Technical decisions about algorithms, architecture, data and execution, checked against documentation and commit history.
2026-08-09 · d96ccb0fad · Change verified; trade-off inferred
Unify document signals instead of duplicating the collection stack
- Before
- Demand expansion had separate source profiles, scope processing, extraction jobs and relational tables beside freight collection.
- After
- One source registry, retrieval run, extraction call and observations model cover freight, demand and explicit lead-time signals; structured indicators keep their artifact-specific path.
Share source collection, retrieval and observation processing instead of maintaining parallel document pipelines.
2026-08-10 · e3dfbdea07 · Documented change
A failed observation should not discard the whole document
- Before
- Unified extraction rejected an entire response if any proposed observation failed validation.
- After
- Keep observations that individually pass the same strict validators; discard nonconforming observations.
Whole-document rejection caused repeated retries that did not converge. Keep independently valid observations while retaining the strict validators.