Lesson overview | Previous part | Next part

Capability Benchmarks: Part 5: Statistical Reliability

5. Statistical Reliability

Statistical Reliability is the part of capability benchmarks 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.

5.1 Bootstrap intervals

Bootstrap intervals is part of the canonical scope of capability benchmarks. In this chapter, the object under study is not merely a dataset or a model, but the full benchmark protocol: 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

$$\hat{\mu}_{m,t} = \frac{1}{n}\sum_{i=1}^n s_m(z_i).$$

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 bootstrap intervals, 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 bootstrap intervals 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 bootstrap intervals:

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, bootstrap intervals 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. Bootstrap intervals is one place where that habit becomes concrete.

5.2 Paired model comparisons

Paired model comparisons is part of the canonical scope of capability benchmarks. In this chapter, the object under study is not merely a dataset or a model, but the full benchmark protocol: 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

$$\hat{\mu}_{m,t} = \frac{1}{n}\sum_{i=1}^n s_m(z_i).$$

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 paired model comparisons, 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 paired model comparisons 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 paired model comparisons:

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, paired model comparisons 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. Paired model comparisons is one place where that habit becomes concrete.

5.3 Leaderboard uncertainty

Leaderboard uncertainty is part of the canonical scope of capability benchmarks. In this chapter, the object under study is not merely a dataset or a model, but the full benchmark protocol: 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

$$\hat{\mu}_{m,t} = \frac{1}{n}\sum_{i=1}^n s_m(z_i).$$

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 leaderboard uncertainty, 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 leaderboard uncertainty 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 leaderboard uncertainty:

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, leaderboard uncertainty 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. Leaderboard uncertainty is one place where that habit becomes concrete.

5.4 Multiple comparisons

Multiple comparisons is part of the canonical scope of capability benchmarks. In this chapter, the object under study is not merely a dataset or a model, but the full benchmark protocol: 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

$$\hat{\mu}_{m,t} = \frac{1}{n}\sum_{i=1}^n s_m(z_i).$$

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 multiple comparisons, 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 multiple comparisons 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 multiple comparisons:

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, multiple comparisons 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. Multiple comparisons is one place where that habit becomes concrete.

5.5 Benchmark power and sample size

Benchmark power and sample size is part of the canonical scope of capability benchmarks. In this chapter, the object under study is not merely a dataset or a model, but the full benchmark protocol: 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

$$\hat{\mu}_{m,t} = \frac{1}{n}\sum_{i=1}^n s_m(z_i).$$

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 benchmark power and sample size, 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 benchmark power and sample size 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 benchmark power and sample size:

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, benchmark power and sample size 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. Benchmark power and sample size is one place where that habit becomes concrete.