Production and Evidence Method
This edition was assembled and revised with extensive AI-agent assistance operating against pinned repositories, source code, tests, project records, and generated verification artifacts. Agents helped inventory APIs, compare claims across chapters, inspect implementation paths, draft and edit prose, generate structured diagrams, run validators, and challenge the resulting manuscript. AI output was not treated as technical authority. The cited repository state, external specification or implementation, test oracle, and retained artifact determine each claim.
The edition is deliberately revision-bounded. CNA behavior is described at v0.1.0-alpha.1 @ 1bb2145d99ed572dd4eb15009c34e2e5f410fcf0, while sibling behavior is described at the pins in Appendix D; executed results additionally name their configuration and host. Automated source scans, builds, PDF checks, and adversarial review passes reduce drift but cannot make a changing codebase timeless. Later editions should correct stale statements when the repositories or their evidence change. Responsibility for the published selection, wording, and claims remains with the named human author and editor.
Evidence labels used in this edition
| Label | Meaning |
|---|---|
| Source-proven | A reachable implementation, fallback, refusal, or absence was established from the pinned source. No successful build or execution is implied. |
| Compile-proven | Names, signatures, includes, templates, and required link closure were exercised in the named configuration. |
| Runtime-observed | The named executable path ran and produced the recorded state, value, file, trace, or lifecycle event. |
| Renderer-engaged | The intended renderer family and native or translation route were shown to execute, rather than a substitute or early skip path. |
| Oracle-compared | The observation was compared with an identified authority under a stated normalization or tolerance policy. |
| Pixel-verified | Selected pixels or a complete frame were compared at the stated readback boundary, renderer, host, and tolerance. |
| Manually observed | A human inspected the result, but no retained discriminating oracle supplied an automated verdict. |
| Historically recorded | A dated project artifact reports the result; this edition did not rerun it. |
| Blocked / unsupported / not attempted | Three distinct outcomes: an external precondition prevented the attempt, the named path has no implementation, or no attempt was made. |
These labels are coordinates, not a single quality ladder. A compile probe can answer an API question that a screenshot cannot; a hostile-input test and a pixel oracle exercise different contracts. Appendix G gives the full evidence vector and preferred authority for each kind of claim.
Repository paths and revisions serve as the primary references, so the book does not maintain a separate bibliography database. External links are few and reviewed. Source locations describe the pin and may move later; names, revision identifiers, and the repository map provide the durable reproduction route.
PDF accessibility boundary
The source supplies a consistent heading hierarchy, document language and metadata, bookmarks, selectable text, Unicode mappings, and descriptions for figures. The current pdfLaTeX pipeline does not produce a fully tagged PDF, so this edition does not claim PDF/UA conformance. Moving to a tagged-PDF toolchain remains publishing work for a later release and requires a separate structural and assistive-technology review.