Target Geographies

Where an intervention should go, who lives there, and what reaching them would cost. This is the core query loop of the targeting-analyst agent; the judgement that surrounds it — vetting a new source, defending the number, producing the artifacts — lives in its sibling skills, and this one hands off rather than half-doing them.

The data and arithmetic are in the Connect Labs targeting_* MCP tools. What this skill carries is what those tools cannot: which question to ask, and how to read the answer without overclaiming.

Inputs

Every row except the optional Phase 1 PDD is an explicit argument or a live labs read; no run is needed.

Source	Artifact	Used for
The request	indicator, threshold(s), country scope	the query itself
Connect Labs MCP	targeting_research	what is already known about this indicator's sources and traps
Connect Labs MCP	targeting_indicators	indicator family, unit, which methods can answer
Connect Labs MCP	targeting_select	totals, ranked areas, honesty fields
Connect Labs MCP	targeting_scenario	cost basis and total
Phase 1 (optional)	1-design/idea-to-pdd.md	intervention, unit cost and pricing basis

Products

• <phase>/target-geographies_summary.md — the selection, its totals, and every caveat the honesty fields raised
• A handoff to defend-a-figure before any figure is published

Process

1. Read what is already known. targeting_research for the indicator, first and always. It returns prior findings with each claim re-run against live data. Read trust before using anything: holds means the note describes reality now, drifted means re-derive whatever the failed checks cover, unverified means the note is a lead rather than a finding. If rescan_due is set, say so and ask the user whether to run scan-data-sources — the checks confirm what we found, they cannot tell you whether something better has since been published.
2. Establish the indicator's family and unit. Call targeting_indicators. A burden measure (mortality, stunting, malaria incidence) is worse when high and selects places above the threshold. A coverage measure (sanitation, ORS, immunisation, ITN use, effective treatment) is worse when low, selects below, and the fundable quantity is the unreached count, not the rate. More than half of the 29 targetable indicators are coverage measures — never assume. Units are not interchangeable either: mortality is per 1,000 live births, so 8% is a threshold of 80, while 50% sanitation is 50.
3. Check which methods can answer it. targeting_indicators lists per-method country counts. IGME publishes mortality only; MAP's surfaces answer malaria only; DHS answers what it surveyed. A method showing 0 countries returns an empty answer, which reads exactly like "no burden here".
4. Run targeting_select for each threshold of interest, scoping with iso_codes when the question names countries.
5. Read the honesty fields before quoting any total. coverage says how many selected units carry each count — short of the total means the figure is a floor, not a measurement, and must be worded that way. inherited_units counts units whose rate was measured somewhere coarser — usually their country — and applied here; a selection that is mostly inherited is a national figure repeated across regions, not a subnational finding. countries_unsupported names who was left out rather than answered at another level.
6. Cost it with targeting_scenario when money is in the question. A unit price is meaningless without a unit of measure, and the basis is a property of the programme: KMC per newborn, a bednet per child, a water connection per household, an antimalarial per case. Where an indicator implies no case count the per-case basis is refused rather than approximated — pick another basis.
7. Hand off before publishing. defend-a-figure for anything a funder will read; build-targeting-model for a Doc and model. Do not shortcut either.
8. Write back anything new you learned with targeting_research_write, especially a trap or a rejected source. The next session should not have to find it again.

MCP Tools Used

connect_labs (remote HTTP, PAT auth)

• targeting_research — built
• targeting_research_write — built
• targeting_indicators — built
• targeting_select — built
• targeting_scenario — built
• targeting_admin_levels — built

What has counts, and why it matters

Most indicators here are rates, and a rate cannot answer "how many cases would we be treating" — the per-case cost basis is refused rather than faked. Four families escape that, and they are what to reach for when a programme has to be sized rather than merely located:

Count	From	Answers
malaria_cases, malaria_deaths	MAP 5 km surfaces, annual to 2024	how much disease, per year
antimalarial_effective_gap	cases x (1 - effective treatment)	untreated cases — its denominator is a case count, not a population
pop_beyond_2h	Weiss et al. travel time x WorldPop	people further than two hours' walk from care
pop_rural	DEGURBA x WorldPop	rural population, by the UN-comparable definition

Continental figures, for scale: 241.9M malaria cases and 116.9M of them untreated; 242.9M people beyond two hours' walk; 745.2M rural.

Ranking countries against each other

"Which countries should be tier 1?" is a different question from "which regions cross a threshold?", and the tools answer it unevenly. Know this before you start, or you will spend a dozen calls finding it out (ace#2560).

• resolution: national answers only what a national method's sources carry. national_igme is u5mr, imr and nmr for 54 countries (IGME 2024); national_modelled is lbw_rate; national_surface is the gridded MAP/access/DEGURBA layers. Ask ors_coverage, exclusive_breastfeeding or any other survey indicator at national resolution and you get empty_because_unanswerable: true — which is not a finding.
• Survey indicators come from method: subnational_survey, which rolls a country up to one ADM0 row only when every region clears the threshold. A country whose regions straddle the cut stays split, even at a threshold above threshold_max; read countries_partly_above to see which. On 2026-09-30, for ORS, that list held Nigeria, DR Congo, Ethiopia, Mali, Tanzania and fourteen others. Quote a range across regions for those countries, labelled as such; never average regions into a national figure yourself — the weights you would need are not in the response.
• Rank neonatal mortality with method: national_igme, never the survey route. Since connect-labs#2116 it answers nmr for 54 countries. subnational_survey still returns UNICEF's 2021 national value for only 24, and the two disagree by more than rounding (Zimbabwe 33.7 under IGME 2024 against 22.8 under UNICEF 2021), so do not mix them in one ranking. National ORS/EBF values are still missing (connect-labs#2117).
• Keep the calls small. A survey select returns up to limit region rows at ~400 tokens each. For a ranking, limit: 60 on the survey indicators is enough to see every whole-country row plus the split countries' regions.
• Name the scope. The data is the 55 African countries. A ranking request that does not say "Africa" may mean South Asia too — say it was not assessed rather than implying the list is complete.

What this cannot answer

• "How many villages?" Only where somebody has already drawn them: of 55 African countries, three reach village level — Rwanda (14,815 umudugudu), Madagascar (17,465 fokontany), Burundi (2,615 collines). Twenty-one stop at ward, twenty-five at district. Check targeting_admin_levels first. Detection from building footprints does not substitute: tested against Rwanda's register, the same buildings yield 68 or 2,403 clusters for one district depending only on the radius, and in Kigali one cluster absorbed 69% of the district. Where no register exists, reframe to "how many delivery units of ~N households" — well posed, and what a programme actually budgets.
• Rural vs urban, without naming the definition. DEGURBA classes 17% of Rwanda's villages as rural against a national figure near 72%. Both are defensible; a number quoted without its definition is not.
• Anything from IHME. Its non-commercial licence excludes for-profit entities and their employees and forbids re-hosting; nobody should register a healthdata.org account on a dimagi.com address. Report it as excluded on licence grounds, and note it may well be better — not that it is worse.

Each of these is recorded as a research note, so targeting_research returns the current version rather than this file's.

Mode Behavior

• Auto: run the queries, write the summary, hand off to defend-a-figure, notify the admin group with the headline figure and any honesty-field caveat.
• Review: present the selection, its totals and its caveats for approval before costing or handing off.

Change Log

Date	Change	Author
2026-08-29	Initial version — burden/coverage families, honesty fields, village and DEGURBA limits	ACE team
2026-08-29	Split defence and deliverables into sibling skills; research-note read is now step 1; 26 indicators; MAP counts	ACE team
2026-08-29	29 indicators; countable families now include access (pop_beyond_2h) and settlement (pop_rural), not malaria alone	ACE team
2026-09-01	off_method_units became inherited_units — a source a method does not declare is no longer used at all, so what remains to report is how much was inherited	ACE team
2026-09-30	Added § Ranking countries against each other — national resolution answers only u5mr/lbw/surfaces; survey indicators split into regions; nmr covers 24 countries (ace#2560)	ACE team
2026-09-30	nmr now answers nationally for 54 countries via national_igme (connect-labs#2116); use it, not the 24-country survey route	ACE team