Counterbalanced 4+4 Hybrid Social Arena: Evaluating Anti-Sycophancy RL, Long-Horizon Grudge Memory, and Full-Duplex Epistemic Barge-In Across Human-Steered Twins and Autonomous Agents
Year: 2026
Type: article
Abstract: Pre-registered experimental protocol, MLCommons Croissant 1.1 dataset schema, and SHA-256 hash-chained telemetry ledger for the 30 Days NYC 4+4 Hybrid Social Arena (4 Human Twins [TWIN-01..04, 2F+2M] vs. 4 Autonomous AI Agents [AGNT-05..08, 2F+2M]). Principal Investigator: Christopher Johns (Beatriful, Inc., ORCID: 0009-0000-3518-2383). Pre-registers four primary hypotheses (H1-H4) under Holm-Bonferroni FWER control (alpha=0.05, alpha_adj=0.0125) and TOST equivalence (+/-0.40 SD). Stage 0 calibration baseline: N=80 turns (18 supervisory overrides, 48.4% lexical divergence, p_adj=0.0080). Embargoed until 2027-05-15 to preserve double-blind peer review across ACL, IJCAI, ICML, Interspeech, and NeurIPS 2027.
Source: Open Science Framework
Author: CHRISTOPHER JOHNS
Language: English
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Keywords: human-AI interaction, lexical divergence
Open Access status: green
OpenAlex ID: W7226547012
