Scan Data Sources

Finding out what could answer an indicator is the expensive half of targeting work, and almost none of it survives the session that did it. This skill makes the sweep repeatable and makes its output durable: a scan ends by writing a research note that the next session reads instead of starting again — including the sources it ruled out, which is the part nobody writes down and everybody re-derives.

A scan is not the same as a check. The stored checks on a note confirm that what we found still holds. Only a scan can tell you whether something better has been published since. Run one when a note reports rescan_due, when an indicator's coverage is embarrassing, or when a funder question depends on a figure the current source answers thinly.

Inputs

Source	Artifact	Used for
The request	indicator, or the gap that prompted the scan	what to sweep for
Connect Labs MCP	targeting_research	what was already ruled in or out, and when
Connect Labs MCP	targeting_indicators	current coverage and which methods answer
Connect Labs MCP	targeting_admin_levels	the geography the source would have to match

Products

• A research note written back via targeting_research_write, carrying the verdict, the alternatives table, and scanned_now: true
• <phase>/scan-data-sources_summary.md — the sweep, for a human reader
• Optionally a loader in connect_labs/labs/indicators/sources/, when a source is adopted

Process

1. Read the existing note before searching anything. targeting_research for the indicator. A prior scan's rejected list is the most valuable thing in it: re-litigating a source that was ruled out on licence wastes the scan. Note its scan_age_days — what you are doing is refreshing that clock.
2. State what would have to be true for a new source to win. Coverage of more units, a more recent vintage, a count where we only have a rate, or a licence that unblocks something. A scan without this becomes a list of links.
3. Sweep the four families that actually publish at this scale. Household surveys (DHS, MICS); UN and agency series (IGME, WHO GHO, UNICEF SDMX); modelled geospatial surfaces (MAP, WorldPop, GHSL/DEGURBA, Weiss et al. accessibility); and humanitarian aggregators (HDX HAPI). Most gaps are filled by the third family, because only a modelled surface has a value everywhere.
4. Check the licence before the data. This is the step that gets skipped and it is the one that invalidates the work. Find the actual licence page and quote it — do not infer a licence from a site being free to browse. A non-commercial licence (IHME, WHO GHO) means the source can serve as an external cross-check but must not be carried into a commercially-framed deliverable.
5. Test the access path for real before believing it exists. Fetch one country, or one boundary, and look at the numbers. An OGC endpoint that answers GetCapabilities may still refuse GetCoverage; an API may return a styled image instead of the data band. A source is not available until you have held a value from it in your hand.
6. Validate against something external. A new source's national total against a published figure — WHO, a national statistical office, the World Bank. Agreement within a few percent is the evidence that adoption is safe; a large gap is not automatically disqualifying, but it has to be named and carried forward into the methodology.
7. Decide the aggregation before adopting. A count sums; a rate needs a declared weight. A rate aggregated by area mean lets empty land vote and is frequently indefensible — if there is no population grid to weight by, say so and consider deferring rather than shipping a weak number that looks strong.
8. Write the note back with targeting_research_write and scanned_now: true. Include every source considered with its verdict (adopted, rejected, candidate) and the why. Give the note checks that would fail if the conclusion stopped holding.

MCP Tools Used

connect_labs (remote HTTP, PAT auth)

• targeting_research — built
• targeting_research_write — built
• targeting_indicators — built
• targeting_admin_levels — built

Writing checks that are worth having

A check earns its place by being able to fail. Four kinds exist:

• coverage — how many units carry the indicator at a level. Passes if coverage has since grown, because a backfill is not drift.
• value — the figure you reasoned from, for one country, within a tolerance.
• source — whether a named source still supplies the indicator at all.
• measure — whether the indicator still has the shape you assumed. This is the one that catches the nastiest drift: prose reasoning about a rate after somebody redefined it as a count.

Record the figure your argument rests on, not a figure that happens to be handy. If your note says "MAP's national total tracks WHO", the check is MAP's national total.

Verdicts, and what they oblige

• adopted — in use. Say how it is aggregated and what it was validated against.
• rejected — say whether it failed on licence, coverage, vintage, or quality, and never imply a licence rejection is a quality judgement. IHME is excluded because its terms exclude us, not because it is worse.
• candidate — good, and blocked on something nameable. Say what would unblock it, so the next scan can pick it up rather than rediscover it.

Mode Behavior

• Auto: run the sweep, write the note, notify the admin group with the verdict and any adopted source's validation gap.
• Review: present the alternatives table and the recommendation for approval before adopting a source or writing a loader.

Change Log

Date	Change	Author
2026-08-29	Initial version — sweep procedure, licence-first vetting, verdict vocabulary, check design	ACE team