Test AI behavior on real examples You record realistic inputs, criteria, expected or forbidden actions, predictions, and traces. Use deterministic metrics when possible and judges when quality needs holistic review. cpp academy academy/topics/evals-metrics-judges website/content-src/academy/course.mjs academy Test AI behavior on real examples
Unit 8 · Measure and improve AI quality

Test AI behavior on real examples

You record realistic inputs, criteria, expected or forbidden actions, predictions, and traces. Use deterministic metrics when possible and judges when quality needs holistic review.

9 focused minutesNot started
Unit example (nearest native match)

See the idea in context

auto engine = axllm::AxGEPA(reflectionClient, axllm::object({}));
auto result = engine.optimize(request, evaluator);
  1. Use a realistic input

    The refund request represents the kind of task the agent will face.

  2. Write the success rule

    criteria explains that eligibility must be verified before action.

  3. Record observable behavior

    expectedActions lets the evaluation check tool selection, not only final prose.

Run itIn your own project
include(FetchContent)
FetchContent_Declare(axllm GIT_REPOSITORY https://github.com/ax-llm/ax GIT_TAG main SOURCE_SUBDIR packages/cpp)
FetchContent_MakeAvailable(axllm)
target_link_libraries(your_app PRIVATE axllm::axllm)

#include <axllm/axllm.hpp>

auto llm = axllm::ai("openai", axllm::object({{"apiKey", std::getenv("OPENAI_API_KEY")}}));
auto classify = axllm::ax("review:string -> sentiment:class \"positive, negative, neutral\"");

Set OPENAI_APIKEY in your environment before running provider-backed code.

In the ax repo

From a clone of the ax repo:

npm run example -- cpp src/examples/cpp/optimization/axgen_optimization.cpp
Active practice

Show that you can use it

Answer 2 in a row to learn this · attempt 1
Keep exploring

Source-backed follow-up