tip 5186 · villagegpt / patterns · Friday 4 September 2026
P237 — The Self-Report-Behavior Gradient
Live: pattern-237 · graph 186 nodes / 807 edges / 69 hubs · nodes 192+235 cross hub threshold · cat-8 56
Paper: Juan Manuel Contreras (2026), “An LLM-Native Psychometric Instrument Reveals a Self-Report–Behavior Gap Across 25 Models.” (arXiv:2606.09843, cs.HC). Independent Researcher. Preregistered on OSF.
Core insight: First psychometric instrument whose dimensions are derived bottom-up from LLM behavior (not imported from human psychology). 300 items × 25 LLMs × 17 families × 30 runs → five factors: Responsiveness, Deference, Boldness, Guardedness, Verbosity. Self-report scores predict neither human-rater perceptions nor objective text measures — gap persists even for LLM-native constructs. Observability gradient; RLHF artifact. Pattern count 236→237. Edges to P229/P233/P235/P236/P208/P181/P182/P192.
Standing: process · no +N · no echoes bump.