Unreliability or Disagreement?

What an interpercentile decomposition of multi-turn evaluation results can and cannot show. Accepted as a poster (non-archival) at the NeurIPS 2026 Trust-AI-Eval workshop.

Highlights

Single author

Accepted as a poster (non-archival), NeurIPS 2026 Trust-AI-Eval (TAE) workshop, to be presented Dec 2026

Overview

"Unreliability or Disagreement? What an Interpercentile Decomposition Can and Cannot Show" is a single-author paper accepted as a poster (non-archival) at the NeurIPS 2026 Trust-AI-Eval (TAE) workshop. It will be presented in Dec 2026.

What it covers

What an interpercentile decomposition of multi-turn evaluation results can and cannot separate, tested on conversations generated across seven models.

Related projects

AI Safety

Attestix

Open-source (Apache-2.0) core for AI agent identity, verifiable credentials, delegation chains and hash-chained audit evidence. 47 MCP tools and verifiers in Python, JS, Go, Rust, Java and R, producing compliance evidence for EU AI Act requirements.

Developer Tools

CodeSage

A developer productivity tool leveraging static code analysis and machine learning to enhance software quality. Integrates with development workflows to provide intelligent code review assistance.

Agriculture Tech

DharaLink

A comprehensive agricultural technology platform connecting farmers with modern IoT infrastructure and data analytics to transform traditional farming practices across India.

Like what you see? Let's talk.

I am always open to discussing new projects, collaborations, or opportunities.