Test Strategy

Three gate layers

Archon's verification surface is deliberately layered so different failure modes are caught by different mechanisms.

L1 — Portable governance contract

Consumed from .archon/contracts/governance-contract.yaml by two equivalent checkers:

• scripts/archon-check.py — Python stdlib-only (runs on any CI image with Python 3).
• scripts/archon-check.sh — Bash port.

What it verifies:

• File shape — required headings / sections exist in soul.md, manifest.md, decisions.md, etc.
• Cross-references — every <file> §<heading> link resolves mechanically.
• Line caps — .archon/soul.md, .archon/manifest.md, decisions.md, drift / debt hot indices obey their declared budget.
• Archive hygiene — drift / debt / memos archive headers carry the expected row counts.
• Forbidden substrings — host-project terms do not leak into universal modules (Universal Module Guard).
• Preservation Axis probes (ADR-28) — load-bearing headings keep their body shape.

L2 — Export manifest round-trip

scripts/test-archon-export.mjs runs a dry export for every platform target and asserts:

• Every image referenced by a bundled markdown is listed in DOC_ASSET_FILES.
• Platform path rewrites (.cursor/ → .claude/, .mdc → .md) round-trip cleanly.
• Required core files are present.

This gate catches the most common export regression: adding a new comic to architecture.md without updating the export manifest.

L3 — Project validate pipeline

The adopter declares this in .archon/manifest.md § Validation Command. Example (Node stack):

npm run validate  # lint + typecheck + test + bundle-budget

Archon does not dictate what goes here — it only insists that some validation command exists and passes before close-out.

Gate composition

The three layers compose via the npm run validate entry on the authoring repo:

validate
├── archon:records:check   # records-folder fold integrity (ADR-22)
├── archon:check           # archon-check.py (L1)
├── archon:verify          # claim verifier (ADR-27)
├── archon:export:test     # test-archon-export.mjs (L2)
└── <project-specific>     # eslint + tsc + vitest + bundle budget (L3)

Every pre-commit and pre-push hook runs the same chain (via .husky/).

What is not tested mechanically

Intentional gaps:

• Design quality — visual polish, copy tone. These use Plan-mode human review, not CI.
• ADR trade-off quality — a rejected alternative can be technically well-framed yet strategically wrong. Captured in ADR-N rows (negative ADRs) + re-evaluation conditions.
• Domain correctness in user code — Archon governs the governance, not the product logic. The product team owns product tests.

The line is: if a mistake can be mechanized, it must be. If it cannot, the governance surface is upgraded until it can (Lint-Rule Bridge — ADR-20).