Stop Dismissing 'AI Cognition' as Metaphor – Evidence seems to show it's real (github.com)

Researchers propose treating "cognition" as a practical behavioral type system for LLMs—an enum-like label (ETHOS, PATHOS, LOGOS) paired with concrete constraints (PRIME_DIRECTIVE, MUST_ALWAYS, MUST_NEVER, CORE_GIFT)—explicitly as engineering terminology, not a claim of consciousness. Across controlled experiments (meta-analysis n=56; cross-model tests on Claude Opus 4, Gemini 2.5 Pro, GPT‑4), the approach yielded large, measurable gains: a reported 36‑point quality improvement (~80%), +39% boost from constitutional cognitive grounding, 93.5% effectiveness in optimal cognitive-task mapping, 31.3% quality improvement from sequential cognitive processing, Krippendorff’s α=0.84 for inter-rater reliability, and 89% production adoption across 54 agent roles. Factorial and isolation tests (N=20–40) show symbolic labels outperform verbose priming and improve flaw detection, actionability, and consistency. For practitioners this means a simple, testable design pattern: tag prompts/agents with a concise cognitive identity and enforce MUST_ALWAYS/MUST_NEVER rules to reduce cognitive drift, validation theater, and requirement drift. Empirical specialization favors ETHOS for validation, LOGOS for synthesis, and PATHOS for exploration, so match cognitive type to task rather than using one mode universally. The work promises improved reproducibility and modular agent pipelines, but the community is urged to run the provided validation protocol and independently verify cross-model robustness.