Skip to content

Reference — Claim matrix

Generated from the maintainer-curated claim registry in scripts/generate_reference_docs.py: every public claim carries its maturity, implementation location, automated evidence, and known limitation. Documentation must never describe an experimental, optional, or owner action surface as universally operational.

Public claim Maturity Implementation location Automated evidence Known limitation
CLI lifecycle commands stable src/knovaryn/interfaces/cli/ tests/test_cli_conformance.py (tree snapshot + docs scan) Local run drives one job in-process; durable multi-worker mode uses worker.
REST API + web console stable src/knovaryn/interfaces/rest/ tests/test_rest_api.py, tests/test_rest_webconsole.py Bearer-token auth only when KNOVARYN_API_TOKEN is set.
Python SDK (Workspace) stable src/knovaryn/application/workspace.py tests/test_workspace_control.py APIs may still gain parameters within the 0.x line.
MCP stdio server experimental src/knovaryn/interfaces/mcp/server.py tests/mcp/, scripts/mcp_acceptance_matrix.py (4 cells: mcp 1.28.0–2.1.0), ADR-0007 Tool surface may be reshaped before 1.0.
MCP streamable HTTP (authenticated) experimental src/knovaryn/interfaces/mcp/server.py scripts/mcp_acceptance_matrix.py (both SDK majors, 4 pins), ADR-0007 Deploy behind the bearer-token guard; not exposed by default.
Provenance lineage + location precision stable src/knovaryn/application/, src/knovaryn/pipeline/export/gate.py tests/provenance/ Precision reflects the parser: markdown/text caps at section/chunk.
Docling parsing (bounding-box spans) optional src/knovaryn/infrastructure/docling/ tests marked docling; ADR-0003 resource guard Requires the docling extra and its runtime dependencies.
Offline semantic validation (offline-fast) stable src/knovaryn/pipeline/quality/validators.py tests/semantic/ Deterministic checks catch specific contradiction classes only.
Certified semantic validation (certified-semantic) optional src/knovaryn/pipeline/quality/profiles.py tests/semantic/ benchmark corpus Needs a reachable judge provider; unavailable judge ⇒ unverified.
Preference pairs + certified-pairwise profile alpha src/knovaryn/pipeline/quality/validators.py tests/preference/ Judge-based certification inherits provider limitations.
Information-gain / preference-signal gates alpha src/knovaryn/pipeline/quality/ tests/test_validators.py Heuristic scoring; not a quality guarantee.
Durable jobs (claim/heartbeat/checkpoint/resume) stable src/knovaryn/pipeline/jobs/, src/knovaryn/infrastructure/database/ tests/chaos/, tests/deployment/ SQLite mode is single-writer; concurrent workers need PostgreSQL.
PostgreSQL multi-worker queue optional src/knovaryn/infrastructure/database/repositories.py (SKIP LOCKED) tests/deployment/test_compose_e2e.py claim-exclusivity race Requires a deployed Postgres; SQLite dialect ignores FOR UPDATE.
S3-compatible artifact storage optional src/knovaryn/infrastructure/artifacts/s3.py tests marked s3 Requires the s3 extra and object-store credentials.
Hugging Face publishing optional src/knovaryn/infrastructure/publish/hf.py publication-gate tests (dry-run default) Never automatic; license/privacy gate plus explicit confirmation.
PyPI installation stable .github/workflows/publish.yml (OIDC trusted publishing) scripts/package_ci.py clean-env battery First-time setup of PyPI trusted publisher is an owner action.
Release integrity (checksums, SBOM, notes) stable scripts/release_governance.py, src/knovaryn/pipeline/export/verify_release.py tests/test_release_governance.py, tests/test_release_integrity.py GitHub-side settings (labels, environments) are owner actions.
Docker Compose topology stable deploy/docker/ .github/workflows/deploy-e2e.yml exact-SHA suite Weekly/dispatch execution; always runs at release time.
Kubernetes manifests alpha deploy/kubernetes/ kustomize build in CI lint path No automated cluster E2E yet; backup/restore documented as runbook.
Documentation site & search metadata stable mkdocs.yml, overrides/, docs/ ci.yml docs job (strict build + links + assets) Search-engine indexing itself is an owner action (§7.2).
Benchmarks (canonical methodology + current report) alpha benchmarks/run_report.py, docs/reference/benchmark-methodology.md benchmarks/results/0.2.1/ (3 runs, checksums, H(R1)=H(R2)=H(R3) digest equality) Offline fake-provider framework metrics ≠ generation throughput or model quality; live-provider evidence not measured.
Release-bundle reproducibility (reproducible mode) alpha src/knovaryn/domain/ids.py (seeded IdGenerator), benchmarks/run_report.py tests/test_reproducible_release_digest.py; results/0.2.1/aggregate.json digest equality Requires seeded IDs (reproducible mode); production IDs stay unguessable UUIDv7s.