AI Model Evaluation: Fact-Grounded Tests for Frontier Models

tech_projects

Overview

As an AI Training Expert at Micro1, I’ve designed knowledge-seeking evaluation tasks for frontier AI models. Each task asks whether a model can answer a question grounded in one indisputable, source-verifiable fact, and each is built to reveal where models fail. Project and client details are confidential, so this page describes the method rather than the content.

My Role

  • Designed question-answer evaluation tasks, each anchored to a single, indisputable, source-verifiable fact
  • Engineered and tested prompts against live model outputs over multiple trials, refining each question until it reliably exposed a genuine failure mode
  • Vetted niche, specialized sources against strict accessibility and reliability criteria
  • Wrote quote-supported justifications that tie every claim to primary-source text

Workflow & Quality

  • Every task went through a multi-stage peer quality review: initial review, independent verification, and sign-off
  • Tasks were accepted only when they met shared evidentiary standards
Lee-Ann Vidal Covas
Authors
Language Scientist (PhD, Boston University) bridging sociolinguistic research and AI, with expertise in model evaluation, dataset curation, and applied data science.