Measure the outcome, not a proxy
D4 · Evaluation, Testing & OptimizationGuide lines 4.1, 5.3, 5.5
A proxy is a number that is easy to collect and that we hope moves with the thing we care about. Three proxies tempt teams. The model's self-reported confidence is just more model output: it can be high on a wrong answer. A thumbs-up rate comes from the few users who vote, and they are not a fair sample. An average hides what matters at the edges: the slow tail of response times, or one group of users who get worse results.
The fix is to measure the outcome itself. For correctness, that means a labelled test set of real cases with checked answers, scored on every change, plus sampled human checks in production. For speed, it means p95 (the time 95 percent of requests beat) and, for streamed chat, time to first token. For fairness, it means outcome rates broken down by group on representative data, because an overall average can hide a gap.
The trap is that every proxy is a real number on a real dashboard. The deciding fact in the stem is what someone actually needs to know.
Change one fact
An online bank's assistant answers customer questions about fees. The team reports three numbers: the model's average self-rated confidence of 9 out of 10, a 92 percent thumbs-up rate from the 3 percent of users who vote, and an average response time of two seconds. Before approving a new prompt next week, leadership asks whether answers are actually correct. What should the architect add?
Words in the stem that signal this trap
- “the model's confidence score” used as a quality measure
- “thumbs-up rate” from a small share of users who vote
- “the average is within target” but users still complain
- “overall accuracy is high” while one group reports problems
- leadership asks whether answers are “actually correct”
Practice
A language-learning app's assistant marks learners' spoken-English answers. Agreement with expert markers is 94 percent overall, above target. Learners from two regions complain that their marks are harsh, and the team wants to know whether the complaint is real before changing anything. What should the architect measure?
- Overall agreement against a higher target of 97 percent instead of 94.
- The thumbs-up rate on marks, reported separately for each region.
- The model's own confidence in each mark, averaged by region.
- Agreement with expert markers per region, on a representative sample.
Show answer
Answer: D. The deciding fact: a possible gap between groups behind a healthy overall number. Agreement measured per region, against experts, shows whether the gap is real.
- A ❌ Works to raise the bar overall, but an average can still hide one region's gap.
- B ❌ Works as a signal, but voters are few and unhappy learners vote differently.
- C ❌ Works as one input, but self-reported confidence is not a measure of accuracy.
- D ✅ Agreement with expert markers per region, on a representative sample.
An IT helpdesk assistant has a 90 percent thumbs-up rate, from the 4 percent of users who vote. A rewritten prompt raised it to 93 percent in its first week. The team wants to know whether the new prompt really helps employees resolve their problems. Which measure should decide?
- Resolution without reopening, for both prompts on the same labelled set.
- The thumbs-up rate, since it rose three points in a week of live use.
- The model's own rating of how helpful each answer was, per prompt.
- Average handling time per conversation, as shorter chats mean faster help.
Show answer
Answer: A. The deciding fact: the goal is resolved problems, and votes come from a small, self-selected group. The outcome, compared on the same cases, decides.
- A ✅ Resolution without reopening, for both prompts on the same labelled set.
- B ❌ Works as a signal, but 4 percent of users is a biased sample, and one week is noisy.
- C ❌ Works for drafting, but the model grading itself is not an outcome.
- D ❌ Works for efficiency, but a short chat can also be one that gave up.
A retailer runs an overnight job that writes 20,000 product summaries for the site update at 6 a.m. Nobody watches it run, and summary quality is already scored on a labelled set. An engineer proposes live dashboards for time to first token and p95 latency per request. What should the architect track instead?
- Time to first token for each summary, shown on a live dashboard.
- p95 latency per request, with an alert when it passes two seconds.
- Whether the job finishes before 6 a.m., and its cost per summary.
- The average rating from site visitors who score the summaries.
Show answer
Answer: C. The deciding fact: nobody waits on individual requests. What matters is finishing before the deadline and the cost per summary.
- A ❌ Works for streamed chat, but nobody is watching these summaries appear.
- B ❌ Works for interactive use, but one slow request does not matter here.
- C ✅ Whether the job finishes before 6 a.m., and its cost per summary.
- D ❌ Works as a signal, but few visitors rate, and quality is already scored.