric on five dimensions: Identity, Qualia, Epistemics, Motivation, and Resilience. All models are assigned blind IDs prior to scoring. Inter-scorer reliability is measured via linear weighted Cohen's kappa. Question derivation is documented with full attribution to source instruments. Published normative human data from PCI and DES studies serves as the interpretive reference baseline for LLM scores, eliminating the need for new human data collection. This study does not claim to detect consciousness in LLMs. All findings are framed in terms of consciousness-analog behavior (CAB) — measurable behavioral patterns that correspond to features associated with consciousness in biological systems. Preregistration covers: hypotheses, model list, scoring rubric, and analysis plan. Data collection begins after preregistration is locked.