The architecture is optimized for:
-
determinism
-
reproducibility
-
auditability
-
extraction safety
-
trust-boundary clarity
-
documentation navigability
These requirements matter more than feature count. A smaller but provable system is preferable to a broader system with ambiguous boundaries.
Quality Scenarios
| Quality | Scenario | Required behavior | Evidence |
|---|---|---|---|
Determinism |
The same committed source tree is bundled twice. |
Planning, ordering, render plan hashes, and checksums remain stable. |
|
Reproducibility |
A release package is rebuilt from the certified commit. |
Generated runtime output and release integrity metadata match the expected commit and version. |
release integrity checks, double-build reproducibility lane. |
Auditability |
A reviewer asks why a bundle was trusted or dirty. |
Manifest metadata records trust labels, dirty-state override evidence, note governance context, and derived review exports. |
manifest schema, bundle manifest, audit log where applicable. |
Extraction safety |
A downstream process extracts files from a bundle. |
Extraction follows manifest truth and verifies when requested; degraded recovery is explicit and never silent. |
|
Trust-boundary clarity |
An agent produces useful live reasoning. |
Track B output remains untrusted until Track A validation or verification proves an artifact. |
system contracts, MCP policy, Friday-to-Monday workflow. |
Documentation navigability |
A new operator needs to choose the right command without reading every architecture page. |
Manual pages provide entry points, while architecture pages explain the model and link to stable contracts. |
Antora navigation, docs assurance tests, architecture overview. |
Priority Rule
When a feature conflicts with a quality requirement, the quality requirement wins unless a documented, reviewed exception records the tradeoff. This is why dirty-state refusal, strict overlap handling, and latest-only schema publishing are allowed to be inconvenient: they keep the public contract honest.