Modern Quality Engineering Lab · reports Repository ↗ · ozgurcetintas.dev ↗

Published by CI · commit 5b80442 · 2026-09-16

The evidence, not the claim.

Built by Özgür Çetintaş, Senior Test Automation Engineer / SDET. A production-style demonstration of how I design quality systems: choosing the cheapest reliable test layer for each risk and publishing the evidence through CI.

Every push to main runs the gates below and republishes this page. Each layer exists because it is the cheapest reliable place to prove a specific risk, not because the tool belongs in the stack; the test strategy and the decision records explain each placement.

Quality signals · run #16
SignalCurrent evidenceContext
CI✓ passingevery gate green in run #16
Mutation score96.4%133 of 138 injected faults caught · CI breaks under 80%
Coverage100% lines · 100% branchesthresholds enforced: 90% lines, 85% branches
E2E + accessibility7 passed · 0 failed0 flaky · Chromium · axe WCAG A/AA
Performancep95 2.22 ms · 0% errors1,000 requests · CI-local workload, indicative, not a production benchmark
AI evaluation3/3 passeddeterministic assertions · anthropic:messages:claude-opus-5
Core gates · the risk-to-layer decisions
  1. 01

    Unit and API tests

    Pure validation logic at every boundary, then in-process HTTP semantics: 400 versus 500, malformed JSON, a failing repository. No browser, milliseconds.

    CI job ↗
  2. 02

    Mutation testing

    Do those suites notice a bug, or merely execute the code? Stryker injects faults into src/; surviving mutants are highlighted in source and CI fails under 80%.

    Report ↗
  3. 03

    Contract tests

    Consumer and provider can change independently: the UI client's Pact is generated by the consumer test and replayed against the real Express app in the same job.

    CI job ↗
  4. 04

    Integration with PostgreSQL

    Real database semantics only where they matter: a disposable postgres:16 proves the CHECK and primary-key constraints and the full HTTP-to-database path.

    CI job ↗
  5. 05

    Browser E2E and accessibility

    Only the critical journey, verified through the API rather than by trusting the text, two failure paths mocked at the network layer, and axe-core WCAG scans before and after interaction.

    Report ↗
Supporting evidence · breadth, on their own cadence
  1. 06

    Performance

    k6 protocol test against the lab booted inside the runner, gated on p95 and error-rate thresholds. Weekly and on demand, never a per-commit gate.

    1,000 requests · p95 2.22 ms · 0% failed · thresholds met · run #2 on 2026-09-14 · CI-local workload, indicative, not a production benchmark

    Summary ↗
  2. 07

    Observability

    The instrumented server is started during publishing, three requests are sent, and the exported OpenTelemetry spans are kept as evidence.

    12 spans exported by Node auto-instrumentation for three requests (health, checkout, order lookup)

    Spans ↗
  3. 08

    AI evaluation

    Promptfoo checks a fictional support assistant for valid JSON, mandatory escalation and refusal of data requests. On demand with an Anthropic or OpenAI key; the assertions are provider-agnostic.

    3 assertions passed · 0 failed · anthropic:messages:claude-opus-5 · run #2 on 2026-09-15

    Runs ↗
Also in the repository

Test strategy · Decision records · CI runs · TestPlan Studio, which uses this lab as its demo target