Simon AxelssonandClaude Sonnet 5
9740e20818
CI (SIAX Cloud) / security (push) Successful in 23s
Security Scan (SIAX Cloud) / security (push) Successful in 22s
CI (SIAX Cloud) / integration (push) Successful in 29s
CI (SIAX Cloud) / quality (push) Successful in 45s
CI (SIAX Cloud) / security (pull_request) Successful in 27s
CI (SIAX Cloud) / integration (pull_request) Successful in 31s
Security Scan (SIAX Cloud) / security (pull_request) Successful in 29s
CI (SIAX Cloud) / quality (pull_request) Successful in 48s
Continues P010 (branch renamed from feat/p010-source-registry-persistence):
closes the three remaining BEFORE_STATE.md gaps (§2 TS/Python split-brain,
§3 missing provenance fields, §4 untested failure modes), all additive to
the production pipeline — the Python QA filter's delivery decision is
unchanged by any of this.
TS/Python bridge (§2):
- New route surface app/api/0pt/quality/* (rules/runs/issues CRUD +
POST /api/0pt/quality/bridge-run) over the shared QualityService
singleton (src/services/quality.ts), fixing the same "no shared
instance" defect the Source Registry had.
- src/quality/quality-bridge.ts auto-provisions a deterministic,
tenant-scoped baseline rule set on first use so the Python engine never
pre-registers rules through the CRUD surface.
- services/api/workers/quality_bridge.py: the Python-side HTTP client.
Feature-flagged off by default (OPT_QUALITY_BRIDGE_ENABLED), fails open
on any error (never raises into run_job, never changes `delivered`).
Advisory only — real per-row gating against the TS engine isn't
implementable yet (ExpectationEngine returns dataset-level pass/fail
counts + samples, not a per-row boolean array); documented, not faked.
- Live-verified end-to-end: `next dev` (AUTH_ENFORCE=false) + the real
Python pipeline hitting the real brreg.no API, round-tripping through
the real bridge route, receiving a real QualityService score.
Provenance (§3): sources/connectors.py's Provenance dataclass gains
retrieved_at, transform, connector_version, producer, quality,
entity_resolution, artifact_ref (all previously absent from the real
dataflow entirely) via a make_provenance() factory used by every connector.
resolve.py stamps entity_resolution on merged records' provenance (never on
unmerged ones). pipeline.py stamps artifact_ref at delivery and quality
once the bridge succeeds.
Failure tests (§4), real fault injection against real code (services/api,
new pytest + respx suite, not vitest mocks):
- dedup-determinism + duplicate-input + out-of-order: resolve.py under
shuffled/interleaved/exhaustive-permutation input, plus an explicit,
verified test of the one case that IS legitimately order-sensitive
(non-transitive similarity in a greedy clusterer) instead of hiding it.
- schema-evolution: respx-mocked upstream responses with added/removed/
alien-shaped fields through the real brreg connector + pipeline.
- restart: connector exceptions mid-job (including one from respx itself)
never crash run_job.
- backfill: the real property that holds given no persistence layer
exists (replay is compute-idempotent) — NOT a fabricated dedup-store
test.
- CDC-interruption and true backfill/delivery-store idempotency:
explicitly documented as not implementable against this codebase
(no CDC pipeline, no delivery store exist anywhere) via skip markers
with reasons, in tests/test_known_gaps.py, rather than silently
dropped or faked.
Testing: 256/256 vitest (249 pre-existing + 7 new bridge route/service
tests) and 28/28 pytest (2 documented skips) green locally. tsc -p
tsconfig.check.json has zero new errors (14 pre-existing errors, all in
the untracked/unaudited src/fleet, src/graph, src/metrics, src/projections
paths this pass does not touch, per BEFORE_STATE.md).
No PR opened, per instructions.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
0pt data engine (services/api)
The legal data motor (docs/06). FastAPI orchestrator + worker pipeline that turn
compliant sources into delivered, provenance-tagged data. Runs on a VPS via
../../infra/docker-compose.yml — not on Vercel (long jobs).
Sourcing boundary (CLAUDE.md §1 — non-negotiable)
Per field × market, in strict order: official API/bulk → authorized public web (robots + rate limits) → licensed partner data → otherwise level 0 "ej tillgänglig". We never invent values.
Deliberately excluded (never add here): WAF/bot-detection bypass, FlareSolverr, stealth/undetected browsers, proxy/fingerprint rotation to evade bans, scraping behind login walls, automated fake-account creation. This is the honest-accuracy moat.
Layout
sources/connectors.py— compliant connectors (GLEIF, Brønnøysund, Companies House live; PRH/CVR/Bolagsverket gated on verified terms → level 0 until wired).sources/sources.yaml— source registry (access/licence/cadence; VERIFY terms).orchestrator/main.py— REST API (health, sources, estimates, orders, jobs) per docs/04.workers/adaptive_router.py— picks the cheapest compliant method per market (api > bulk > authorized_web > licensed). No bypass tiers.workers/pipeline.py— sourcing → extraction → structuring (dedup) → enrichment/fusion → QA (drops rows below the order's confidence target). Writes provenance per field.
Run locally
pip install -r requirements.txt
uvicorn orchestrator.main:app --reload # API on :8000
python -m workers.run # run one demo job
Reconciliation note
Supersedes the earlier nordic-eu-data-platform backend: the compliant connectors +
pipeline are migrated here; the FlareSolverr/stealth/proxy/fake-account paths and the
test_bypass_chain/test_proxy_pool examples are intentionally not carried over.