This section refines the quality goals
into scenarios. Every scenario names the guarantee the code makes today and the test
or check that would fail if it broke. A scenario without a guard is listed under
risks instead.
flowchart LR
q["Kionga quality"]
q --> sec["Security and isolation"]
q --> truth["Truthful state"]
q --> ops["Operability"]
q --> open["Openness and portability"]
q --> repro["Reproducibility"]
sec --> s1["Default deny on every request"]
sec --> s2["Fail closed on unsafe production config"]
sec --> s3["Throttled local sign-in"]
sec --> s4["Project isolation for normal users"]
sec --> s5["No secrets in history or images"]
truth --> t1["Atomic change, audit and outbox"]
truth --> t2["Engine failure recorded as failure"]
truth --> t3["At-least-once lifecycle delivery"]
truth --> t4["Schedules fire at most once per slot"]
ops --> o1["Stack resumes after a restart"]
ops --> o2["Gateway latency under load"]
ops --> o3["No lost writes under concurrency"]
ops --> o4["Local ports on loopback"]
open --> p1["Excluded proprietary products"]
open --> p2["Standard interfaces"]
repro --> r1["Digest-pinned server and cluster images"]
repro --> r2["Identical ML pins"]
repro --> r3["Network-free gate tests"]
A credential is committed, or a dependency gains a reachable vulnerability
CI fails
gitleaks, govulncheck, gosec, pip-audit and bandit in the shared workflows (CI gates)
S8
A vendored browser file changes without its hash
The UI test fails
platform/tests/ui/vendor-integrity.test.cjs
S9
The Helm chart renders
Every Deployment is non-root with a read-only root file system and no capabilities, every Service is ClusterIP, no cluster-wide RBAC or Secret is rendered
test_managed_chart_security_and_availability in deploy/tests/test_production_deploy.py
A model trained in a notebook or pipeline is served
It loads, because MLflow and scikit-learn match across images
Pins in pipeline-runner/requirements.txt, deploy/mlflow/Dockerfile, workspace/jupyter/Dockerfile; test_training_images_match_the_server in deploy/tests/test_mlflow_server.py
R3
Gate tests run without network access or API keys
They pass: the mock LLM and hashing embedding function make them deterministic
sdk-python/tests, agent-runtime/tests; the paid evaluation lane runs only with MLAIOPS_RUN_EVALS=1
Throughput and latency of the agent runtime, the trace proxy and model containers have
no thresholds. The Kubernetes pipeline executor, log export and external S3
connections are tested against fakes only. See the
implementation status.