THE SENTRIAL BLOG
Evaluations, regression tests, and production experiments.
Build a production evaluation workflow that measures false alarms, audits missed failures, and separates biased review queues from population estimates.
Read articleSeparate application tests, production evaluation and evaluator calibration so passing scores do not hide missing evidence or unseen failures.
Read articleKeep assignment, tool conditions and evaluation rules stable so a prompt comparison measures the agent change you intended to test.
Read articleTurn an observed failure into a reproducible test without losing the tool conditions or accepting a score change as proof of a fix.
Read article