[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"navigation-en":3,"index":24,"latest-posts-en":126},[4,8,12,16,20],{"title":5,"path":6,"stem":7},"The Humanitarian AI Chatbot","\u002Fprojects\u002Fai-chatbot-project","1.projects\u002F1.ai-chatbot-project",{"title":9,"path":10,"stem":11},"Analytics Dashboard for ReliefWeb Jobs Data","\u002Fprojects\u002Freliefweb-jobs-dashboard","1.projects\u002F2.reliefweb-jobs-dashboard",{"title":13,"path":14,"stem":15},"Understanding Cash Assistance in Ukraine","\u002Fprojects\u002Fhumanitarian-data-platform","1.projects\u002F3.humanitarian-data-platform",{"title":17,"path":18,"stem":19},"Humanitarian Data Analyst Portfolio","\u002Fprojects\u002Fweb-development-project","1.projects\u002F4.web-development-project",{"title":21,"path":22,"stem":23},"AI-Powered Humanitarian Legal Advisor for Ukraine","\u002Fprojects\u002Fhumanitarian-paralegal-advisor","1.projects\u002F5.humanitarian-paralegal-advisor",{"id":25,"title":26,"body":27,"cta":28,"description":41,"extension":42,"features":43,"hero":71,"meta":82,"navigation":83,"path":84,"sections":85,"seo":121,"stem":124,"testimonials":27,"__hash__":125},"index\u002F0.index.yml","Practical support for stretched humanitarian teams: operations, coordination, data, AI",null,{"title":29,"description":30,"links":31},"Ready to transform your data systems?","Let's discuss how baena.ai can help your organization leverage data for greater humanitarian impact.",[32,36],{"label":33,"to":34,"icon":35},"Get in touch","\u002Fcontact","i-lucide-message-circle",{"label":37,"to":38,"trailingIcon":39,"variant":40},"View demos","\u002Fdemos","i-lucide-arrow-right","subtle","Flexible, task-based consulting and Professional Services for social organizations. I deliver targeted technical expertise and data systems that drive humanitarian impact.","yml",{"title":44,"description":45,"items":46},"My Approach, A Partnership for Impact","My work is built on a foundation of collaboration, field-informed design, and a deep commitment to creating solutions that are not only powerful but also sustainable and ethical.",[47,51,55,59,63,67],{"title":48,"description":49,"icon":50},"Subject-Matter Expertise","Every solution is informed by years of direct humanitarian experience, technical decisions are grounded in operational reality.","i-lucide-brain-circuit",{"title":52,"description":53,"icon":54},"Collaborative & Transparent Process","I work as a partner, not just a consultant. You'll have full visibility into the project's progress through the client portal.","i-lucide-handshake",{"title":56,"description":57,"icon":58},"Sustainable & Empowering Solutions","The goal is to build capacity, not dependency. I deliver well-documented solutions and provide the training needed for your team.","i-lucide-sprout",{"title":60,"description":61,"icon":62},"Security & Ethics by Design","Security and ethical data principles are integrated from the very beginning of the design process, not as an afterthought.","i-lucide-shield-check",{"title":64,"description":65,"icon":66},"Built for the Field","Solutions architected for resilience and performance, prioritizing accessibility and usability even in low-bandwidth settings.","i-lucide-compass",{"title":68,"description":69,"icon":70},"Dedicated & Managed Infrastructure","For projects requiring robust performance and data sovereignty, I offer secure managed hosting on a private hybrid cloud.","i-lucide-server",{"links":72},[73,79],{"label":74,"icon":75,"size":76,"color":77,"to":38,"target":78},"Demos","i-lucide-projector","xl","primary","_blank",{"label":80,"icon":35,"size":76,"color":81,"variant":40,"to":34,"target":78},"Contact","neutral",{},false,"\u002F",[86,104],{"title":87,"description":88,"id":89,"orientation":90,"image":91,"features":92},"Actionable Solutions from the Ground Up","Solutions built upon deep subject-matter expertise, allowing your team to benefit directly from years of experience gained in the world's most complex contexts.","features","horizontal","\u002Fsubject_expertise.svg",[93,97,100],{"name":94,"description":95,"icon":96},"Field-Ready & User-Centric","Solutions are designed with the end-user in mind, ensuring they are intuitive and effective for humanitarian staff, regardless of their technical background.","i-lucide-users",{"name":98,"description":99,"icon":62},"Built for Resilience","Leveraging robust infrastructure and best practices to ensure tools remain available and performant, even in low-bandwidth or challenging operational settings.",{"name":101,"description":102,"icon":103},"Ethical Data Handling","Adhering to the highest standards of data protection and security, ensuring the privacy and safety of vulnerable populations are always the top priority.","i-lucide-lock",{"title":105,"description":106,"orientation":90,"reverse":107,"image":108,"features":109},"Precision Execution - The Task-Based Consultancy Model","While organizations need long-term strategic evaluations, the immediate bottleneck is usually operational - the report that is due next week or the dataset that needs cleaning. I offer Task-Based Consultancy designed to plug directly into your workflow.",true,"\u002Ffrom_pilot.svg",[110,114,117],{"name":111,"description":112,"icon":113},"Defined Deliverables, Not Open-Ended Strategy","I don't sell broad advice — I sell finished products. You define the need, I execute the solution. No open-ended strategy, no scope creep.","i-lucide-rocket",{"name":115,"description":116,"icon":54},"Donor-Ready Reporting & Compliance","The gap between field data and donor compliance is often where funding stalls. You provide the raw data; I process, analyze, and deliver a polished, compliant document ready for submission.",{"name":118,"description":119,"icon":120},"Tailor-Made IM Tools","AI has lowered the cost and complexity of building custom software. For NGOs this is especially relevant — off-the-shelf tools are built for companies, not for us.","i-lucide-chart-bar",{"title":122,"description":123},"baena.ai - Data Systems for Humanitarian Action","Flexible, task-based consulting for social organizations. I deliver targeted technical expertise and data systems that drive humanitarian impact.","0.index","uJDHT6wgiieKMHpjD7ZRHZeCnmS1vQjvd6hksoVMLd0",[127,821,1134],{"id":128,"title":129,"authors":130,"badge":136,"body":138,"date":810,"description":811,"extension":812,"external_links":27,"image":813,"meta":816,"navigation":107,"path":817,"seo":818,"stem":819,"__hash__":820},"posts\u002F3.articles\u002F11.humanitarian-flagship-reports.md","The Cost of Humanitarian Evidence: $60 Million for the Whole Planet",[131],{"name":132,"to":133,"avatar":134},"Jesus Baena","https:\u002F\u002Flinkedin.com\u002Fin\u002Fjbaenanet",{"src":135},"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002Favatar-jesus-color.jpg",{"label":137},"Humanitarian Data | Sector Analysis",{"type":139,"value":140,"toc":794},"minimark",[141,145,148,153,156,289,300,304,457,462,465,511,515,524,533,543,549,557,563,569,573,576,602,609,613,624,628,631,634,640,646,653,657,660,666,673,676,682,689,694,698,705,715,724,738,748,757,760,764,770,776,782,788],[142,143,144],"p",{},"With the help of Claude, I pulled every edition of thirteen humanitarian flagship publications from 2022 to 2026 into one folder — sixty PDFs. Then I went looking for what they cost to produce, which turned out to be the more interesting half of the exercise.",[142,146,147],{},"This is a guide to the instruments we use to describe the sector every year, and which shape decisions organisations take: which one answers which question, how to quote them without getting caught out, where the holes are, and what the whole apparatus costs relative to what industry spends knowing about far smaller things.",[149,150,152],"h2",{"id":151},"the-publishing-calendar","The publishing calendar",[142,154,155],{},"Most of these reports are annual and they land in a predictable order. As a consequence, between January and April you are working with figures that are already a year old and an appeal written the previous November. And if you are drafting anything in May or June, wait: displacement, food security and funding all refresh within about six weeks of each other. Not a big deal, but something to have in mind.",[157,158,159,175],"table",{},[160,161,162],"thead",{},[163,164,165,169,172],"tr",{},[166,167,168],"th",{},"When",[166,170,171],{},"What lands",[166,173,174],{},"Covers",[176,177,178,190,206,219,237,247,257,275],"tbody",{},[163,179,180,184,187],{},[181,182,183],"td",{},"December",[181,185,186],{},"OCHA Global Humanitarian Overview; IRC Emergency Watchlist",[181,188,189],{},"The year ahead",[163,191,192,195,203],{},[181,193,194],{},"March–April",[181,196,197,198,202],{},"CRED ",[199,200,201],"em",{},"Disasters in Numbers"," (EM-DAT)",[181,204,205],{},"Previous calendar year",[163,207,208,211,217],{},[181,209,210],{},"April–May",[181,212,213,214],{},"Global Report on Food Crises; IDMC ",[199,215,216],{},"Global Report on Internal Displacement",[181,218,205],{},[163,220,221,224,235],{},[181,222,223],{},"June",[181,225,226,227,230,231,234],{},"UNHCR ",[199,228,229],{},"Global Trends","; Global Humanitarian Assistance report; NRC ",[199,232,233],{},"most neglected crises","; GHO mid-year review",[181,236,205],{},[163,238,239,242,245],{},[181,240,241],{},"August",[181,243,244],{},"Aid Worker Security Report",[181,246,205],{},[163,248,249,252,255],{},[181,250,251],{},"October",[181,253,254],{},"Grand Bargain self-reporting highlights",[181,256,205],{},[163,258,259,262,272],{},[181,260,261],{},"Irregular",[181,263,264,265,268,269],{},"IFRC ",[199,266,267],{},"World Disasters Report","; UNDRR ",[199,270,271],{},"Global Assessment Report",[181,273,274],{},"Thematic",[163,276,277,280,286],{},[181,278,279],{},"Every four years",[181,281,282,283],{},"ALNAP ",[199,284,285],{},"State of the Humanitarian System",[181,287,288],{},"The previous four years",[142,290,291],{},[292,293],"img",{"alt":294,"className":295,"src":299},"Grid of the thirteen publications by edition year, 2022 to 2026, showing full editions, partial or substitute editions, and years with nothing published",[296,297,298],"rounded-lg","w-full","my-8","https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002Fflagship-coverage-2022-2026.png",[149,301,303],{"id":302},"which-report-answers-which-question","Which report answers which question",[157,305,306,319],{},[160,307,308],{},[163,309,310,313,316],{},[166,311,312],{},"Theme",[166,314,315],{},"Source \u002F publication",[166,317,318],{},"Notes and timing",[176,320,321,335,345,355,368,378,388,401,411,424,434,444],{},[163,322,323,329,332],{},[181,324,325],{},[326,327,328],"strong",{},"Needs, risk and crisis prioritisation",[181,330,331],{},"Global Humanitarian Overview (GHO)",[181,333,334],{},"December release with June mid-year review; an appeal document rather than a pure needs assessment",[163,336,337,339,342],{},[181,338],{},[181,340,341],{},"IRC Emergency Watchlist",[181,343,344],{},"Focuses on 20 countries, with the top 10 ranked",[163,346,347,349,352],{},[181,348],{},[181,350,351],{},"NRC Most Neglected Displacement Crises",[181,353,354],{},"Weighs funding, media coverage and political neglect; now incorporates scale",[163,356,357,362,365],{},[181,358,359],{},[326,360,361],{},"Sectoral crises and population movement",[181,363,364],{},"Global Report on Food Crises (GRFC)",[181,366,367],{},"Built systematically on IPC\u002FCH analyses",[163,369,370,372,375],{},[181,371],{},[181,373,374],{},"UNHCR Global Trends",[181,376,377],{},"Tracks end-of-year population stocks as well as annual movement flows",[163,379,380,382,385],{},[181,381],{},[181,383,384],{},"IDMC Global Report on Internal Displacement (GRID)",[181,386,387],{},"Distinguishes displaced people (stocks) from displacement movements (flows)",[163,389,390,395,398],{},[181,391,392],{},[326,393,394],{},"Disaster impact, risk reduction and economics",[181,396,397],{},"CRED Disasters in Numbers \u002F EM-DAT",[181,399,400],{},"Covers natural hazards only",[163,402,403,405,408],{},[181,404],{},[181,406,407],{},"UNDRR Global Assessment Report (GAR)",[181,409,410],{},"Released roughly biennially, alongside thematic special reports",[163,412,413,418,421],{},[181,414,415],{},[326,416,417],{},"Financing, operational performance and reform",[181,419,420],{},"Global Humanitarian Assistance (GHA) report",[181,422,423],{},"The only consolidated financing picture; produced by ALNAP since 2025",[163,425,426,428,431],{},[181,427],{},[181,429,430],{},"ALNAP State of the Humanitarian System (SOHS)",[181,432,433],{},"Comprehensive sector performance review, every four years",[163,435,436,438,441],{},[181,437],{},[181,439,440],{},"Grand Bargain reporting",[181,442,443],{},"Tracks system commitments; the independent annual review ended after 2023",[163,445,446,451,454],{},[181,447,448],{},[326,449,450],{},"Security and duty of care",[181,452,453],{},"Aid Worker Security Report (AWSR) \u002F AWSD",[181,455,456],{},"Annual report backed by an open, year-round queryable database",[458,459,461],"h3",{"id":460},"other-annual-publications-worth-tracking","Other annual publications worth tracking",[142,463,464],{},"These are not in the library, but they answer questions the thirteen do not.",[466,467,468,475,481,487,493,499,505],"ul",{},[469,470,471,474],"li",{},[326,472,473],{},"UN Secretary-General — Protection of Civilians"," (annual, May). The UN's minimum count of civilian deaths in conflict.",[469,476,477,480],{},[326,478,479],{},"UN Secretary-General — Children and Armed Conflict"," (annual, June). Verified grave violations and the \"list of shame\" annexes.",[469,482,483,486],{},[326,484,485],{},"Save the Children — Stop the War on Children"," (annual). With PRIO, the standard source on children living in conflict zones.",[469,488,489,492],{},[326,490,491],{},"ACAPS — Humanitarian Access Overview"," (twice a year). The only systematic cross-country ranking of access constraints.",[469,494,495,498],{},[326,496,497],{},"INFORM Risk"," (EC JRC \u002F IASC; annual plus mid-year). The standard crisis-risk index used for prioritisation.",[469,500,501,504],{},[326,502,503],{},"CERF Annual Results Report"," (annual). Pooled-fund allocations, including the underfunded-emergency rounds.",[469,506,507,510],{},[326,508,509],{},"FAO et al. — State of Food Security and Nutrition in the World"," (SOFI, annual). Chronic hunger; complements the GRFC's acute focus.",[149,512,514],{"id":513},"six-rules-for-quoting-these-numbers","Six rules for quoting these numbers",[142,516,517,520,521],{},[326,518,519],{},"The edition year is not the data year."," The GHO 2026 was published in December 2025 and plans for 2026. The GRFC 2026 came out in April 2026 and describes 2025. NRC and CRED name editions after the data year, so NRC's 2025 list appeared in June 2026. ",[326,522,523],{},"Your citation has to distinguish publication from reference period.",[142,525,526,529,530],{},[326,527,528],{},"Figures are revised between editions."," International humanitarian assistance for 2022 appears as US$46.9bn in the GHA 2023, US$43.9bn in Development Initiatives' 2024 report, US$47.5bn in the SOHS 2026 executive summary and US$46.1bn in the same report's performance table. Deflators, revised FTS data, different inclusion rules. ",[326,531,532],{},"Quote a figure with its edition.",[142,534,535,538,539,542],{},[326,536,537],{},"Never mix editions inside one series."," Do not build a five-year trend by taking each year's figure from the edition published that year. Aid worker deaths are the clean example: 2023 was printed as 280 and is now 297, 2024 as 383 and now 387. A trend built from printed figures shows a smaller 2023→2024 jump than a trend built from the current database. ",[326,540,541],{},"Take every point in a series from the same source vintage"," — ideally the latest edition, which restates all prior years on one basis.",[142,544,545,548],{},[326,546,547],{},"Coverage changes move the trend line."," The GRFC's country set expanded from 48 to 73 for the 2023 edition, inflating the year-on-year rise. In 2026 the reverse: eighteen countries had no usable data, the lowest coverage in a decade, so part of the fall from 295 to 266 million people in acute food insecurity is missing data rather than improvement. IDMC says the same of its own 2025 figures — reduced data availability in fifteen per cent of monitored countries, triple the previous year.",[550,551,554],"insight-box",{"icon":552,"title":553},"i-ph-warning-diamond-duotone","Declines in a year of cuts should raise an alarm",[142,555,556],{},"When the measuring apparatus is losing funding at the same time as the thing it measures, a falling number is ambiguous by default. Check the coverage note before you write the sentence — this is more relevant now than it has been at any point in the period.",[142,558,559,562],{},[326,560,561],{},"Methodology changes break series."," The Aid Worker Security Database began systematically recording detentions in 2025 and backfilled earlier years, so 2025's record total is partly definitional even though deaths fell. NRC added scale of displacement as a fourth criterion in 2026, which is largely why Sudan entered at number one.",[142,564,565,568],{},[326,566,567],{},"Know which denominator you are using."," People in need, targeted, prioritised and reached are four different numbers. For 2026: 239 million in need, 135 million targeted, 87 million in the hyper-prioritised core. SOHS 2026 records the compression — the share of people in need who were targeted fell from 67% in 2023 to 60% in 2025, with 38% hyper-prioritised. Using \"people in need\" where you mean \"people we plan to reach\" overstates coverage by a factor of nearly three.",[149,570,572],{"id":571},"where-the-holes-are","Where the holes are",[142,574,575],{},"Four of the thirteen series have gaps in this period, of four different kinds.",[466,577,578,584,590,596],{},[469,579,580,583],{},[326,581,582],{},"By design."," The SOHS appears every four years — ALNAP says the interval lets it capture major shifts while tracking long-term trends — and each edition covers the four years since the last, although this periodicity has changed since its inception. UNDRR alternates full GAR editions with special reports.",[469,585,586,589],{},[326,587,588],{},"By governance decision."," The Grand Bargain's independent review, written by ODI\u002FHPG, ended after 2023 when signatories agreed to simplify reporting under the Grand Bargain 3.0 framework. What replaced it is a Secretariat summary of self-reported data that the Secretariat says it cannot verify. The Grand Bargain's mandate expires in 2026.",[469,591,592,595],{},[326,593,594],{},"Unexplained."," IFRC published no World Disasters Report between the 2022 edition (which came out in January 2023) and March 2026.",[469,597,598,601],{},[326,599,600],{},"By attrition."," The GHO 2026 exists only as an online e-report plus summary PDFs, in a year when OCHA cut about a fifth of its staff against a US$58 million shortfall. OCHA has not linked the two. But if you are building an archive, the full-report PDF you are used to filing no longer exists.",[142,603,604],{},[292,605],{"alt":606,"className":607,"src":608},"Grid of the thirteen publications by edition year showing which funders each edition acknowledges, with USAID-funded editions highlighted",[296,297,298],"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002Fflagship-funders-2022-2026.png",[149,610,612],{"id":611},"what-usaid-was-paying-for","What USAID was paying for",[142,614,615,616,619,620,623],{},"USAID co-funded the Global Report on Food Crises in 2022 and 2023, and was a named funding partner of IDMC's displacement report from 2022 to 2024. It appears in neither now. It paid for roughly ninety per cent of EM-DAT's €300,000 budget — ",[326,617,618],{},"the four-researcher operation behind almost every disaster mortality figure in our proposals","; CRED was notified in February 2025 and still had no replacement in writing fifteen months later. Humanitarian Outcomes states that the 2025 Aid Worker Security Report was almost not produced after its database lost US government funding. ",[326,621,622],{},"FEWS NET, four decades of famine early warning, went dark for six months."," Development Initiatives, which produced the Global Humanitarian Assistance report for twenty-five years, closed — ALNAP's own State of the Humanitarian System calls it a casualty of the cuts. And ALNAP, which inherited that report, had USAID as its single largest funder at twenty-eight per cent of income.",[149,625,627],{"id":626},"what-the-whole-thing-costs","What the whole thing costs",[142,629,630],{},"Nobody appears to publish what the humanitarian sector spends on knowing things. Funding is coded by donor, recipient, crisis and channel — never by function — so there is no \"assessment, information management and analysis\" line anywhere in FTS, OECD DAC or the GHA report's own taxonomy.",[142,632,633],{},"So I reconstructed it from published accounts, grant records and procurement data, in two layers.",[142,635,636,639],{},[326,637,638],{},"Producing the thirteen reports costs roughly US$8–17 million a year",", midpoint about US$12 million. Some of that is anchored in the publishers' own accounts: IDMC spent US$7.0 million in 2023 with 64 staff; ALNAP £2.0 million in 2024–25 with 17 staff, producing both the SOHS and, since 2025, the GHA report; Humanitarian Outcomes received CAD 400,000 from Canada for sixteen months of core research including the Aid Worker Security Database; CRED runs EM-DAT on about €300,000 a year with four researchers. The rest — the GHO, the GAR — is estimated, because OCHA and UNDRR have never published what their flagships cost.",[142,641,642,645],{},[326,643,644],{},"The data systems underneath add about US$52 million a year",", and here the figures are mostly public, because they are grant-funded programmes rather than in-house activities: FEWS NET at roughly US$30 million a year (US$213 million obligated between 2019 and 2026), the IPC's global support programme at about US$8 million, IDMC's monitoring operation, the World Bank–UNHCR Joint Data Center at US$4.7 million, and OCHA's Centre for Humanitarian Data at around US$1.8 million.",[550,647,650],{"icon":648,"title":649},"i-ph-coins-duotone","On the order of US$60–70 million a year",[142,651,652],{},"Against the 2025 appeal of US$47 billion that is about 0.13%; against the US$33.3 billion actually spent, about 0.19%. And it is a floor — it excludes needs assessments, the JIAF, cluster information staff, REACH's multi-sector assessments, UNHCR's statistics function and agency monitoring, none of which publish a figure.",[149,654,656],{"id":655},"what-the-world-spends-knowing-about-other-things","What the world spends knowing about other things",[142,658,659],{},"What else in the world costs about US$60 million a year to know about?",[142,661,662,665],{},[326,663,664],{},"Barb Audiences, the company that measures what British households watch on television, ran on £40.7 million in 2024 with fourteen employees and a seven-thousand-home panel."," Knowing what one country watches on TV costs roughly four-fifths of what the world spends on its entire shared humanitarian evidence base.",[142,667,668],{},[292,669],{"alt":670,"className":671,"src":672},"Horizontal bar chart comparing annual spend on producing analysis about one domain, from global coffee statistics at three million dollars to US media audience measurement at two hundred and fifty-four million, with the humanitarian evidence base at sixty-two million",[296,297,298],"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002Fcost-of-knowing-by-subject.png",[142,674,675],{},"Just above the line: WGSN, which tells clothing brands what colours will be fashionable in eighteen months, turned over £91.3 million at a 45% EBITDA margin. Mintel and Euromonitor, two publishers of consumer market reports, took about US$200 and US$220 million. Comscore's syndicated audience business, counting who watched which advertisement, took US$254 million. On the public side, the United States spends US$135 million a year on energy statistics through the EIA and US$188 million on crop and livestock statistics through USDA NASS — each, for one country and one commodity class, more than the world spends on humanitarian analysis for everywhere.",[142,677,678,679],{},"The UK Office for National Statistics cost £374.8 million in 2024–25. Sportradar, selling sports data to bookmakers, earned €1.29 billion. L'Oréal spent €1.38 billion on research and innovation with four thousand scientists — ",[326,680,681],{},"about twenty-two times the humanitarian evidence base, on cosmetics.",[142,683,684,685,688],{},"In the other direction, to avoid dramatising: the International Coffee Organization maintains global coffee statistics on £2.7 million a year with fifteen posts, and the International Cocoa Organization runs global cocoa statistics on €3.3 million, partly funding itself by selling the data. Both are an order of magnitude ",[199,686,687],{},"below"," the humanitarian evidence base.",[142,690,691],{},[326,692,693],{},"A system this cheap is not robust, and it breaks when one funder leaves.",[149,695,697],{"id":696},"the-five-year-review","The five-year review",[142,699,700,701,704],{},"Below is what each year's editions actually said. Years here are ",[326,702,703],{},"edition years"," (the year of publication). Each set describes the previous calendar year, except the GHO and the IRC Watchlist, which look forward.",[142,706,707,710,711,714],{},[326,708,709],{},"2022 editions — reporting on 2021, published as Ukraine began."," The GHO asked for US$41 billion for 183 million of 274 million people in need, up from 235 million a year earlier. UNHCR counted 89.3 million forcibly displaced at end-2021 and IDMC a then-record 59.1 million IDPs, with Ethiopia's 5.1 million conflict displacements the highest ever recorded for a single country in one year. The GRFC put 193 million people in IPC Phase 3 or worse — nearly 40 million more than 2020 — and 570,000 in Catastrophe across Ethiopia, South Sudan, Yemen and southern Madagascar. ",[326,712,713],{},"Funding was flat at US$31.3 billion, and direct funding to local actors fell to 1.2%, the lowest in the series."," The SOHS 2022 reported that the system reached about 46% of the people it had itself identified as in need.",[142,716,717,720,721],{},[326,718,719],{},"2023 editions — reporting on 2022: Ukraine, the Horn of Africa drought, the largest displacement increase ever recorded."," The GHO launched at a record US$51.5 billion for 230 million of 339 million in need — 65 million more people than the year before. Forced displacement rose 21% to 108.4 million; the IDP stock rose 20% to 71.1 million on a record 60.9 million internal displacements, of which Ukraine accounted for roughly 16.9 million movements and Pakistan's floods 8.2 million. Food insecurity rose for a fourth consecutive year to 258 million, with 214,100 of the 376,400 people in Catastrophe in Somalia alone. ",[326,722,723],{},"International humanitarian assistance hit an all-time high of US$46.9 billion while the appeals shortfall hit an all-time high of US$22.1 billion.",[142,725,726,729,730,733,734,737],{},[326,727,728],{},"2024 editions — reporting on 2023: Sudan, Gaza, and the first deliberately leaner appeal."," The GHO asked for less than the previous year — US$46.4 billion for 180.5 million of 299.4 million in need — the first explicit narrowing of ambition. Gaza accounted for 576,600 of the 705,000 people in IPC Catastrophe, the highest total in GRFC history, and for 3.4 million displacements in the last quarter of 2023 alone. Sudan reached 9.1 million IDPs, the most ever recorded for one country. ",[326,731,732],{},"Humanitarian Outcomes counted 280 aid workers killed — the deadliest year on record at the time — 163 of them in Gaza in the war's first three months",", ",[326,735,736],{},"with 57% of all deaths caused by aerial bombardment, a method almost exclusively available to states."," Funding stalled at US$43.4 billion (−1.1%), appeal coverage fell to 45%, and Development Initiatives projected a further 11% fall.",[142,739,740,743,744,747],{},[326,741,742],{},"2025 editions — reporting on 2024, published into the collapse."," Every indicator set a record in the wrong direction: 83.4 million IDPs, above 80 million for the first time; 1.9 million people in Catastrophe, the highest since GRFC reporting began; 383 aid workers killed; Sudan at 11.6 million IDPs; food insecurity up for a sixth consecutive year to 295.3 million. ALNAP's first GHA edition reported that assistance had fallen 10% in 2024 — \"the largest cut ever recorded\" — and modelled public donor funding down 34–45% from the 2023 peak. Then the January 2025 US freeze arrived mid-cycle. OCHA rewrote the appeal in June under the title ",[199,745,746],{},"The Cruel Math of Aid Cuts",": 114.4 million people identified for hyper-prioritised assistance out of 305 million in need, needing US$29.1 billion — of which 18.5% had been received by 10 June.",[142,749,750,753,754],{},[326,751,752],{},"2026 editions — reporting on 2025: a smaller system."," The GHO was built around hyper-prioritisation from the start: 239 million in need, 135 million targeted, 87 million in the prioritised core, US$33 billion requested, about 30% below the 2025 ask. Assistance fell 20% to US$33.3 billion, the US share of government funding dropped to about 23%, and the UAE and Saudi Arabia entered the top five donors. Two famines were confirmed in the same year, in Gaza and Sudan. For the first time in a decade both global forced displacement and the IDP stock declined — but largely through returns under duress and through data that no longer exists: the GRFC 2026 has its lowest coverage in a decade, with eighteen countries lacking usable analysis, so part of the fall from 295 to 266 million is arithmetic rather than improvement. The SOHS 2026 supplies the summary judgement: the share of people in need who were even targeted fell from 67% in 2023 to 60% in 2025, and ",[326,755,756],{},"the system entered 2026 smaller, more fragile, and facing open questions about its role.",[142,758,759],{},"The 274 million people in need in the 2022 edition and the 239 million in the 2026 edition are not the same measurement. Needs did not fall by 35 million; the definition of who counts tightened under funding pressure, the GHO moved from planning for everyone in need to planning for a prioritised core, and assessment coverage shrank. The series that runs most cleanly is the funding one, because it is counted rather than estimated — and taken from a single edition, the SOHS 2026, it reads as a peak of US$46.1 billion in 2022, a 13% fall in 2024 and a further 20% in 2025, to US$33.3 billion. Note that this is not the same 2024 decline as the 10% reported by the GHA 2025 above: different edition, revised data, same direction.",[149,761,763],{"id":762},"what-i-would-actually-do-with-this","What I would actually do with this",[142,765,766,769],{},[326,767,768],{},"Keep local copies."," Publisher URLs rot faster than you expect: the FSIN links for older GRFC editions are dead, and the only reliable copy of the 2022 report I found was in a CGIAR repository. Download the PDF on publication day, one folder per series.",[142,771,772,775],{},[326,773,774],{},"Cite the edition, not the report."," \"GRFC 2025, p. 12\" ages well. \"The GRFC says 295 million\" will be changed in the next revision.",[142,777,778,781],{},[326,779,780],{},"Go to the annexes and the data."," The GHO master dataset is on HDX, EM-DAT and the AWSD are queryable, IDMC publishes country tables, and the GHA figures come from FTS and OECD DAC. For anything you will reuse, take the number from the data and use the report for the framing.",[142,783,784,787],{},[326,785,786],{},"Write down which denominator you used."," In the same year, the gap between people in need and people prioritised is 152 million, so it does matter to say what you are counting.",[789,790],"closing-cta",{"description":791,"icon":792,"link-label":33,"link-url":34,"title":793},"Thirteen series, sixty editions, 2022 to 2026, with a year-by-year summary, the funding and coverage charts and the cost workings. Assembled for internal use — happy to share it as a reference for your team.","i-ph-books-duotone","Want the library?",{"title":795,"searchDepth":796,"depth":796,"links":797},"",2,[798,799,803,804,805,806,807,808,809],{"id":151,"depth":796,"text":152},{"id":302,"depth":796,"text":303,"children":800},[801],{"id":460,"depth":802,"text":461},3,{"id":513,"depth":796,"text":514},{"id":571,"depth":796,"text":572},{"id":611,"depth":796,"text":612},{"id":626,"depth":796,"text":627},{"id":655,"depth":796,"text":656},{"id":696,"depth":796,"text":697},{"id":762,"depth":796,"text":763},"2026-09-14T00:00:00.000Z","A practitioner's guide to the humanitarian sector's flagship publications — which one answers which question, how to quote them without getting caught out, and what the whole evidence base costs compared with what the world spends knowing about television ratings.","md",{"src":814,"alt":815},"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002FComfyUI_00053_.png","Hand-drawn illustration of a tall stack of annual reports with two groups of people standing on either side of it",{},"\u002Farticles\u002Fhumanitarian-flagship-reports",{"title":129,"description":811},"3.articles\u002F11.humanitarian-flagship-reports","Y0R-q264ixMDnz2yWTDocG3JxnFLwyuzmfO-9O1eHag",{"id":822,"title":823,"authors":824,"badge":827,"body":829,"date":1118,"description":1119,"extension":812,"external_links":1120,"image":1126,"meta":1129,"navigation":107,"path":1130,"seo":1131,"stem":1132,"__hash__":1133},"posts\u002F3.articles\u002F10.data-sovereignty-ngo.md","Do No Harm Starts With Where You Keep the Data",[825],{"name":132,"to":133,"avatar":826},{"src":135},{"label":828},"Data Sovereignty | Governance",{"type":139,"value":830,"toc":1108},[831,834,843,846,853,856,860,867,876,879,882,886,889,892,895,902,905,908,911,915,922,925,928,931,934,938,941,944,950,956,962,965,969,972,978,992,995,999,1005,1008,1014,1021,1025,1028,1031,1034,1042,1045,1059,1062,1065,1069,1081,1084,1087,1094,1097,1102],[142,832,833],{},"In June 2025, a senior director from Microsoft France was asked under oath, in front of the French Senate, whether he could guarantee that French public data hosted in Microsoft's French datacentres would never be handed to US authorities. He could not. He explained that Microsoft challenges such requests where it can, and publishes the numbers. But the guarantee does not exist.",[142,835,836],{},[837,838,842],"a",{"href":839,"rel":840},"https:\u002F\u002Fwww.senat.fr\u002Fcompte-rendu-commissions\u002F20250609\u002Fce_commande_publique.html",[841],"nofollow","The transcript is public.",[142,844,845],{},"That was the day a comfortable assumption died. Many of us had been operating on the hopeful presumption that data residency was enough — that if the servers were in Frankfurt or Paris, our beneficiaries were covered and our operations were safe. Microsoft and Google both offer EU hosting and both take GDPR compliance seriously. That part is real.",[550,847,850],{"icon":848,"title":849},"i-ph-scales-duotone","Geography vs. jurisdiction",[142,851,852],{},"Residency is about geography. Sovereignty is about jurisdiction. The CLOUD Act of 2018 reaches data controlled by US-incorporated companies regardless of where it physically sits. You do not need to invoke the Patriot Act or anything exotic. The ordinary law is sufficient, and the company itself has now said so in a European parliament.",[142,854,855],{},"Nothing against US services in particular. I use several of them myself and some are excellent. But we work in a sector with a first principle, and the principle is Do No Harm. In 2026, Do No Harm starts with where you keep the data.",[149,857,859],{"id":858},"two-risks-not-one","Two risks, not one",[142,861,862,863,866],{},"When we talk about this, we usually mean confidentiality — who can read our beneficiary lists. That is real. But there is a second risk that gets much less attention and hits harder in the field: ",[326,864,865],{},"continuity",".",[142,868,869,870,875],{},"US providers must comply with the Office of Foreign Assets Control (OFAC), and this is a very real concern. In August 2026, EU leaders had to publicly rally behind the International Criminal Court ",[837,871,874],{"href":872,"rel":873},"https:\u002F\u002Fwww.euronews.com\u002Fmy-europe\u002F2026\u002F08\u002F19\u002Feu-leaders-rally-behind-international-criminal-court-after-fresh-us-sanctions",[841],"after a fresh round of US sanctions",". A mandated international body, facing service exposure because of a decision taken in another jurisdiction entirely.",[142,877,878],{},"Now imagine that logic applied to a country office. One day just an email account that stops working on a Tuesday because a compliance team somewhere applied a rule you were never party to and cannot appeal. This is a more real possibility these days.",[142,880,881],{},"That is a programme risk, and it belongs on a risk register next to access negotiations and supply chain.",[149,883,885],{"id":884},"so-what-is-actually-in-your-cloud","So what is actually in your cloud?",[142,887,888],{},"Start here, before any discussion about vendors or migrations.",[142,890,891],{},"Ask the question honestly: what is in there? Not what the policy says should be in there. What is actually in there.",[142,893,894],{},"In my experience the answer is always the same and always worse than expected. Beneficiary lists in Excel on SharePoint. Case notes in Word documents on OneDrive. Protection information in email threads. Registration data in a folder someone created for a 2019 response and never closed.",[550,896,899],{"icon":897,"title":898},"i-ph-folder-lock-duotone","A category error",[142,900,901],{},"Beneficiary data and case management do not belong in a general-purpose file system. Not in SharePoint, not in OneDrive, not in Dropbox — and not in Nextcloud either, before anyone thinks this is only a Microsoft problem. It is not. It is a category error.",[142,903,904],{},"A file system cannot do what case data requires. Permissions are almost impossible to manage properly at scale. Changes are not tracked in any way that survives an audit. There is no archiving discipline, no retention that actually executes, no meaningful record of who looked at what and when. For protection data, that access history is often the part that made the holding defensible in the first place.",[142,906,907],{},"We got away with it for years because nobody could search across all of it at once.",[142,909,910],{},"That just ended.",[149,912,914],{"id":913},"copilot-did-not-create-the-problem-but-it-made-it-visible","Copilot did not create the problem, but it made it visible.",[142,916,917,918,921],{},"Copilot and Gemini surface anything the person asking has permission to see. Not what they knew they could see. What they ",[199,919,920],{},"can"," see.",[142,923,924],{},"Most NGO tenants have a decade of permission sprawl. Sites shared with \"everyone in the organisation\" during a surge. Folders inherited from a closed project. Files technically readable by three hundred people and practically buried where nobody would ever look.",[142,926,927],{},"Copilot un-buries them. A programme officer asking a reasonable question can now surface a protection case file that permissions always allowed and no one ever expected them to find.",[142,929,930],{},"And the feature arrived by release note, not by procurement decision. Nobody chose this. It appeared.",[142,932,933],{},"Before enabling any AI assistant on your tenant, run a permissions audit. It is cheap, it takes days not months, and it is valuable whether or not you ever switch the AI on — because the exposure already exists. Copilot only lit it up.",[149,935,937],{"id":936},"the-shortcut-that-does-not-work","The shortcut that does not work",[142,939,940],{},"Some organisations reach for what feels like the safe answer: hyper-anonymisation. If we cannot protect the data, we will not hold the data. Strip the identifiers and sleep well.",[142,942,943],{},"I understand the instinct. It is still wrong, for three reasons.",[142,945,946,949],{},[326,947,948],{},"Anonymisation breaks."," Re-identification from combined attributes is well established and it gets easier every year. Location, age band, household composition, date of assistance — put enough weak identifiers together in a small caseload and you have a name. If your \"anonymised\" dataset leaks, or Copilot surfaces it, you are not off the hook.",[142,951,952,955],{},[326,953,954],{},"Your donors will not accept it."," Try explaining to an institutional donor that you cannot identify the people you served. Accountability to affected populations, verification, audit — all of it assumes you know who received what.",[142,957,958,961],{},[326,959,960],{},"And it destroys coordination."," Without identity you cannot deduplicate. You cannot deconflict with another agency working the same district, or with the line ministry, or with the cluster. So you get double registration in one place and gaps in another. That is not a privacy win. That is an operational failure with protection consequences, and the people who pay for it are the ones who were missed.",[142,963,964],{},"The answer is not to hold less. It is to hold it properly.",[149,966,968],{"id":967},"extraordinary-databases-for-extraordinary-cases","Extraordinary databases for extraordinary cases",[142,970,971],{},"This is a solved problem, and it was solved by our own sector.",[142,973,974,975],{},"If you run medical consultations, you need an electronic medical record — not a folder structure. If you run nutrition case management, DHIS2 exists and is the de facto standard across ministries of health. ",[326,976,977],{},"These tools are open source, field-proven at national scale, and they were built with the authentication, role separation, audit logging and archiving that caseload data actually requires.",[142,979,980,981,986,987,866],{},"I keep working demos if you want to see them rather than read about them: ",[837,982,985],{"href":983,"rel":984},"https:\u002F\u002Fdhis2.baena.info\u002F",[841],"dhis2.baena.info"," and ",[837,988,991],{"href":989,"rel":990},"https:\u002F\u002Ffemr.baena.info\u002F",[841],"femr.baena.info",[142,993,994],{},"The point is not these specific tools. The point is that purpose-built beats general-purpose, and that our sector already has purpose-built options that we routinely ignore in favour of a folder on SharePoint because it was already there.",[149,996,998],{"id":997},"two-traps-on-the-way-out","Two traps on the way out",[142,1000,1001,1004],{},[326,1002,1003],{},"Trap one: adapting a commercial ERP."," I have watched organisations contract enterprise ERP implementations or invest seriously in Power Apps development. Good luck.",[142,1006,1007],{},"The mismatch is structural. Commercial software is built for the business economy, which counts and invoices earnings. NGOs spend money. The whole logic runs in the opposite direction — grants, budget lines, donor restrictions, eligibility, reporting against a proposal. Every one of those has to be forced onto a data model designed for revenue. It always takes more work than the estimate, and the estimate was already the reason you chose it.",[142,1009,1010,1013],{},[326,1011,1012],{},"Trap two: building your own from scratch."," We have been here before. Enormous development cost, and then a dependency — an Oracle licence, a proprietary framework — that made changing course mid-project impossible. So the organisation ran forward and poured in more money, because stopping felt like admitting that years of work and investment were wasted. Outsourced Power Apps development lands in the same place, just with a shorter runway.",[142,1015,1016,1017,1020],{},"What has genuinely changed is the cost of building. ",[326,1018,1019],{},"AI has made custom tooling dramatically cheaper to produce."," NGO processes are more modular than they used to be. And mature open-source databases like PostgreSQL mean your data stays interoperable and portable instead of locked inside someone's licence.",[149,1022,1024],{"id":1023},"so-should-you-leave-microsoft-and-google","So should you leave Microsoft and Google?",[142,1026,1027],{},"No. And I say that as someone who self-hosts almost everything.",[142,1029,1030],{},"For a mid-size organisation that has run on M365 or Workspace for a decade, migrating the whole estate is not realistic and not a good use of your political capital. Email, documents, finance admin, HR — leave them. Take the EU data boundary options, document what remains exposed honestly, and move on.",[142,1032,1033],{},"But do not leave it as it is either. Split the problem:",[466,1035,1036],{},[469,1037,1038,1041],{},[326,1039,1040],{},"Office tools stay. Caseload data moves."," That is the line. General organisational data can live with your current provider. Beneficiary and case data goes to purpose-built systems on infrastructure you control, in a jurisdiction you chose deliberately.",[142,1043,1044],{},"Two things follow from that, and both are about continuity rather than confidentiality.",[466,1046,1047,1053],{},[469,1048,1049,1052],{},[326,1050,1051],{},"Have a plan B, and write it down."," If the tenant became unavailable organisation-wide tomorrow, what happens? Who holds a contact list that is not in Outlook? Which country office can still authorise a funds transfer? Most organisations have never asked the question. It takes an afternoon, it costs nothing, and it is the cheapest business continuity work available to you.",[469,1054,1055,1058],{},[326,1056,1057],{},"And keep a copy under the mattress."," Not everything — but the archive, the registration records, the closed-project files. The cloud should be redundancy, not the only copy.",[142,1060,1061],{},"In March 2026, Nine PBS, the public broadcaster in St. Louis, lost access to roughly 50 terabytes of material covering seventy years of local television history. Their cloud storage vendor went quiet, then defunct. In this case, it was not one of the hyperscalers but a small specialist provider they had renewed with annually since 2019. Fifty terabytes fits on a commercial NAS for less than the price of a laptop. That is the whole lesson.",[142,1063,1064],{},"Each organization operates differently, but the underlying principles apply universally. The world is evolving, and while a unified Humanitarian Cloud Service may eventually be the solution, its development does not excuse anyone from their immediate obligations.",[149,1066,1068],{"id":1067},"there-is-a-clock-on-the-rest-of-it","There is a clock on the rest of it",[142,1070,1071,1072,1075,1076,866],{},"The Microsoft grants already changed once. On 1 July 2025 the free Business Premium and Office 365 E1 grants for nonprofits ended, replaced by 300 free Business Basic seats and discounts of up to 75%. Organisations that waited until their renewal date got hurt. The lesson is not about pricing — it is that your entire operational stack was resting on a donation, not a contract, and donations are modified unilaterally. This transition is not going exactly smoothly: ",[326,1073,1074],{},"over 170,000 nonprofits have been affected",", and Slate has asked whether ",[837,1077,1080],{"href":1078,"rel":1079},"https:\u002F\u002Fslate.com\u002Ftechnology\u002F2026\u002F08\u002Fmicrosoft-software-nonprofit-data-delete.html",[841],"Microsoft is to blame for the data loss",[142,1082,1083],{},"Microsoft's recent licensing changes retire the free Business Premium tier, forcing NGOs to either pay for desktop applications and advanced security or downgrade to a web-only free tier that severely hinders offline field operations. Furthermore, Microsoft now strictly enforces an 85% active usage rule, where failing to maintain consistent account activity can trigger the revocation of an organization's entire donated license allocation. When these licenses are downgraded or revoked, the associated data is permanently purged within 30 to 90 days, making independent, self-hosted backup infrastructure essential to protect institutional memory.",[142,1085,1086],{},"Meanwhile the EU Data Act removes cloud switching charges entirely from 12 January 2027. The contract you sign this quarter will still be running on that date. If it bakes in old-world egress terms, multi-year lock-ins and auto-renewals, the deadline will arrive and find you still bound by the economics it was meant to dismantle.",[550,1088,1091],{"icon":1089,"title":1090},"i-ph-list-checks-duotone","The four things to do this quarter",[142,1092,1093],{},"Audit your permissions, move the caseload data, write the lockout plan, and read your contract before renewal.",[142,1095,1096],{},"I am writing a separate piece on AI sovereignty for NGOs. That one is a harder problem, and the good news is that it is also the layer where you still have room to move. Stay tuned.",[142,1098,1099],{},[199,1100,1101],{},"I work with humanitarian organisations on data systems, sovereign infrastructure and applied AI. If you are looking at this and not sure where to start, the permissions audit is the first move and you can do it yourself.",[789,1103],{"description":1104,"icon":1105,"link-label":1106,"link-url":34,"title":1107},"The permissions audit is the first move, and you can do it yourself. If you want a second pair of eyes on it, or on splitting office tools from caseload data, let's talk.","i-ph-chats-circle-duotone","Let's talk","Not sure where to start?",{"title":795,"searchDepth":796,"depth":796,"links":1109},[1110,1111,1112,1113,1114,1115,1116,1117],{"id":858,"depth":796,"text":859},{"id":884,"depth":796,"text":885},{"id":913,"depth":796,"text":914},{"id":936,"depth":796,"text":937},{"id":967,"depth":796,"text":968},{"id":997,"depth":796,"text":998},{"id":1023,"depth":796,"text":1024},{"id":1067,"depth":796,"text":1068},"2026-08-31T00:00:00.000Z","NGOs, sovereignty and the cloud. This is not a drill.",[1121,1124],{"label":1122,"url":983,"icon":1123},"DHIS2 Demo","i-ph-database-duotone",{"label":1125,"url":989,"icon":1123},"FEMR Demo",{"src":1127,"alt":1128},"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002FComfyUI_00043_.png","Illustration of two people reaching through a cloud between rows of server racks to exchange documents, a laptop and a key",{},"\u002Farticles\u002Fdata-sovereignty-ngo",{"title":823,"description":1119},"3.articles\u002F10.data-sovereignty-ngo","nIgloOg_Yu2LQ6tVL9TSTp7Smsd1pZ00aN4A6yIefQ8",{"id":1135,"title":1136,"authors":1137,"badge":1140,"body":1142,"date":1635,"description":1636,"extension":812,"external_links":27,"image":1637,"meta":1640,"navigation":107,"path":1641,"seo":1642,"stem":1643,"__hash__":1644},"posts\u002F3.articles\u002F9.multilingual-humanitarian-rag.md","Adding Bangla, Swahili and Hausa to a sovereign humanitarian assistant",[1138],{"name":132,"to":133,"avatar":1139},{"src":135},{"label":1141},"AI | Evaluation",{"type":139,"value":1143,"toc":1625},[1144,1147,1154,1157,1161,1164,1195,1201,1205,1208,1211,1245,1249,1256,1334,1341,1345,1352,1410,1413,1416,1434,1441,1445,1448,1451,1479,1486,1490,1497,1504,1508,1515,1554,1560,1575,1579,1586,1593,1610,1613,1618,1621],[142,1145,1146],{},"Our humanitarian assistant answers questions about the normative canon of the sector — the Sphere Handbook, the Geneva Conventions, the 1951 Refugee Convention and their siblings — from a corpus we ingest and embed ourselves, behind an egress boundary where every model call is logged, redacted and attributable. The corpus is English. The people who need it often aren't English speakers.",[142,1148,1149,1150,1153],{},"So we decided to add three languages: ",[326,1151,1152],{},"Bangla, Swahili and Hausa",". Together they cover humanitarian contexts from Cox's Bazar to the Sahel. Individually, they are three very different engineering problems — which we only know because we measured each one before building anything.",[142,1155,1156],{},"This is the story of those measurements: what they cost (single-digit euros), what they found, and the working method that made the architecture emerge from the data rather than from a preconceived opinion.",[149,1158,1160],{"id":1159},"four-layers-four-ways-to-fail","Four layers, four ways to fail",[142,1162,1163],{},"A language isn't \"supported\" because the model card says so. For a user writing in Bangla, four independent layers have to work:",[1165,1166,1167,1173,1183,1189],"ol",{},[469,1168,1169,1172],{},[326,1170,1171],{},"Detect"," — know which language the message is in. Our existing heuristic, built for a Ukrainian\u002FRussian\u002FEnglish world, classified all three new languages as English.",[469,1174,1175,1178,1179,1182],{},[326,1176,1177],{},"Retrieve"," — find the right ",[199,1180,1181],{},"English"," passages from a query written in another language. Cross-lingual retrieval is the embedding model's job.",[469,1184,1185,1188],{},[326,1186,1187],{},"Generate"," — answer in the user's language, grounded in English evidence, without reaching for a provider that breaks the sovereignty posture.",[469,1190,1191,1194],{},[326,1192,1193],{},"Redact"," — catch names, phone numbers and national IDs before anything crosses the perimeter. Fail closed, per language.",[142,1196,1197,1198,866],{},"Each layer can fail on its own, but it happens that ",[326,1199,1200],{},"they fail differently per language",[149,1202,1204],{"id":1203},"measure-first-rent-nothing","Measure first, rent nothing",[142,1206,1207],{},"Before committing to any architecture (or any GPU), we ran two probes on a serverless, per-token EU inference tier. About 2,900 model calls, less than five euros, no instance ever provisioned.",[142,1209,1210],{},"The probes inherited the evaluation discipline we'd already paid tuition for on the English stack:",[466,1212,1213,1219,1225,1239],{},[469,1214,1215,1218],{},[326,1216,1217],{},"Repeat everything."," Our eval has a known noise floor (~4 points on a 100-point scale between identical runs). Any single-run comparison smaller than that is measuring the generator's mood. Every number below is a mean of three runs, with the spread reported.",[469,1220,1221,1224],{},[326,1222,1223],{},"Score retrieval separately from generation."," Doing it together, it prevents observability.",[469,1226,1227,1230,1231,1234,1235,1238],{},[326,1228,1229],{},"Verify your translations."," The eval questions were machine-translated, then round-trip-checked by an ",[199,1232,1233],{},"independent"," model with a third model judging semantic equivalence. Survival rates: Bangla 32\u002F40, Swahili 25\u002F40, Hausa 22\u002F40 — and one Bangla flag was a silent measles–cholera substitution. That's a finding in itself: machine-translated ",[199,1236,1237],{},"content"," needs review, whatever you do about conversation.",[469,1240,1241,1244],{},[326,1242,1243],{},"Distrust your judge."," The Hausa results were re-judged in full by a second model from a different family. The ranking survived; the absolute numbers moved. Report both.",[149,1246,1248],{"id":1247},"finding-1-generation-one-model-is-language-uniform","Finding 1 — Generation: one model is language-uniform",[142,1250,1251,1252,1255],{},"We gave three candidate models the ",[199,1253,1254],{},"correct"," English evidence and the question in each language, requiring the answer in that language. This isolates cross-lingual generation from retrieval entirely.",[157,1257,1258,1276],{},[160,1259,1260],{},[163,1261,1262,1265,1267,1270,1273],{},[166,1263,1264],{},"Model",[166,1266,1181],{},[166,1268,1269],{},"Bangla",[166,1271,1272],{},"Swahili",[166,1274,1275],{},"Hausa",[176,1277,1278,1299,1316],{},[163,1279,1280,1283,1286,1291,1294],{},[181,1281,1282],{},"gemma-4-26b",[181,1284,1285],{},"0.79",[181,1287,1288],{},[326,1289,1290],{},"0.88",[181,1292,1293],{},"0.81",[181,1295,1296],{},[326,1297,1298],{},"0.78",[163,1300,1301,1304,1307,1310,1313],{},[181,1302,1303],{},"qwen3.6-35b",[181,1305,1306],{},"0.80",[181,1308,1309],{},"0.73",[181,1311,1312],{},"0.84",[181,1314,1315],{},"0.65",[163,1317,1318,1321,1324,1327,1329],{},[181,1319,1320],{},"mistral-small-3.2",[181,1322,1323],{},"0.85",[181,1325,1326],{},"0.82",[181,1328,1293],{},[181,1330,1331],{},[326,1332,1333],{},"0.31",[142,1335,1336,1337,1340],{},"One model — Gemma — brackets its own English score in all three languages. One model fails Hausa outright. And the most instructive failure wasn't in the scores at all: ",[326,1338,1339],{},"mistral-small answered Bangla questions in English 32% of the time",", despite explicit instructions. Its content was fine; its language compliance was not. An LLM judge missed this consistently — a five-line script checking the answer's Unicode script caught it every time. Cheap deterministic checks beat model judges wherever you can write one.",[149,1342,1344],{"id":1343},"finding-2-retrieval-the-mirror-image","Finding 2 — Retrieval: the mirror image",[142,1346,1347,1348,1351],{},"Then the same questions went against the live corpus as ",[199,1349,1350],{},"queries",", testing whether our embedding model (bge-m3) can retrieve English passages from non-English text. The English control reproduced our production baseline exactly — always validate the harness before believing it.",[157,1353,1354,1367],{},[160,1355,1356],{},[163,1357,1358,1361,1364],{},[166,1359,1360],{},"Query language",[166,1362,1363],{},"MRR",[166,1365,1366],{},"Questions never retrieved",[176,1368,1369,1378,1386,1396],{},[163,1370,1371,1373,1375],{},[181,1372,1181],{},[181,1374,1293],{},[181,1376,1377],{},"0 of 40",[163,1379,1380,1382,1384],{},[181,1381,1269],{},[181,1383,1285],{},[181,1385,1377],{},[163,1387,1388,1390,1393],{},[181,1389,1272],{},[181,1391,1392],{},"0.66",[181,1394,1395],{},"2 of 40",[163,1397,1398,1400,1405],{},[181,1399,1275],{},[181,1401,1402],{},[326,1403,1404],{},"0.37",[181,1406,1407],{},[326,1408,1409],{},"15 of 40",[142,1411,1412],{},"Bangla retrieves at English parity. Hausa collapses — and machine-translation noise was ruled out by re-scoring on only the verified-clean translations (same result). The embedding model simply hasn't seen enough Hausa, which matches what the African-language retrieval literature (AfriMTEB) reports for this model class.",[142,1414,1415],{},"Put the two findings side by side and the architecture designs itself:",[550,1417,1420],{"icon":1418,"title":1419},"i-ph-arrows-left-right-duotone","The pivot",[142,1421,1422,1425,1426,1429,1430,1433],{},[326,1423,1424],{},"Hausa generates fine and retrieves badly — the exact opposite of our assumption going in."," So Hausa needs help only where the query meets the embedder. Translate the ",[199,1427,1428],{},"query"," for retrieval; answer ",[199,1431,1432],{},"natively"," from the evidence. The answer path never touches machine translation.",[142,1435,1436,1437,1440],{},"We measured that pivot the same afternoon: translating Hausa queries to English with NLLB-200 (the 1.3B distilled model, on CPU, seconds per query) lifted retrieval from ",[326,1438,1439],{},"0.37 MRR to 0.70 — above native Swahili"," — and cut never-retrieved questions from 15 to 2. A language we'd have called unservable in the morning had a proven end-to-end architecture by the evening, and the fix costs a CPU-seconds translation step on one language's queries.",[149,1442,1444],{"id":1443},"finding-3-redaction-the-traps-are-all-silent","Finding 3 — Redaction: the traps are all silent",[142,1446,1447],{},"The privacy layer was the biggest gap: the industry-standard PII framework has no models at all for these languages. We added a second analyzer service — transformer NER models trained on African-language and Bangla corpora — while leaving the existing, already-measured analyzer untouched. Two services, so that expanding coverage can never silently regress the thing you've already measured.",[142,1449,1450],{},"Measuring recall before deployment caught three defects, each of which would have been invisible in code review:",[1165,1452,1453,1464,1470],{},[469,1454,1455,1458,1459,1463],{},[326,1456,1457],{},"The redactor masked the word \"beneficiary\" and leaked the actual name."," The Bangla model's tokenizer strips vowel marks; under the default span aggregation the name fragments were dropped in alignment. One config value (",[1460,1461,1462],"code",{},"aggregation: max",") fixed it — but only a recall measurement would ever have shown it.",[469,1465,1466,1469],{},[326,1467,1468],{},"The framework's context booster never fired."," ID patterns scored just below the redaction threshold while sitting next to a perfect context word (\"NID\", \"kitambulisho\"), because the multilingual tokenizer provides no lemmas for the booster to match. The numbers looked plausible; they were systematically one notch too low.",[469,1471,1472,1478],{},[326,1473,1474,1475,866],{},"\"Common Article 3\" became ",[1460,1476,1477],{},"\u003CPERSON_1>"," In Hausa, the NER model tagged the citation opener of a Geneva Conventions question as a person — which would have quietly destroyed one of the most important queries in the corpus. Domain-vocabulary allowlists exist for exactly this.",[142,1480,1481,1482,1485],{},"The deployed system now measures 120\u002F120 recall on our labelled corpus — and we publish that number with its caveat attached: the corpus is built from the same format rules the recognizers encode, so part of that score is tautological. When our Ukrainian\u002FRussian redaction met a genuinely held-out test set, ~100% became ",[326,1483,1484],{},"84.2%",". That's the number we quote for those languages, and the three new languages don't get a public number until their held-out sets exist. A recall figure without its provenance is marketing, not evidence.",[149,1487,1489],{"id":1488},"finding-4-measurement-pays-for-itself-in-bugs-you-werent-looking-for","Finding 4 — Measurement pays for itself in bugs you weren't looking for",[142,1491,1492,1493,1496],{},"Making the lexical search layer Unicode-aware (it previously dropped every Bengali character), we found the tokenizer had ",[326,1494,1495],{},"never matched the Ukrainian letter \"ї\""," — a one-character gap in a regex that had been silently fragmenting most Ukrainian words in keyword search since the feature shipped. Nobody noticed, because hybrid retrieval degraded gracefully instead of failing. That bug was found not by a bug report but by a probe refusing to produce a sensible number.",[142,1498,1499,1500,1503],{},"The same pattern recurred all day. Back-translation QA flagged 32 of 40 Hausa translations as broken — until we checked the ",[199,1501,1502],{},"checker"," and found the back-translation model was the one misreading Hausa (\"toilet\" became \"distance from home\"). Swap the reader, and 22 of 40 survive. Every measurement instrument is itself a thing to measure.",[149,1505,1507],{"id":1506},"the-self-hosting-question-finally-with-a-number","The self-hosting question, finally with a number",[142,1509,1510,1511,1514],{},"The last experiment of the session addressed a different standing question: what does self-hosting the production model actually cost in quality? We served the ",[199,1512,1513],{},"same model"," both ways — the hosted API at full precision, and a 4-bit quantized copy on our own GPU — and ran the identical evaluation, three times per arm.",[157,1516,1517,1530],{},[160,1518,1519],{},[163,1520,1521,1524,1527],{},[166,1522,1523],{},"Arm",[166,1525,1526],{},"Score",[166,1528,1529],{},"Run spread",[176,1531,1532,1543],{},[163,1533,1534,1537,1540],{},[181,1535,1536],{},"Hosted API (bf16)",[181,1538,1539],{},"0.872",[181,1541,1542],{},"0.008",[163,1544,1545,1548,1551],{},[181,1546,1547],{},"Self-hosted (int4)",[181,1549,1550],{},"0.809",[181,1552,1553],{},"0.038",[142,1555,1556,1559],{},[326,1557,1558],{},"Quantization to int4 costs about 6 points, outside the noise of both arms."," That converts a fuzzy architectural debate into a decision rule: don't route quality-sensitive traffic to an int4 arm at price parity. The self-hosted route is sold on what it's actually for — a sovereignty requirement a client pays for — with its measured quality cost stated up front, and FP8 on production-class hardware queued as the parity test.",[142,1561,1562,1563,1566,1567,1570,1571,1574],{},"One more instrumentation lesson: the first run of this comparison reported the delta as ",[199,1564,1565],{},"inside"," the noise floor. Twelve answers had scored zero because the ",[326,1568,1569],{},"judge's"," API rate-limited mid-run — an infrastructure failure masquerading as a quality signal. Only per-record failure logging made it visible. Re-judge those twelve, and the verdict flips. If your eval can't tell you ",[199,1572,1573],{},"why"," a record scored zero, it will eventually lie to you.",[149,1576,1578],{"id":1577},"what-it-cost-and-what-we-still-dont-know","What it cost, and what we still don't know",[142,1580,1581,1582,1585],{},"The entire investigation — two probes across four languages and three models, a retrieval pivot experiment, a deployed and verified redaction expansion, and a quantified self-hosting decision — cost ",[326,1583,1584],{},"single-digit euros in API tokens and zero GPU-hours rented",". The expensive-looking question (\"do we need a €1,000\u002Fmonth GPU?\") was answered for the price of a coffee, and the answer was \"not yet, and here is the number that will tell us when.\"",[142,1587,1588,1589,1592],{},"Just as important is what the evidence does ",[199,1590,1591],{},"not"," yet support, because publishing that list is what makes the rest credible:",[466,1594,1595,1598,1601,1604,1607],{},[469,1596,1597],{},"The 120\u002F120 redaction recall is a wiring proof, not a production claim — held-out sets per language come first.",[469,1599,1600],{},"Probe scores are not production scores; gold evidence makes generation easier than the live retrieval loop.",[469,1602,1603],{},"Translations passed machine QA only; native-speaker review is pending for anything user-facing.",[469,1605,1606],{},"Detection is measured on our evaluation register, not on colloquial chat, code-switching, or romanised Bangla.",[469,1608,1609],{},"Every number above is a mean over repeated runs with its spread — and the Hausa generation numbers carry more uncertainty than the others, because even two judges are still just two judges.",[142,1611,1612],{},"The method is the takeaway. Nothing here required rare infrastructure or a research budget. It is based on architectural decisions made from measurements — and treating every surprising number as a question about the instrument before believing it as a fact about the world.",[142,1614,1615],{},[199,1616,1617],{},"The full evidence pack behind this article — claim ledger, raw results, and the list of what each claim does and doesn't demonstrate — is part of our sovereign agent reference architecture.",[142,1619,1620],{},"As a bonus, we turned the same measure-first pipeline on the site itself: baena.ai is now machine-translated into Bangla, Swahili and Hausa — the same three languages this article is about — so this piece, and the rest of the site, are easier to reach for the people they're actually for.",[789,1622],{"description":1623,"icon":1105,"link-label":1106,"link-url":34,"title":1624},"If you're building AI for humanitarian or public-sector use where data residency and measurable safeguards are contractual, this is the working method we bring.","Building something similar?",{"title":795,"searchDepth":796,"depth":796,"links":1626},[1627,1628,1629,1630,1631,1632,1633,1634],{"id":1159,"depth":796,"text":1160},{"id":1203,"depth":796,"text":1204},{"id":1247,"depth":796,"text":1248},{"id":1343,"depth":796,"text":1344},{"id":1443,"depth":796,"text":1444},{"id":1488,"depth":796,"text":1489},{"id":1506,"depth":796,"text":1507},{"id":1577,"depth":796,"text":1578},"2026-08-21T00:00:00.000Z","Three Languages, One Measured Pipeline — and what it taught us about measuring before building.",{"src":1638,"alt":1639},"https:\u002F\u002Fqecdwuwkxgwkpopmdewl.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fbaena-ai-assets\u002FArticles\u002FComfyUI_00030_.png","Illustration of three people browsing shelves in a library",{},"\u002Farticles\u002Fmultilingual-humanitarian-rag",{"title":1136,"description":1636},"3.articles\u002F9.multilingual-humanitarian-rag","Zt10wLKaUlVB7gEJdjakLvdmxapX6x2HbDIRNAnPkVA"]