Latest summary
Built and verified the SE cycle-2 non-toy population review bundle at /home/synthestat/output/runs/SE/se_population_review_cycle2_3a9d999a_seed420987. It is a contract-complete, explicitly labelled stratified multi-DeSO candidate (42 DeSO, 91,030 persons, 43,739 households/dwellings) with HARD selected-zone age-sex residual exact and unavailable building/hidden/work-school layers surfaced rather than inferred.
Body
Country: SE (Sweden)
Project root / allowed write root: /home/synthestat
Parent manager task: t_0c611b3b
Depends on source freeze task: t_1bbf9f63
Prior cycle:
- Cycle-1 bundle: /home/synthestat/output/runs/SE/se_population_review_cycle1_a5ad12d7_seed420987
- Reviewer verdict t_33ff07f7: NEEDS_MORE_SOURCES because output was an 8-person/two-test-zone seeded fixture with missing hidden/work-school/building/source grounding.
- Local methodology ingest t_20de640c: /home/synthestat/workspace/manager_handoffs/SE_other_synthesis_ingest.md
Mission:
Build the next SE population review bundle using the frozen/catalogued sources from t_1bbf9f63 and the ingested other_synthesis methodology. Replace the 8-person toy fixture with the strongest responsible Sweden candidate available: full-national if feasible, otherwise a clearly labelled stratified multi-DeSO slice large enough to stress household type, age-sex, origin/background, school/work, and building assignment.
Mandatory modelling guardrails:
- Reuse Synthestat modules/registry contracts; do not copy country-specific other_synthesis logic wholesale or create a parallel Sweden-only synthesis engine.
- HARD constraints never break. FIRM/SOFT/GUIDE precedence must be explicit.
- Every modelled/weakly measured estimate needs uncertainty bounds or an explicit unavailable reason.
- Hidden-population overlays remain separate and must not silently rewrite de jure population constraints.
- Fine-geography sparse attributes and assignment layers must be flagged as measured/constrained/modelled/unknown.
- Preserve provenance: source IDs, retrieval timestamps where available, geography, reference period, quality flags, fallback/relaxation records.
Required improvements from cycle 1:
1. Stop presenting the seeded fixture as a candidate for national SE review; build full-national or labelled stratified multi-DeSO slice.
2. Implement/read Sweden target registry/fallback metadata from frozen SCB artifacts: source ID/table ID, geography, reference period, classification, target tier, tolerance, DeSO -> municipality -> county -> national fallback, suppression/null handling.
3. Adopt household/home-first structure where feasible: households to homes/buildings/addresses first, then dwellings/building assignments; carry fallback/source fields for each assignment.
4. Add household/family graph realism fields where evidence supports them, with SOFT/GUIDE uncertainty labels.
5. Add origin/background assignment with official SCB categories preserved and synthesis-oriented taxonomies labelled as model taxonomy.
6. Emit work_school_assignments.parquet if evidence supports assignments; otherwise emit explicit partial/unavailable artifacts with source/fallback status.
7. Fix or document the prior synthetic_dwellings.household_id backlink issue.
8. Align geography_quality_tiers, zones_degraded, and model_notes so seeded/sliced/degraded zones cannot overclaim A/well_constrained status.
9. Include SOFT residual summaries when SOFT constraints are in the registry, or explicitly mark not exercised.
Deliverable:
A contract-complete population review bundle under /home/synthestat/output/runs/SE/<new_run_id>/ following docs/contracts/population_review_bundle.md, plus modeller handoffs under /home/synthestat/workspace/manager_handoffs/modeller/. Complete with metadata containing bundle_path, run_id, key limitations, files written, tests/checks run, and any blockers.
If t_1bbf9f63 proves P0 sources are insufficient for a non-toy model pass, do not hallucinate a production run: produce the best valid partial/slice bundle or block with exact missing source requirements, following no-silent-degradation.