Run 001 - deterministic composition mechanics

Run date: 2026-07-13
Environment: Windows, Python 3.14.0, Node 22.21.1
Raw output: .cache/results/harborworks-composition/run-001/ (local only)

Reproduction

From the repository root:

python .cache\src\harborworks-composition\build_fixtures.py
python .cache\src\harborworks-composition\test_composer.py -v
python .cache\src\harborworks-composition\compose.py run --root memory\2026-07\13\harborworks-composition-experiment\fixtures\repos\harborworks-operations-workbench --query memory\2026-07\13\harborworks-composition-experiment\fixtures\scenarios\prefilter-query.json --output .cache\results\harborworks-composition\run-001
python .cache\src\harborworks-composition\compose.py compare-order --first memory\2026-07\13\harborworks-composition-experiment\fixtures\scenarios\order-a --second memory\2026-07\13\harborworks-composition-experiment\fixtures\scenarios\order-b
python .cache\src\harborworks-composition\compose.py expect-failure --root memory\2026-07\13\harborworks-composition-experiment\fixtures\scenarios\revision-conflict
python .cache\src\harborworks-composition\compose.py expect-failure --root memory\2026-07\13\harborworks-composition-experiment\fixtures\scenarios\policy-conflict

Observations

The repository filter removed 2 of 9 repositories and 12 of 54 skills.

Artifact filtering removed 6 additional remediation skills contained in the three selected domain repositories.

The naive physical overlay retained 54 relative paths and performed 27 overwrites while processing 81 logical artifacts.

The logical catalog retained unique identities for all 81 artifacts.

No .cache/ path was exported.

Reversing import declaration order returned identical sorted identity sets: 72 artifacts in each scenario.

The incompatible core revisions failed with: revision conflict for org.harborworks.context.core: core-r1 vs core-r2.

The read-only root importing the executable briefing context failed with: root does not allow ['execute'] required by org.harborworks.methods.briefing.

All nine dependency-free CLIs returned valid JSON identifying their context, revision, requested skill, and knowledge path.

Six focused tests passed, covering full-catalog retention, scope filtering, import order, revision conflict, policy conflict, and skill-to-owned-surface routing.

Interpretation boundary

These results establish deterministic behavior for the synthetic corpus. They do not establish that repository prefiltering improves model selection, that the proposed labels generalize, or that the policy rule is sufficient for real tools. Those claims require separate experiments.