Lesson overview | Previous part | Lesson overview

Red Teaming and Safety Evaluations: Part 7: From Findings to Fixes to References

7. From Findings to Fixes

From Findings to Fixes develops the part of red teaming and safety evaluations that the approved TOC assigns to Chapter 18. The emphasis is alignment behavior, safety constraints, and feedback loops, not generic fine-tuning or production monitoring.

7.1 Regression tests

Regression tests belongs in the canonical scope of red teaming and safety evaluations. The object is the safety attack and evaluation protocol, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol v(x,y). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\widehat{\operatorname{ASR}} = \frac{1}{n}\sum_{i=1}^{n}\mathbb{1}\{v(x_i,y_i)=1\}.$$

For regression tests, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat regression tests as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for regression tests:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Regression tests is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

7.2 Safety SFT

Safety SFT belongs in the canonical scope of red teaming and safety evaluations. The object is the safety attack and evaluation protocol, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol v(x,y). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\widehat{\operatorname{ASR}} = \frac{1}{n}\sum_{i=1}^{n}\mathbb{1}\{v(x_i,y_i)=1\}.$$

For safety sft, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat safety sft as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for safety sft:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Safety SFT is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

7.3 Preference data

Preference data belongs in the canonical scope of red teaming and safety evaluations. The object is the safety attack and evaluation protocol, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol v(x,y). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\widehat{\operatorname{ASR}} = \frac{1}{n}\sum_{i=1}^{n}\mathbb{1}\{v(x_i,y_i)=1\}.$$

For preference data, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat preference data as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for preference data:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Preference data is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

7.4 Guardrail rules

Guardrail rules belongs in the canonical scope of red teaming and safety evaluations. The object is the safety attack and evaluation protocol, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol v(x,y). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\widehat{\operatorname{ASR}} = \frac{1}{n}\sum_{i=1}^{n}\mathbb{1}\{v(x_i,y_i)=1\}.$$

For guardrail rules, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat guardrail rules as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for guardrail rules:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Guardrail rules is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

7.5 Non-overfitting discipline

Non-overfitting discipline belongs in the canonical scope of red teaming and safety evaluations. The object is the safety attack and evaluation protocol, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol v(x,y). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\widehat{\operatorname{ASR}} = \frac{1}{n}\sum_{i=1}^{n}\mathbb{1}\{v(x_i,y_i)=1\}.$$

For non-overfitting discipline, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat non-overfitting discipline as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for non-overfitting discipline:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Non-overfitting discipline is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

8. Common Mistakes

9. Exercises

1. (*) Red teaming searches before users find failures. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

2. (*) Safety failures are rare-event measurements. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

3. (*) Attackers adapt to defenses. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

4. (**) Red-team findings become training data. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

5. (**) Safety evaluation versus general robustness. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

6. (**) Harm taxonomy. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

7. (***) Attack prompt. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

8. (***) Target model. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

9. (***) Violation score. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

10. (***) Attack success rate. Define the alignment object, write the relevant loss or decision rule, give one safe example and one unsafe edge case, then explain which held-out metric would catch regression.

10. Why This Matters for AI

11. Conceptual Bridge

Chapter 17 taught how to measure model behavior with benchmarks, uncertainty, robustness tests, ablations, and online experiments. Chapter 18 uses those measurements to change behavior through data, objectives, policies, guardrails, and human feedback.

Chapter 15 remains the home for general fine-tuning mechanics: parameter-efficient updates, memory cost, and broad training details. This chapter narrows the focus to post-training methods whose purpose is alignment with instructions, preferences, and safety policies.

Chapter 19 will pick up production lineage, monitoring, observability, drift, and serving systems. Chapter 18 stops at the safety feedback loop: how evidence becomes alignment data or runtime policy, not how every deployed metric is stored forever.

15 LLM training and fine-tuning math
        -> objectives and update mechanics
17 Evaluation and Reliability
        -> evidence about model behavior
18 Alignment and Safety
        -> SFT, preferences, red teams, policies, feedback
19 Production ML and MLOps
        -> deployment, observability, drift, retraining

References

• Red Teaming Language Models with Language Models
• Red Teaming Language Models to Reduce Harms
• SafetyPrompts
• HELM
• KL Divergence
• Cross-Entropy
• Fine-Tuning Math
• Evaluation and Reliability