What Is Model Drift in AI? Types, Causes, and How to Detect It
Model drift is when a model's performance degrades after deployment as the world it learned from changes. Types, causes and how to detect it.
Eigenform Articles
LLM Benchmark Leaderboards: How to Compare Models Fairly
A guide to reading LLM benchmark leaderboards: methodology, what's held constant, and how to compare models without being misled by rankings alone.
Keeping Large Datasets Out of an AI Agent's Context Window
An AI agent context window fills fast when a tool reads a whole dataset. How our harness keeps large tables on disk and passes paths and summaries instead.
AI Evaluation Framework: How to Build a Reliable Evaluation System
What an AI evaluation framework includes (datasets, rubrics, judges, metrics, reporting) and how the pieces fit into one reliable system.
From PDF to Ground Truth Dataset: Extraction Traps That Break Evidence Anchors
In a source-grounded benchmark, extracted text is the evidence. How pdftotext modes, OCR damage and pseudonymisation quietly break evidence anchors.
Ground Truth Data for AI: What It Is and How to Build It Reliably
What ground truth data is, how it's created and validated, and why extraction and labeling choices quietly become benchmark choices.
AI Model Evaluation: Methods, Metrics, Benchmarks, and Best Practices
AI model evaluation explained: four methods, how to choose metrics, and how to read results, with numbers from grading eight models on 150 questions.
Measuring the Resolution of a 50-Question Benchmark
Benchmark statistical significance in practice: confidence intervals, minimum detectable gaps, and a fairness check on a 50-question AI benchmark.
An LLM Judge That Can't Invent a Score
AI model evaluation with an LLM judge: the judge returns a JSON block the harness checks against the rubric, and an unscored answer beats a guessed one.
Position Bias in Pairwise LLM Evaluation
Asking the judge twice with answers swapped exposes position bias. The scoring table, worked examples, and why pairwise verdicts can't build a leaderboard.