Regression Testing

AI Regression Testing

You are setting up a regression testing framework for an AI feature — a systematic process that catches quality degradations caused by model changes, prompt changes, or data/context changes before they reach users.

Framework: Hamel Husain + Shreya Shankar (Building eval systems, 2025), software testing principles applied to AI.

Step 1 — Load Context

Read memory/user-profile.md for the AI feature being protected. Read the eval suite design if available — regression tests are a subset of the broader eval suite, focused on the specific failure modes the team has already identified and fixed.

Step 2 — What Triggers a Regression?

AI regressions can be caused by:

• Model updates: The underlying LLM changes (e.g., Claude version update)
• Prompt changes: The system prompt or few-shot examples are modified
• Context changes: The data passed to the model (retrieved documents, user context) changes
• Tool/API changes: External tools the model calls change their behavior
• Distribution shift: The types of inputs coming from users have changed over time

The regression test suite should catch all of these, not just obvious changes.

Step 3 — Building the Regression Test Set

The regression test set is a curated collection of (input, expected behavior) pairs. "Expected behavior" means the output should PASS a specific eval.

Sources for the test set:

1. Past failures that were fixed: If you fixed a failure in v1.0, add a test that would have caught it. This prevents it from recurring silently.
2. Edge cases from error analysis: The failure categories identified in error analysis become test cases.
3. Representative happy path examples: The most common, important use cases should have at least one test each.
4. Boundary cases: Inputs that are at the edge of the system's capability or scope.

Size guidance:

• Minimum viable: 20–30 carefully chosen test cases (covers major failure categories)
• Good: 50–100 test cases (covers failure categories + representative happy paths)
• Comprehensive: 200+ test cases (needed for high-stakes AI features)

Step 4 — Regression Test Execution

When to run:

• On every PR that changes the system prompt, model version, or retrieval pipeline
• On every scheduled model update (when the underlying model is upgraded)
• Weekly (to catch silent regressions from distribution shift)

Pass/fail definition: The test suite passes if: (1) every individual test case passes its specific eval, AND (2) the aggregate pass rate doesn't drop by more than [threshold]% from the baseline.

Set the threshold based on the feature's criticality:

• Core feature: < 2% regression acceptable
• Secondary feature: < 5% regression acceptable
• Experimental feature: < 10% regression acceptable

Step 5 — Regression Report Structure

When a regression is detected, produce a report:

What changed: Which eval(s) failed? What does the failure pattern look like?

Affected input types: Are regressions concentrated on certain types of inputs (short inputs, specific user segments, specific task types)?

Severity: How many test cases failed? What's the regression % vs. baseline?

Root cause hypothesis: What change (model, prompt, context) most likely caused this?

Rollback recommendation: Should the change be reverted immediately, or is this a degradation that can be fixed forward?

Fix plan: If fixing forward, what changes to the prompt or system would address the regression?

Step 6 — CI/CD Integration

Connect regression tests to the deployment pipeline:

pythonCopy code
# Pseudocode: regression gate in deployment pipeline
def run_regression_gate(eval_suite, test_cases, baseline_pass_rate, threshold=0.02):
    results = [run_eval(test_case, eval_suite) for test_case in test_cases]
    current_pass_rate = sum(1 for r in results if r.passed) / len(results)
    regression = baseline_pass_rate - current_pass_rate

    if regression > threshold:
        raise DeploymentBlockedError(
            f"Regression detected: {regression:.1%} quality drop vs. baseline. "
            f"Blocking deployment. Review failing cases: {[r for r in results if not r.passed]}"
        )

    return {"pass_rate": current_pass_rate, "regression": regression, "status": "PASS"}

Step 7 — Regression Triage Process

When a regression is flagged:

1. Identify which test cases failed (which failure categories?)
2. Compare failing outputs to the passing outputs from before the change
3. Determine root cause: is this a prompt issue, model issue, or retrieval issue?
4. Decide: revert (fastest) or fix forward (if the cause is known and the fix is simple)
5. After fixing, run the full regression suite before deploying
6. Update the test set: add the new failing cases as permanent regression tests

Step 8 — Output

Produce:

• Regression test set design (how many cases, which failure categories to cover, sources)
• Pass/fail thresholds and regression % tolerances
• Deployment gate integration plan
• Regression triage process (who does what when a regression is caught)
• The first 10 regression test cases to implement (highest-priority failures from error analysis)