Eval-Driven Development

A formal evaluation framework implementing eval-driven development (EDD) principles. Evals are the "unit tests of AI development."

Philosophy

• Define expected behavior BEFORE implementation
• Run evals continuously during development
• Track regressions with each change
• Use pass@k metrics for reliability measurement

Usage

/eval define <feature-name>   # Create eval definition
/eval check <feature-name>    # Run and check evals
/eval report <feature-name>   # Generate full report
/eval list                    # Show all evals

Eval Types

Capability Evals

Test if a new capability works:

[CAPABILITY EVAL: feature-name]
Task: Description of what should be accomplished
Success Criteria:
  - [ ] Criterion 1
  - [ ] Criterion 2
Expected Output: Description of expected result

Regression Evals

Ensure changes don't break existing functionality:

[REGRESSION EVAL: feature-name]
Baseline: SHA or checkpoint name
Tests:
  - existing-test-1: PASS/FAIL
  - existing-test-2: PASS/FAIL
Result: X/Y passed

Grader Types

Code-Based Grader (Deterministic)

# Check if file contains expected pattern
grep -q "export function handleAuth" src/auth.ts && echo "PASS" || echo "FAIL"

# Check if tests pass
npm test -- --testPathPattern="auth" && echo "PASS" || echo "FAIL"

Model-Based Grader (Open-ended)

[MODEL GRADER PROMPT]
Evaluate the following code change:
1. Does it solve the stated problem?
2. Is it well-structured?
3. Are edge cases handled?
Score: 1-5

Human Grader

[HUMAN REVIEW REQUIRED]
Change: Description
Risk Level: LOW/MEDIUM/HIGH

Metrics

pass@k — "At least one success in k attempts"

• pass@1: First attempt success rate
• pass@3: Success within 3 attempts
• Target: pass@3 > 90%

pass^k — "All k trials succeed"

• pass^3: 3 consecutive successes
• Use for critical paths

Eval Workflow

1. Define (Before Coding)

## EVAL DEFINITION: feature-xyz

### Capability Evals
1. Can create new user account
2. Can validate email format

### Regression Evals
1. Existing login still works
2. Session management unchanged

### Success Metrics
- pass@3 > 90% for capability evals
- pass^3 = 100% for regression evals

2. Implement

Write code to pass the defined evals.

3. Evaluate

Run each eval, record PASS/FAIL.

4. Report

EVAL REPORT: feature-xyz
========================
Capability: 3/3 passed (pass@3: 100%)
Regression: 3/3 passed (pass^3: 100%)
Status: SHIP IT

Eval Storage

.codex/evals/
  feature-xyz.md      # Eval definition
  feature-xyz.log     # Eval run history
  baseline.json       # Regression baselines

Best Practices

1. Define evals BEFORE coding
2. Run evals frequently
3. Track pass@k over time
4. Use code graders when possible (deterministic > probabilistic)
5. Human review for security
6. Keep evals fast
7. Version evals with code