AI Engineer · November 5, 2026

Cameron Wolfe: calibrate subjective AI evaluations against humans and production

Cameron Wolfe: calibrate subjective AI evaluations against humans and production video thumbnail
Why it matters

Cameron Wolfe describes a three-part evaluation loop for tasks such as creative writing, where there is no single correct answer. Start with subject-matter experts defining a small set of quality criteria, rating instructions and concrete examples. Have multiple reviewers assess the same outputs, examine disagreements and revise the guidance before treating those labels as a reference. Use that calibrated human data to check automatic judges or task-specific rubrics, comparing judge agreement with the level humans achieve with each other. Offline evaluations can then support faster experimentation, but both the generator and judge introduce variability. Wolfe recommends realistic, discriminative task samples, sample-size planning, repeated rollouts and uncertainty estimates. Promising candidates return to human review and then controlled production A/B tests; user feedback, abandonment or other relevant measures provide additional evidence. Production failures feed back into the evaluation set. This is a method for making judgments more defensible, not a claim that an LLM judge turns subjective quality into ground truth.

My takeaway: Define quality and reviewer agreement before tuning an automatic judge. Keep a human reference set, report uncertainty alongside scores, and require both offline evidence and an appropriate production comparison before broad rollout. Add newly observed failure cases to the next evaluation cycle.
Keep exploring

More curated notes connected through Model Evaluation and AI Engineering.

OpenAI News · framework

Pacing model development in an era of cyber-critical capabilities

OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.

OpenAI News · framework

A blueprint for democratic governance of frontier AI

OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.

OpenAI News · framework

OpenAI’s Frontier Governance Framework

OpenAI's 22-page Frontier Governance Framework maps its frontier-model processes to California's Transparency in Frontier AI Act and the EU AI Act's general-purpose AI code. It documents lifecycle risk assessment, cyber-offense and other risk tiers, mitigation and residual-risk decisions, critical-incident handling, security risk management, model reporting, external review, responsibility allocation, and change control.