Lesson overview | Previous part | Next part

Online Experimentation and AB Testing: Part 2: Formal Definitions

2. Formal Definitions

Formal Definitions is the part of online experimentation and ab testing that turns the approved TOC into a concrete learning path. The subsections below keep the focus on Chapter 17's canonical job: measurement, reliability, uncertainty, and decision support for AI systems.

2.1 Treatment and control

Treatment and control is part of the canonical scope of online experimentation and ab testing. In this chapter, the object under study is not merely a dataset or a model, but the full online randomized experiment: the items, prompts, outputs, graders, uncertainty statements, and decision rules that turn model behavior into evidence.

The basic mathematical pattern is an empirical estimator. For a model or system $m$ evaluated on items $z_1,\ldots,z_n$, the local estimate is written

$$\widehat{\operatorname{ATE}} = \frac{1}{n}\sum_{i=1}^n Y_i(1)-Y_i(0).$$

The formula is intentionally simple. The difficulty lies in deciding what counts as an item, which loss or score is meaningful, whether the items are independent, and whether the estimate answers the real product or research question. For treatment and control, those choices determine whether the reported number is evidence or decoration.

A useful invariant is that every evaluation claim should be reproducible as a tuple $(m,\mathcal{T},\pi,g,\rho)$, where $m$ is the system, $\mathcal{T}$ is the task sample, $\pi$ is the prompt or intervention policy, $g$ is the grader, and $\rho$ is the aggregation rule. If any part of this tuple is missing, the number cannot be audited.

Examples of correct use:

• Report treatment and control with item count, prompt protocol, grader version, and a confidence interval.
• Use paired comparisons when two models answer the same evaluation items.
• Inspect at least one meaningful slice before concluding that the aggregate result is reliable.
• Store raw outputs so future graders can be replayed without querying the model again.
• Document whether the metric is measuring capability, reliability, user value, or risk.

Non-examples:

• A leaderboard point estimate without sample size.
• A benchmark score produced with an undocumented prompt template.
• A model-graded result without judge identity, rubric, or agreement check.
• A robustness claim measured only on the easiest in-distribution examples.
• An online win declared before the randomization and logging checks pass.

Worked evaluation pattern for treatment and control:

1. Define the evaluation population in words before writing code.
2. Choose the smallest metric set that answers the decision question.
3. Compute the point estimate and an uncertainty statement together.
4. Run a slice or paired analysis to check whether the aggregate hides structure.
5. Archive raw outputs, scores, and seeds before changing the prompt or grader.

For AI systems, treatment and control is especially delicate because the same model can be used with many prompts, decoding policies, tools, retrieval contexts, and safety filters. The measured quantity is therefore a property of the system configuration, not just the base weights.

Implementation checklist:

• Use deterministic seeds for synthetic or sampled evaluation subsets.
• Print metric denominators, not only percentages.
• Keep missing, invalid, timeout, and refusal outcomes explicit.
• Prefer typed result records over loose CSV columns.
• Separate raw model outputs from normalized grader inputs.
• Track the smallest reproducible command that generated the result.
• Record whether the estimate is item-weighted, token-weighted, user-weighted, or domain-weighted.
• Write the decision rule before seeing the final score whenever the result will guide a release.

The mathematical habit to build is skepticism with structure. A score is not ignored because it is noisy; it is interpreted through the design that produced it. Treatment and control is one place where that habit becomes concrete.

2.2 Randomization unit

Randomization unit is part of the canonical scope of online experimentation and ab testing. In this chapter, the object under study is not merely a dataset or a model, but the full online randomized experiment: the items, prompts, outputs, graders, uncertainty statements, and decision rules that turn model behavior into evidence.

The basic mathematical pattern is an empirical estimator. For a model or system $m$ evaluated on items $z_1,\ldots,z_n$, the local estimate is written

$$\widehat{\operatorname{ATE}} = \frac{1}{n}\sum_{i=1}^n Y_i(1)-Y_i(0).$$

The formula is intentionally simple. The difficulty lies in deciding what counts as an item, which loss or score is meaningful, whether the items are independent, and whether the estimate answers the real product or research question. For randomization unit, those choices determine whether the reported number is evidence or decoration.

A useful invariant is that every evaluation claim should be reproducible as a tuple $(m,\mathcal{T},\pi,g,\rho)$, where $m$ is the system, $\mathcal{T}$ is the task sample, $\pi$ is the prompt or intervention policy, $g$ is the grader, and $\rho$ is the aggregation rule. If any part of this tuple is missing, the number cannot be audited.

Examples of correct use:

• Report randomization unit with item count, prompt protocol, grader version, and a confidence interval.
• Use paired comparisons when two models answer the same evaluation items.
• Inspect at least one meaningful slice before concluding that the aggregate result is reliable.
• Store raw outputs so future graders can be replayed without querying the model again.
• Document whether the metric is measuring capability, reliability, user value, or risk.

Non-examples:

• A leaderboard point estimate without sample size.
• A benchmark score produced with an undocumented prompt template.
• A model-graded result without judge identity, rubric, or agreement check.
• A robustness claim measured only on the easiest in-distribution examples.
• An online win declared before the randomization and logging checks pass.

Worked evaluation pattern for randomization unit:

1. Define the evaluation population in words before writing code.
2. Choose the smallest metric set that answers the decision question.
3. Compute the point estimate and an uncertainty statement together.
4. Run a slice or paired analysis to check whether the aggregate hides structure.
5. Archive raw outputs, scores, and seeds before changing the prompt or grader.

For AI systems, randomization unit is especially delicate because the same model can be used with many prompts, decoding policies, tools, retrieval contexts, and safety filters. The measured quantity is therefore a property of the system configuration, not just the base weights.

Implementation checklist:

• Use deterministic seeds for synthetic or sampled evaluation subsets.
• Print metric denominators, not only percentages.
• Keep missing, invalid, timeout, and refusal outcomes explicit.
• Prefer typed result records over loose CSV columns.
• Separate raw model outputs from normalized grader inputs.
• Track the smallest reproducible command that generated the result.
• Record whether the estimate is item-weighted, token-weighted, user-weighted, or domain-weighted.
• Write the decision rule before seeing the final score whenever the result will guide a release.

The mathematical habit to build is skepticism with structure. A score is not ignored because it is noisy; it is interpreted through the design that produced it. Randomization unit is one place where that habit becomes concrete.

2.3 Overall evaluation criterion and guardrails

Overall evaluation criterion and guardrails is part of the canonical scope of online experimentation and ab testing. In this chapter, the object under study is not merely a dataset or a model, but the full online randomized experiment: the items, prompts, outputs, graders, uncertainty statements, and decision rules that turn model behavior into evidence.

The basic mathematical pattern is an empirical estimator. For a model or system $m$ evaluated on items $z_1,\ldots,z_n$, the local estimate is written

$$\widehat{\operatorname{ATE}} = \frac{1}{n}\sum_{i=1}^n Y_i(1)-Y_i(0).$$

The formula is intentionally simple. The difficulty lies in deciding what counts as an item, which loss or score is meaningful, whether the items are independent, and whether the estimate answers the real product or research question. For overall evaluation criterion and guardrails, those choices determine whether the reported number is evidence or decoration.

A useful invariant is that every evaluation claim should be reproducible as a tuple $(m,\mathcal{T},\pi,g,\rho)$, where $m$ is the system, $\mathcal{T}$ is the task sample, $\pi$ is the prompt or intervention policy, $g$ is the grader, and $\rho$ is the aggregation rule. If any part of this tuple is missing, the number cannot be audited.

Examples of correct use:

• Report overall evaluation criterion and guardrails with item count, prompt protocol, grader version, and a confidence interval.
• Use paired comparisons when two models answer the same evaluation items.
• Inspect at least one meaningful slice before concluding that the aggregate result is reliable.
• Store raw outputs so future graders can be replayed without querying the model again.
• Document whether the metric is measuring capability, reliability, user value, or risk.

Non-examples:

• A leaderboard point estimate without sample size.
• A benchmark score produced with an undocumented prompt template.
• A model-graded result without judge identity, rubric, or agreement check.
• A robustness claim measured only on the easiest in-distribution examples.
• An online win declared before the randomization and logging checks pass.

Worked evaluation pattern for overall evaluation criterion and guardrails:

1. Define the evaluation population in words before writing code.
2. Choose the smallest metric set that answers the decision question.
3. Compute the point estimate and an uncertainty statement together.
4. Run a slice or paired analysis to check whether the aggregate hides structure.
5. Archive raw outputs, scores, and seeds before changing the prompt or grader.

For AI systems, overall evaluation criterion and guardrails is especially delicate because the same model can be used with many prompts, decoding policies, tools, retrieval contexts, and safety filters. The measured quantity is therefore a property of the system configuration, not just the base weights.

Implementation checklist:

• Use deterministic seeds for synthetic or sampled evaluation subsets.
• Print metric denominators, not only percentages.
• Keep missing, invalid, timeout, and refusal outcomes explicit.
• Prefer typed result records over loose CSV columns.
• Separate raw model outputs from normalized grader inputs.
• Track the smallest reproducible command that generated the result.
• Record whether the estimate is item-weighted, token-weighted, user-weighted, or domain-weighted.
• Write the decision rule before seeing the final score whenever the result will guide a release.

The mathematical habit to build is skepticism with structure. A score is not ignored because it is noisy; it is interpreted through the design that produced it. Overall evaluation criterion and guardrails is one place where that habit becomes concrete.

2.4 Average treatment effect

Average treatment effect is part of the canonical scope of online experimentation and ab testing. In this chapter, the object under study is not merely a dataset or a model, but the full online randomized experiment: the items, prompts, outputs, graders, uncertainty statements, and decision rules that turn model behavior into evidence.

The basic mathematical pattern is an empirical estimator. For a model or system $m$ evaluated on items $z_1,\ldots,z_n$, the local estimate is written

$$\widehat{\operatorname{ATE}} = \frac{1}{n}\sum_{i=1}^n Y_i(1)-Y_i(0).$$

The formula is intentionally simple. The difficulty lies in deciding what counts as an item, which loss or score is meaningful, whether the items are independent, and whether the estimate answers the real product or research question. For average treatment effect, those choices determine whether the reported number is evidence or decoration.

A useful invariant is that every evaluation claim should be reproducible as a tuple $(m,\mathcal{T},\pi,g,\rho)$, where $m$ is the system, $\mathcal{T}$ is the task sample, $\pi$ is the prompt or intervention policy, $g$ is the grader, and $\rho$ is the aggregation rule. If any part of this tuple is missing, the number cannot be audited.

Examples of correct use:

• Report average treatment effect with item count, prompt protocol, grader version, and a confidence interval.
• Use paired comparisons when two models answer the same evaluation items.
• Inspect at least one meaningful slice before concluding that the aggregate result is reliable.
• Store raw outputs so future graders can be replayed without querying the model again.
• Document whether the metric is measuring capability, reliability, user value, or risk.

Non-examples:

• A leaderboard point estimate without sample size.
• A benchmark score produced with an undocumented prompt template.
• A model-graded result without judge identity, rubric, or agreement check.
• A robustness claim measured only on the easiest in-distribution examples.
• An online win declared before the randomization and logging checks pass.

Worked evaluation pattern for average treatment effect:

1. Define the evaluation population in words before writing code.
2. Choose the smallest metric set that answers the decision question.
3. Compute the point estimate and an uncertainty statement together.
4. Run a slice or paired analysis to check whether the aggregate hides structure.
5. Archive raw outputs, scores, and seeds before changing the prompt or grader.

For AI systems, average treatment effect is especially delicate because the same model can be used with many prompts, decoding policies, tools, retrieval contexts, and safety filters. The measured quantity is therefore a property of the system configuration, not just the base weights.

Implementation checklist:

• Use deterministic seeds for synthetic or sampled evaluation subsets.
• Print metric denominators, not only percentages.
• Keep missing, invalid, timeout, and refusal outcomes explicit.
• Prefer typed result records over loose CSV columns.
• Separate raw model outputs from normalized grader inputs.
• Track the smallest reproducible command that generated the result.
• Record whether the estimate is item-weighted, token-weighted, user-weighted, or domain-weighted.
• Write the decision rule before seeing the final score whenever the result will guide a release.

The mathematical habit to build is skepticism with structure. A score is not ignored because it is noisy; it is interpreted through the design that produced it. Average treatment effect is one place where that habit becomes concrete.

2.5 Power and Type I or Type II errors

Power and Type I or Type II errors is part of the canonical scope of online experimentation and ab testing. In this chapter, the object under study is not merely a dataset or a model, but the full online randomized experiment: the items, prompts, outputs, graders, uncertainty statements, and decision rules that turn model behavior into evidence.

The basic mathematical pattern is an empirical estimator. For a model or system $m$ evaluated on items $z_1,\ldots,z_n$, the local estimate is written

$$\widehat{\operatorname{ATE}} = \frac{1}{n}\sum_{i=1}^n Y_i(1)-Y_i(0).$$

The formula is intentionally simple. The difficulty lies in deciding what counts as an item, which loss or score is meaningful, whether the items are independent, and whether the estimate answers the real product or research question. For power and type i or type ii errors, those choices determine whether the reported number is evidence or decoration.

A useful invariant is that every evaluation claim should be reproducible as a tuple $(m,\mathcal{T},\pi,g,\rho)$, where $m$ is the system, $\mathcal{T}$ is the task sample, $\pi$ is the prompt or intervention policy, $g$ is the grader, and $\rho$ is the aggregation rule. If any part of this tuple is missing, the number cannot be audited.

Examples of correct use:

• Report power and type i or type ii errors with item count, prompt protocol, grader version, and a confidence interval.
• Use paired comparisons when two models answer the same evaluation items.
• Inspect at least one meaningful slice before concluding that the aggregate result is reliable.
• Store raw outputs so future graders can be replayed without querying the model again.
• Document whether the metric is measuring capability, reliability, user value, or risk.

Non-examples:

• A leaderboard point estimate without sample size.
• A benchmark score produced with an undocumented prompt template.
• A model-graded result without judge identity, rubric, or agreement check.
• A robustness claim measured only on the easiest in-distribution examples.
• An online win declared before the randomization and logging checks pass.

Worked evaluation pattern for power and type i or type ii errors:

1. Define the evaluation population in words before writing code.
2. Choose the smallest metric set that answers the decision question.
3. Compute the point estimate and an uncertainty statement together.
4. Run a slice or paired analysis to check whether the aggregate hides structure.
5. Archive raw outputs, scores, and seeds before changing the prompt or grader.

For AI systems, power and type i or type ii errors is especially delicate because the same model can be used with many prompts, decoding policies, tools, retrieval contexts, and safety filters. The measured quantity is therefore a property of the system configuration, not just the base weights.

Implementation checklist:

• Use deterministic seeds for synthetic or sampled evaluation subsets.
• Print metric denominators, not only percentages.
• Keep missing, invalid, timeout, and refusal outcomes explicit.
• Prefer typed result records over loose CSV columns.
• Separate raw model outputs from normalized grader inputs.
• Track the smallest reproducible command that generated the result.
• Record whether the estimate is item-weighted, token-weighted, user-weighted, or domain-weighted.
• Write the decision rule before seeing the final score whenever the result will guide a release.

The mathematical habit to build is skepticism with structure. A score is not ignored because it is noisy; it is interpreted through the design that produced it. Power and Type I or Type II errors is one place where that habit becomes concrete.