Counted but Unknown

Who are the particularly vulnerable that adaptation finance is for? And did they become less vulnerable because of it? These are the only two questions that matter to the people the money is named for, and the largest adaptation fund in the world can answer neither.

Adaptation finance is the stream of climate money that is about people. Mitigation is measured in carbon. Adaptation is measured, or is supposed to be measured, in the lives of those who bear a climate they did not make. The United Nations Framework Convention on Climate Change (Article 4.4), the Paris Agreement (Articles 7.5 and 9.4) and the Green Climate Fund's own Governing Instrument (paragraphs 2 and 52) all name the same people: the particularly vulnerable.

This paper holds the Fund to those words and delivers its verdict in the order the evidence demands. The good news comes first, stated as it is, because it is true and because funds that do the hard part well deserve to hear it. Then the bad news, which is simpler and worse: the Fund cannot say whom its money reaches, because it never defined who counts. Then the solution.

This paper makes three claims. First, across countries, the Green Climate Fund broadly honors its mandate: the most vulnerable countries are not shortchanged, and its fast track is fast. Reaching the most vulnerable countries is not the same as reaching the most vulnerable people. Second, inside countries, the record is blind by design: a beneficiary is anyone "likely to benefit", counted on assumptions, and the convention lets a project count the whole nation. Third, the remedy that the Fund's harmonized results framework proposes, a deduplicated record by sex, age and vulnerable group, can be delivered without building a register of the poor. Verify the count, not the person.

The analysis was registered and checksummed before any figure was compared with vulnerability, any delay was computed, or any rule was tested. The data are the Fund's open portfolio (372 approved projects, frozen 12 October 2026), the INFORM Risk Index (Mid 2026) as the primary measure of country vulnerability, the ND-GAIN index (2023) as a sensitivity, and World Bank population. Every deviation from the registration was logged; one, a more generous reading of the population test, was added after the first results were seen and is reported beside the stricter reading.

The country floor holds. Of US$10.36 billion in adaptation-weighted funding, 62.1 percent goes to Least Developed Countries, Small Island Developing States and African States, above the floor of 50 percent the Board aimed for in 2014 (Decision B.06/06, paragraph 1(b)).

The registered finding is defeated. The registration predicted that the most vulnerable countries would receive less adaptation funding per person than their less vulnerable neighbors, at one or more levels of a recursive tail. They do not. At no level, under either index, is the tail underfunded; under ND-GAIN, the most vulnerable fifth receives 2.57 times the per-person funding of the rest (90 percent interval 1.09 to 5.75). A registered prediction that fails is reported as a failure. This one failed in the Fund's favor.

The fast track is fast. The Governing Instrument promises "a streamlined programming and approval process to enable timely disbursement" (paragraph 53). Projects approved through the Simplified Approval Process reach their first disbursement in a median of 8.8 months; regular proposals take 14.1 (difference −5.4 months, 90 percent interval −7.3 to −2.9), in every run. And delay does not roll downhill: the most vulnerable third of countries waits 13.3 months, the least vulnerable 12.7, a difference indistinguishable from zero (+0.6, −2.0 to 2.8).

One caution belongs in the good news, because it is part of it. The two indices disagree about which countries are most vulnerable. Under ND-GAIN, which measures climate vulnerability, funding per person is unrelated to vulnerability (Spearman 0.05, −0.11 to 0.20). Under INFORM, which weighs displacement, conflict and food insecurity, the more vulnerable a country, the less it receives per person (−0.29, −0.43 to −0.15): a median of US$12.54 per person in the least vulnerable third, US$3.73 in the most. Climate-vulnerable countries get their share. Fragile ones may not.

The mandate names a who. The record holds a where and a how many. For each project the Fund's portfolio records the countries served and two numbers, direct and indirect beneficiaries. There is no field for the community, the group, the household, the gender, or the vulnerability of a single person reached. The particularly vulnerable are written into the treaty and written out of the data.

The count cannot be read as people, and the fault lies in the rule. Every one of the 288 adaptation and cross-cutting projects reports beneficiaries. Among the 216 single-country projects with population data, 18 score above 100 percent on the Fund's own adaptation indicator. That indicator is not this paper's construction. In 2014 the Board adopted, as its core indicator for adaptation, the "Total number of direct and indirect beneficiaries; number of beneficiaries relative to total population" (Decision B.07/04, paragraph (d)). The Fund adds the two counts and divides by the population; this paper only did the arithmetic the Board asked for.

The Board adopted the ratio and never defined its parts. Its integrated framework of 2021 keeps "direct and indirect beneficiaries reached" as a core indicator, counted in individuals, converted from households "based on average number of people per household" where individual data are missing, and concedes that "the same beneficiaries can be impacted by two outcomes", with the guidance to prevent it deferred to a results handbook. Nowhere does it say whether indirect beneficiaries exclude direct ones. So, before reading the 18 as overcounts, give every implementer the most generous reading. Allow that indirect beneficiaries may already include the direct ones, so that the claim is the larger of the two numbers rather than their sum. Allow that counts are lifetime projections, so that the ceiling is the highest population on record from 2020 to 2025 or the population grown to the project's end year, whichever is larger. Under that reading, 3 of the 216 exceed the population: two in the Marshall Islands, whose population has been falling by about 3 percent a year, and one in Barbados that exceeds it by less than 1 percent. Not one is evidence that anyone invented a beneficiary.

What the other 15 reveal is the convention. Ten projects count 90 percent or more of the national population as indirect beneficiaries: the whole country, by assumption. Direct beneficiaries are then added on top. The Fund's definition treats the two as distinct (direct beneficiaries are targeted, indirect ones are not), so the sum is exactly what its record invites, and its own Board paper calls the result "institutionally sanctioned double-counting". The same convention compounds across projects. In Barbados, the Federated States of Micronesia, the Marshall Islands and Tajikistan, direct beneficiaries summed over the Fund's single-country projects exceed the national population (213, 176, 172 and 110 percent). A person can benefit from four projects. Without a deduplicated record, nobody can say how many people that is.

Across the 289 projects that report direct beneficiaries, GCF funding per direct beneficiary runs from US$0.57 to US$143,182, with a median of US$114. At the bottom, a US$5 million grant to Tajikistan's hydrometeorological agency (FP075) claims 8,735,000 direct beneficiaries, more than 80 percent of the country's population, each reached with 57 cents. A forecast service may well serve a whole nation. But then the word beneficiary means a resident, and it cannot tell a funder whether a single herder in the Pamirs received a warning she could act on.

The money travels a long chain. Accredited entities with international access carry about four-fifths of adaptation funding to the most vulnerable third of countries (79.6 percent) and to the least vulnerable third alike (81.4 percent). The people at the far end of that chain are exactly those the record cannot see. In India, the Fund's adaptation money comes to US$0.15 per person; in Uganda, US$2.13; in Bolivia, US$7.83. International entities carry 25 percent of India's adaptation money, 80 percent of Uganda's and all of Bolivia's. None of their projects says who was reached.

The second question fares worse than the first. If the record cannot say who was reached, it cannot say whether they became less vulnerable. A beneficiary is counted when a project is designed, on assumptions, as a lifetime projection; nothing in the record returns to the same people to ask whether their exposure, their losses or their capacity to recover changed. Whether adaptation finance made anyone less vulnerable is, on this record, unknowable.

The blindness is written into the rules. The Fund's Harmonized Results Management Framework, prepared for the Board meeting of 26 to 29 October 2026 (GCF/B.46/11/Rev.01), states the definition and its gap in a single paragraph:

Direct beneficiaries "receive targeted support"; indirect beneficiaries "are not targeted but are likely to benefit, with their number usually estimated using conservative assumptions. The IRMF did not, however, specify how to establish whether support is targeted, or how substantial the support must be for a person to be counted." (paragraph 28)

The same paper finds that "vulnerability dimensions (gender, vulnerable groups, Indigenous Peoples) are not consistently disaggregated across existing indicators" (Finding 8); that the framework "produces institutionally sanctioned double-counting" (Findings 2 and 3); and that accredited entities name "double counting or ambiguity in indicators (43 per cent)" and "lack of clarity in indicator definitions (43 per cent)" among their top challenges.

Read the definition again. A person counts if she is likely to benefit, on assumptions, with no threshold of support and no rule for telling targeted from untargeted. A record built on that definition was built to count people, never to know them. It has no way to name the particularly vulnerable.

The remedy is on the table, and it is late. The paper proposes a "shared beneficiary record that distinguishes direct and indirect beneficiaries, female and male beneficiaries, age groups and vulnerable groups, with a deduplicated roll-up so that the same person is counted only once". Twelve years earlier, in 2014, the Board decided that results "should be disaggregated by gender where relevant" (Decision B.07/04, paragraph (i), 21 May 2014). Twelve years is a generation of adaptation projects designed, approved and closed without a record of whom they reached.

The obvious fix is the dangerous one. A fund that wants to know who its money reaches could demand names, identity numbers and household lists. That would put the poorest people on earth into a central register held by institutions they cannot hold to account, in countries where being on a list can mean being targeted. The people the mandate protects have a right not to be known.

Confidential reach verification resolves the conflict by separating what must be known from who must know it. Names stay with the household and the implementer. A verifier checks them under seal. The fund receives a certified distribution, with intervals, and never a name. It has four tiers, and a project buys the tiers its claim requires.

Tier 0 needs no personal data at all, and it already sorts the 18 projects above: 15 that exceed the population only when direct and indirect beneficiaries are added, and 3 that exceed it under any reading. Tier 1 sends the code to the data rather than the data to the code: a published tabulation script runs inside the implementer's systems and writes only aggregates, attested by checksum. Tier 2 sends trained enumerators, speaking the respondents' language, to a random sample and a tail sample of the register, with consent, to verify the four things the Board paper says the framework never defined: that the person exists, that the support was targeted, how much support arrived, and where the household sits in the tail (land, tenure, caste or tribe, gender of the household head, disability, age). Tier 3 deduplicates across projects through encoded identifiers (privacy-preserving record linkage with Bloom-filter encodings; Schnell, Bachteler and Reiher 2009) sent to a linkage trustee, such as a national statistics office, which returns only the number of overlaps.

What the fund receives matches the Board's proposed record field for field: verified direct beneficiaries with a 90 percent interval and the share by which the reported figure overstated them; shares by sex, age group and tail group, each tested against that group's share of the local population; the distribution of support per household; a deduplicated total; and indirect beneficiaries reported separately, labeled as estimates, capped by the population ceiling.

The safeguards are part of the design. No published cell covers fewer than 100 households drawn from at least 20 sampling clusters. For small populations, such as small island states, published shares carry calibrated and declared statistical noise (differential privacy; Dwork 2006). Access follows the five safes: safe people, projects, settings, data and outputs (Desai, Ritchie and Welpton 2016). National identity numbers are never collected; in India, no Aadhaar number, consistent with Section 29 of the Aadhaar Act, 2016. Names are kept apart from answers and destroyed six months after the final report. Every sampled household receives a route to complain that does not pass through the implementer.

Implementers gain the most. An implementer that can show whom it reaches can win adaptation finance on that evidence, report in the format the Fund's framework proposes, and never hand anyone a list of poor households. A verified count is also a protected count: it cannot later be shown to exceed the population it serves.

The same design answers the second question. A Tier 2 sample drawn at the start of a project and revisited at its end, with the same consent and the same seal, measures whether the households reached became less vulnerable: fewer losses, faster recovery, a better position in the tail. The fund receives the change, with intervals, and still never a name.

Any one of these would overturn the bad news: a published, deduplicated beneficiary record by sex, age and vulnerable group covering a fund's adaptation portfolio; field verification showing that project registers match the households they list; or reported counts that, project by project, stay within the populations they claim to serve. Any one of these would overturn the good news: a disbursement record showing that the most vulnerable countries receive their funding later or not at all once never-disbursed projects are counted; or a vulnerability index that both measures climate risk and finds the most vulnerable countries underfunded per person.

A companion working paper (Tailwinds 2026-11) reads the same portfolio through the Map of Maps of Governance (Ramprasad 2026a), testing how decisions about adaptation finance propagate from boards through intermediaries to the ground. Of seven propagation rules, one holds (Islands: the simplified track is faster), two fail (Downhill: delay does not fall harder on the most vulnerable; Broker: intermediation is uniform at about four-fifths), and four were declared untestable on these data before the analysis began.

This analysis uses public, project-level records published by the Green Climate Fund, a vulnerability index released by the European Commission's Joint Research Centre, an academic vulnerability index used here for research, and World Bank population. None contains the name of a person or household. No project is accused of misconduct: a count above a population is a statement about a definition, and the Fund's own Board paper is the source for that definition. When Tailenders Analytics verifies reach for a client, it follows the Tailenders Privacy and Confidentiality Policy, the Tail-End Protocol and the safeguards in Section 4, under a written data agreement covering purpose, consent, retention and what each party receives. Applicable law is checked for each engagement; in India, the Digital Personal Data Protection Act, 2023 and its Rules, 2025.

The analysis is at the level of countries; the who inside countries is the record's gap, and these data cannot fill it. Only the Green Climate Fund is examined. Beneficiary counts are lifetime projections reported by accredited entities, compared with current population, so the comparison is approximate; a count above the population is implausible as a count of people, but it is not proof that any person was not reached. Multi-country projects are split equally among their countries (a single-country sensitivity agrees). Sixty-three approved projects have never disbursed and are excluded from the timing tests. The population test in Section 2 was made more generous after the first results were seen, at the author's request; the stricter reading is reported beside it, and both are logged. The two vulnerability indices disagree country by country, and that disagreement is part of the finding. The design in Section 4 has not yet been piloted; a first pilot, Tiers 0 to 2 on a single project with consent and arm's-length payment, will be published whatever it shows.

1. Chen, C., Noble, I., Hellmann, J., Coffee, J., Murillo, M. and Chawla, N. 2015. University of Notre Dame Global Adaptation Index: Country Index Technical Report. University of Notre Dame.
2. Desai, T., Ritchie, F. and Welpton, R. 2016. Five Safes: designing data access for research. Economics Working Paper Series 1601, University of the West of England.
3. Dwork, C. 2006. Differential privacy. In Automata, Languages and Programming (ICALP 2006), Lecture Notes in Computer Science 4052: 1–12. Springer.
4. European Commission, Joint Research Centre. 2026. INFORM Risk Index, Mid 2026, version 0.7.3.
5. Green Climate Fund. 2011. Governing Instrument for the Green Climate Fund. Approved by the Conference of the Parties, Decision 3/CP.17.
6. Green Climate Fund. 2022. GCF Handbook: Decisions, Policies, and Frameworks, as Agreed by the Board of the Green Climate Fund from B.01 to B.30. March 2022.
7. Green Climate Fund. 2026. Harmonized Results Management Framework. GCF/B.46/11/Rev.01. Board meeting of 26 to 29 October 2026.
8. Green Climate Fund. 2026. Open data portfolio, projects endpoint (api.gcfund.org/v1/projects), accessed 12 October 2026.
9. Ramprasad, V. 2026a. The Talisman Test: A talisman for governance protection; enduring the pain of India's governance and enjoying its benefits. Tailwinds Working Paper 2026-02, Tailenders Analytics, version 1.0. doi:10.5281/zenodo.23117021.
10. Ramprasad, V. 2026b. Tail-End Standard, version 0.2. Tailenders Analytics (working draft).
11. Schnell, R., Bachteler, T. and Reiher, J. 2009. Privacy-preserving record linkage using Bloom filters. BMC Medical Informatics and Decision Making 9: 41.
12. United Nations. 1992. United Nations Framework Convention on Climate Change.
13. United Nations. 2015. Paris Agreement.
14. World Bank. 2026. World Development Indicators, Population, total (SP.POP.TOTL), accessed 12 October 2026.