D2 Verification and evidence

Tests, independent checks, AI review, proof a change is safe.

Red flag: One model writes the code, the tests and the review; unreviewed merges rising.

The three statements

Each is rated from 0 (not true) to 4 (fully true). The dimension score is their average.

1. Automated tests give us enough confidence to ship routine changes.
2. Changes are checked by tests or reviewers that are independent of the AI that wrote them.
3. We track escaped defects and use them to improve tests and review.

Evidence to gather

Coverage, flaky-test rate, escaped defects, who writes the tests and the review, sampled human audits of approved changes.

Score from evidence, not impressions. Where two people score the same team, the statements they disagree on are where to look first. The assessment's compare view lists them.

First moves

If D2 is your weakest dimension, start here:

• Put an AI review on every change through a different path from the author agent, and audit a sample of approved changes each week.
• Log every escaped defect with the check that should have caught it, and add that check.

Score this in the assessment

Source: SPEC.md at v0.2.0.