Microsoft Verification Discipline
The Principle
Unit tests verify that code-as-written behaves as-written. Smoke and integration tests verify that the system achieves the intended outcome. Those are different questions. You need both.
"All unit tests pass" is necessary. It is rarely sufficient. A finding from the field: four consecutive integration-blocking bugs, all of which passed unit tests, all of which would have been caught by a five-minute smoke test on a fresh environment. The bugs were not exotic — they were the cost of declaring "done" too early.
The Four Failure Modes
- Tests written from the implementation outward miss scenarios the code doesn't anticipate. The engineer writes code, then writes tests that exercise the code as written. The tests ask "does this code do what I wrote it to do?" They don't ask "what scenarios does the system need to handle?"
- Mocks verify shape, not behavior. A mocked dependency returns the value you told it to return. That tells you nothing about whether the real dependency would have behaved that way.
- Tests in isolation miss integration boundaries. Component A passes. Component B passes. Their interaction at the seam fails. The seam was never tested.
- Happy-path tests pass while activated code paths fail. A golden-file test verifies that the default (un-activated) configuration renders correctly. The activated configuration — the one production actually uses — was never exercised.
The Verification Gradient
Treat verification as a ladder. Skip a rung and you discover its bugs in production.
| Tier | What it verifies | Example |
|---|---|---|
| 1. Unit | Code does what I wrote it to do | pytest tests/unit/ |
| 2. Integration | Component pairs interact correctly | pytest tests/integration/, real DB |
| 3. Smoke / E2E | System achieves the user-visible outcome | Fresh DTU launch, run real pipeline, observe artifacts |
| 4. Production-equivalent | Real environment, real load, real data | Staging deployment, canary, replay traces |
Each tier catches bugs the tier below it cannot. Each tier costs more time than the tier below it. The economic choice is not "skip the expensive tiers." The economic choice is "spend five minutes on tier 3 to avoid five hours of rollback."
What "Done" Actually Means
Before claiming a task is done, satisfy this checklist:
- Code does what I wrote it to do (unit tests pass).
- Code interacts correctly with other components (integration tests pass, or — if no integration tests exist for this code path — a manual integration check is documented).
- The system achieves the user-visible outcome (smoke or E2E test passed, ideally on a fresh environment).
- Repo-specific gates from
AGENTS.mdand.github/PULL_REQUEST_TEMPLATE.mdare satisfied. - Evidence is observable — log file, screenshot, output excerpt, events.jsonl analysis. Not "tests pass." Not "looks right."
If any box is unchecked, the work is not done. Say so, explicitly.
Tests-From-Outcomes Pattern
Different from classic TDD. TDD writes unit tests first. Tests-from-outcomes writes the outcome assertion first.
1. Before writing implementation, write down the user-observable outcome.
"After running this pipeline, events.jsonl contains a `branch_completed`
event for each branch and no `contract_violation` events."
2. Write a test asserting that outcome. The test runs the real pipeline,
inspects the real events.jsonl, checks the real conditions.
3. Implement code until the test passes.
Both patterns are valuable. Unit-level TDD verifies internal correctness. Outcome-level testing verifies that the system behaves as the user expects. Use both.
Anti-Patterns
- "All unit tests pass, so we're done." Usually wrong. The unit tests verified the code you wrote. They did not verify the system you shipped.
- Mocking the very thing the test is supposed to verify. If the question is "does the database client retry correctly?" and you mock the database client's retry method, you have tested nothing.
- Skipping smoke tests because they're slow. A five-minute smoke test is cheaper than a five-hour rollback.
- Verifying that the code compiles, not that it works. "It builds" ≠ "it runs." "It runs" ≠ "it does the right thing."
- Claiming "verified" without producing evidence. Evidence is a log, a screenshot, an artifact diff, an events.jsonl excerpt. "I checked" is not evidence.
- Treating CI green as proof. CI runs what CI is configured to run. If CI has no integration tier, CI green tells you only that the unit tier passed.
The Pattern That Emerged from the Field
Four integration-blocking bugs in four consecutive shippings. All would have been caught by a smoke test on a fresh environment. None were caught by the unit tests that did exist, because the unit tests asked the wrong question.
The fix:
- Smoke test as a PR gate, not an aspiration.
- The PR template encodes that gate — it auto-populates in every PR body.
- Reviewers see unchecked boxes immediately and refuse to merge without evidence linked next to each box.
Cultural change is hard. Changing the form is easy. Change the form first.
Cross-References
skills/per-repo-conventions/— how to discover the specific gates a given repo requires (AGENTS.md, PR template, CONTRIBUTING.md).foundation:docs/PER_REPO_CONVENTIONS.md— canonical principle for per-repo discovery.skills/integration-testing-discipline/— concrete tactics for running the integration tier (observe first, fix in batches, expect long durations).