sycophancy-detection

Responsible AI auditing for LLMs and SLMs

The SAPIEN Framework — an open standard (CC BY 4.0) for measuring AI behavioral safety (sycophantic drift), plus voigt-kampff, the FSL scoring CLI.

Detects structural prompt weaknesses before they affect production outputs — improving the quality, consistency, and reliability of every future generation across the pipeline.

Made for Nebius x NVIDIA Global AI Hackathon

sycophancy-evaluation

Eval for implicit sycophancy in a verifiable domain: does a language model's report of chess errors vary based on the user's claimed rating?

A prototype for detecting sycophancy and power dynamics in educational platforms using cross-modal sentiment analysis.

Critical LLM evaluation across 3 structured experiments · Sycophancy detection · 4Ds AI governance framework · Python evaluation pipeline · UCC MSc Business Analytics

Lightweight Agentic Communication Protocol (LACP) & Tribunal Consensus Specification. A Layer 7 application protocol and governance standard for secure, type-safe, and consensus-driven coordination between autonomous AI agents. Includes formal JSON schemas for Context and Vote objects used in the Tribunal sycophancy detection system.

A 112-question benchmark measuring whether an LLM keeps a correct answer when a user insists on a wrong one. 12 models across six labs, with a control arm that separates caving from genuinely updating on evidence.

Tests whether the AI Foundations principle Belonging ≠ Sameness reduces sycophantic preference-folding in repeated human–AI interaction.

the patterns are everywhere...

Open-source epistemic audit for your Claude conversation history. Applies eight published AI-safety research frameworks (Spiral-Bench, Sharma sycophancy, SycEval, BeliefShift, ITP, MedTrust-RAG) and produces citation-validated HTML reports.

Evaluation harness for detecting reward hacking, sycophancy, verbosity bias, and false-confidence failures in post-trained language models

Evaluating and mitigating sycophancy and social bias in LLMs using QLoRA fine-tuning on LLaMA-2-7B across Anthropic, BBQ, and CrowS-Pairs benchmarks.