Body
Orchestrate the Synthestat population synthesis QA loop for country: AD.
Project root: /home/synthestat
Target geography: finest available official geography
Release mode: internal research review
Goal:
Build the best possible 1:1 synthetic population for AD: persons in households, households in dwellings, dwellings in real houses/buildings where available. Include separate uncertainty-aware overlays for hidden or weakly measured populations where evidence supports them: homelessness, refugees/asylum seekers, Ukrainian displaced people, Syrian refugees, undocumented/seasonal populations, students, institutional populations. Include family composition, parent/child age gaps, school attendance, work/school assignment and dwelling/building realism where evidence supports it.
Mandatory process:
1. Create a synth-modeler task to generate a review bundle according to docs/contracts/population_review_bundle.md.
2. Create a synth-reviewer task dependent on the modeler task.
3. Branch only from the reviewer verdict:
- PASS: create final human review/delivery task and stop automation.
- NEEDS_MODEL_FIX: create modeler fix/rerun task, then reviewer re-review task.
- NEEDS_MORE_SOURCES: create marginal/distribution researcher tasks as needed; downloader freezes/catalogues approved sources; modeler reruns; reviewer re-reviews.
- BLOCKED_INVALID_OUTPUT: send back to modeler for valid bundle.
- EVIDENCE_EXHAUSTED_HUMAN_REVIEW: stop and ask human decision.
- MODEL_IMPROVEMENT_EXHAUSTED_HUMAN_REVIEW: stop and ask human decision.
Stopping rule:
Do not iterate forever. Stop on PASS, documented lack of source data, or model improvement plateau/modeller inability. After two materially similar failed cycles, require human review unless a concrete new source or model change is available.
Non-negotiables:
HARD constraints never break. Hidden-population overlays do not silently rewrite de jure constraints. Weak evidence means relaxed constraints/wider uncertainty, not fake precision. Every degraded zone, missing source, relaxation, and modelled estimate is explicit.