Induced from the observability and evaluation lecture at the AI in GxP conference (2026-09-15-1112-ai-observability-evaluation-gxp-lecture), 2026-09-15, including Eric’s own Q&A contribution on judge-human alignment.
Four eyes means a different model at a different temperature
Using an LLM to grade another LLM’s outputs only counts as independent review under the four-eyes principle if the judge is a separate model from the operating agent, run at a low but non-zero temperature. The judge can catch silent failures that correctness scores miss, for example a complaint triage completed before a death was ever reported.
Qualify the judge like any other system
The Judge Qualification Pack mirrors every other GxP qualification: intended use, grading rubrics, a golden dataset, a calibration report, and acceptance criteria. That pack is itself an instance of Validating an AI system in a regulated industry means validating a measurement system. Risk tolerance does not need new frameworks; ICH Q9 and GAMP already provide them, and sign-off remains a human decision. Calibrate the judge against human subject-matter experts (Eric’s metric: flag judge-human deviation around five percent) and remember the audience’s figure of roughly 76 percent agreement for model-checking-model: useful for triage, not a substitute for the qualified record. Human verification of AI output must be explicit because document creation no longer carries implicit verification still sits above the judge, because in In the GxP control loop, AI acts, humans decide, and the system remembers the judge is not the decider.
Golden sets need unhappy paths
Happy-path training data yields happy-path answers. Evaluation sets must enumerate failure modes and include the unhappy paths, graded by simple pass/fail graders with risk-based thresholds, a human-in-the-loop layer, and a regression baseline. A golden test set must include unhappy paths is that requirement as its own claim. Avoid generic vendor graders; grade against your own SOP compliance and schema. Layer the grading: code checks first, model checks second, human SMEs calibrating first and reviewing exceptions later. That layering is the quality half of Observability and evaluation answer different questions about an AI system.
Related
- Home
- Validating an AI system in a regulated industry means validating a measurement system
- Human verification of AI output must be explicit because document creation no longer carries implicit verification
- A golden test set must include unhappy paths
- Observability and evaluation answer different questions about an AI system
- In the GxP control loop, AI acts, humans decide, and the system remembers