docs(accuracy): retain phase 2 provenance evidence

This commit is contained in:
Jens
2026-08-01 23:46:37 +02:00
parent 5b3c17b494
commit 41f8a54236
7 changed files with 1908 additions and 25 deletions
+53
View File
@@ -12308,3 +12308,56 @@ Open:
`docs/accuracy-program/06-implementation-roadmap.md`: fix CRS/authority/
coverage/lineage and protected-test isolation, complete human corpus review,
rebuild independent data, freeze metrics, then train on `cuda:0`.
## 2026-08-01 - Accuracy Improvement Program Phase 2 source and provenance foundation
### Implemented and verified scope
- Added a server-owned source registry, immutable snapshots, source policies,
dataset/dataset-version provenance fields, lineage edges and durable
quarantine records through Alembic revision `202608010001`.
- Added versioned vector, raster, YOLO label and PyTorch model contracts with
checksum, CRS, bounds, units, resolution, topology, required attributes,
time/freshness and lineage checks. Governed ingestion is idempotent through
ingest keys, can safely replay an identical immutable snapshot across
projects without rewriting its first fetch time, and fails closed into
quarantine when a contract fails.
- Non-WGS84 vector imports now store the exact transformed EPSG:4326 bytes as
the checksummed consumption artifact; original source bytes are retained
only as provenance evidence. Vector operations verify both storage CRS and
byte checksum before reading a dataset.
- Enforced a consumption boundary for production inference, primary reference
QA, derived work, export, coverage and training inputs. GRB is task-bounded
primary authority for suitable Flemish building validation; the Buildings
Register, DHMV, Sentinel-2 and OSM retain distinct roles, and OSM cannot
become automatic truth.
- Production detection/segmentation now resolves model sidecar UUIDs against
the registry and snapshot database records before adapter loading, including
source/snapshot relationship, freshness, quarantine, checksum and version
checks. A structural sidecar alone remains catalogue/preflight evidence.
- Verified the migration in a disposable PostGIS database: upgrade, real
trigger-guard DML checks and downgrade passed. This was not a production
migration or deployment.
### Evidence and test result
- Retained proof is under `artifacts/evidence/accuracy/P2/`: the source/
contract inventory contains 40 server-owned definitions and the PostGIS
guard report covers immutable source/snapshot evidence, exact checksum
binding, lineage cycles and transitive quarantine propagation.
- A final cohesive source-to-training regression set passed 221 tests,
including immutable snapshot replay, canonical CRS/checksum storage and
database-bound model provenance. The governed golden QA benchmark passed.
- Frontend typecheck, 51 frontend unit tests and production build passed.
The complete backend suite now collects but is not green: 1,282 passed and
17 failed. Sixteen failures are historical source-text assertions; one is
an intermittent Windows WSL-backed `bash.exe` host failure in a syntax test
that passes in isolation. Repository Ruff still has 95 findings and
`npm run lint` is still absent.
### Decision
- The source/provenance foundation is implemented and verified in a
disposable environment, but Phase 2 remains **in progress** and Phase 3 is
**not ready**. No training, protected-test release, promotion, national
validation or production migration is authorized by this result.