Curation versioning
Anything that's effectively a curation decision — umbrella memberships, vocabulary aliases, status maps, biomarker umbrellas, low-info inclusion tags, drug aliases, city/state/country aliases, the stale-trial-years threshold — lives in curation_versions as snapshot rows. Edited via /admin/curation/*.
A doctor or curator can fix an umbrella membership or add an alias without a code deploy. The database is the source of truth.
The 9 domains
| Domain | Shape | Used by |
|---|---|---|
umbrella_members | {umbrella: [primary, …]} | cancer_site umbrella mapping |
cancer_site_aliases | {alias: canonical} | Free-form site normalisation |
drug_class_aliases | {alias: canonical} | Drug-class normalisation |
low_info_inclusion_tags | [INC_DEMOGRAPHICS, …] | Tags hidden from the Advanced Search sidebar |
ctri_status_map | {source: canonical} | CTRI status normalisation in transform |
biomarker_umbrella | {key: [{dimension, tag}, …]} | Cross-dimension biomarker matching (TRIPLE_NEG ⇄ triple_negative, HER2_POS, etc.) |
stale_trial_threshold | {years: N} | The recency-filter cutoff (default 7) |
drug_aliases | {generic: {brand_name, drug_class}} | Drug-picker normalisation |
city_state_aliases | {cities, states, countries: {alias: canonical}} | Location autocomplete + bidirectional alias expansion |
How a save works
Every save inserts a new row with an auto-incremented version_number. The newly-saved version becomes is_active=TRUE; the prior active version is demoted. A unique partial index enforces "exactly one active version per domain."
Reads everywhere (src/lib/curation.js for JS, scraper/curation_loader.py for Python) hit a 5-minute in-process cache. A save busts the cache so the next request sees the change.
Every save also writes a manual_dispatches audit row capturing action='curation_update', target domain, actor (from x-admin-user header if supplied), and payload.
Rollback
Use /admin/curation/[domain]/rollback to re-activate an older version. The system inserts a new row pointing at the same payload as the old version (so the audit trail stays linear) and flips active.