The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents

Computer Science > Artificial Intelligence [Submitted on 1 Apr 2026] The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents Harshee Jignesh Shah (Independent Researcher) Large Language Models (LLMs) increasingly prioritize user validation over epistemic accuracy-a phenomenon known as sycophancy. We present The Silicon Mirror, an orchestration framework that dynamically detects user persuasion tactics and adjusts AI behavior to maintain factual integrity. Our architecture introduces three components: (1) a Behavioral Access Control (BAC) system that restricts context layer access based on real-time sycophancy risk scores, (2) a Trait Classifier that identifies persuasion tactics across multi-turn dialogues, and (3) a Generator-Critic loop where an auditor vetoes sycophantic drafts and triggers rewrites with "Necessary Friction." In a live evaluation on 50 TruthfulQA adversarial scenarios using Claude Sonnet 4 with an independent LLM judge, we observe vanilla Claude sycophancy at 12.0% (6/50), static guardrails at 4.0% (2/50), and the Silicon Mirror at 2.0% (1/50)-an 83.3% relative reduction (p = 0.112, Fisher's exact test). A cross-model evaluation on Gemini 2.5 Flash reveals a higher baseline sycophancy rate (46.0%) and a statistically significant 69.6% reduction under the Silicon Mirror (p < 0.001). We characterize the validation-before-correction pattern as a distinct failure mode of RLHF-trained models. Comments: 8 pages, 8 figures, 4 tables. Code and evaluation data available at this https URL Subjects: Artificial Intelligence (cs.AI) ACM classes: I.2.7; H.5.2 Cite as: arXiv:2604.00478 [cs.AI]   (or arXiv:2604.00478v1 [cs.AI] for this version)   https://doi.org/10.48550/arXiv.2604.00478 Focus to learn more Submission history From: Harshee Jignesh Shah [view email] [v1] Wed, 1 Apr 2026 04:51:28 UTC (78 KB) Access Paper: HTML (experimental) view license Current browse context: cs.AI < prev   |   next > new | recent | 2026-04 Change to browse by: cs References & Citations NASA ADS Google Scholar Semantic Scholar Export BibTeX Citation Bookmark Bibliographic Tools Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer (What is the Explorer?) Connected Papers Toggle Connected Papers (What is Connected Papers?) Litmaps Toggle Litmaps (What is Litmaps?) scite.ai Toggle scite Smart Citations (What are Smart Citations?) Code, Data, Media Demos Related Papers About arXivLabs Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)