THE SHORT ANSWER
Three steps, published on Product Talk on September 2, 2026. One, error analysis: read the model's outputs yourself, at volume, and write down the two or three mistakes that recur, ranked by what they cost a customer. Two, match each error to an eval type: a code assertion when the check is deterministic, a golden dataset when there is one right answer, an LLM-as-a-judge when the judgment is semantic, and customer feedback signals such as regeneration and editing for real usage, always trying the cheapest first and using it to filter before the expensive one. Three, the experimentation loop: fixed production-like inputs, a baseline, one change, compare across every eval, iterate. Her rule underneath all of it is that correctness is the product team's definition and never the vendor's. The step I add is the gate: a passing eval on a reversible action buys autonomy, a passing eval on an irreversible one buys a draft and a human, and model confidence never decides.
Torres's guide is the primer most product teams needed and did not have, and the order is the thing to keep. Error analysis before any measurement. Eval type chosen by the error, cheapest first. A loop that changes one thing and compares across every eval.
The part I add is what happens when the eval passes and the system is an agent that acts. A pass on something reversible earns autonomy. A pass on something irreversible earns a draft and a person. Confidence never decides.
The three linked pieces above are the post with the additions, the handbook chapter as a weekly routine, and the toolkit kit with all six templates.