Contact
Have a hard problem at the edge of technology and society? We partner with institutions, agencies, and communities to prototype interventions that hold up outside the lab.
Reach us
Where we are
NYU SPS · CGA · Microsoft · New York
Founding partners
Let's talk
Send a message or a connection request and tell us about the problem you're working on. We read every note and reply to partnership and collaboration inquiries.
Message us on LinkedInContact
Have a hard problem at the edge of technology and society? We partner with institutions, agencies, and communities to prototype interventions that hold up outside the lab.
Reach us
Where we are
NYU SPS · CGA · Microsoft · New York
Founding partners
Let's talk
Send a message or a connection request and tell us about the problem you're working on. We read every note and reply to partnership and collaboration inquiries.
Message us on LinkedInMoved
Taking you there — open the portfolio if nothing happens.
Moved
Taking you there — open the portfolio if nothing happens.
Internal · not linked from the site
The lockup at hero size in every candidate, on the real background. Tell me a number and I will set the wordmark in it — or say none of them and we will go the SVG route instead.
1 · Bebas Neue — Current
What the site uses today, and what every other heading is set in.
2 · Sixtyfour — Dot matrix
Closest to the banner: separate square tiles. You rejected this one.
3 · Press Start 2P — Pixel
The arcade classic. Chunky, square, unmistakably 8-bit.
4 · Silkscreen — Pixel
Small-pixel face. You rejected this one.
5 · Pixelify Sans — Pixel
Solid strokes with stepped edges. You rejected this one.
6 · VT323 — Terminal
Old CRT terminal. Reads as a screen rather than as a grid of blocks.
7 · Orbitron — Wide tech
Squared geometric. Says 'technology' with no pixel grid at all.
8 · Michroma — Wide tech
Very wide, square, calm. Closest to the banner's proportions.
9 · Chakra Petch — Wide tech
Squared with clipped corners — a technical look that still reads fast.
10 · Rajdhani — Wide tech
Condensed and squared: nearest to Bebas's proportions, but technical.
11 · Space Grotesk — Neutral
A quiet modern sans, sibling to the Space Mono already used for body.
Internal · not linked from the site
The lockup at hero size in every candidate, on the real background. Tell me a number and I will set the wordmark in it — or say none of them and we will go the SVG route instead.
1 · Bebas Neue — Current
What the site uses today, and what every other heading is set in.
2 · Sixtyfour — Dot matrix
Closest to the banner: separate square tiles. You rejected this one.
3 · Press Start 2P — Pixel
The arcade classic. Chunky, square, unmistakably 8-bit.
4 · Silkscreen — Pixel
Small-pixel face. You rejected this one.
5 · Pixelify Sans — Pixel
Solid strokes with stepped edges. You rejected this one.
6 · VT323 — Terminal
Old CRT terminal. Reads as a screen rather than as a grid of blocks.
7 · Orbitron — Wide tech
Squared geometric. Says 'technology' with no pixel grid at all.
8 · Michroma — Wide tech
Very wide, square, calm. Closest to the banner's proportions.
9 · Chakra Petch — Wide tech
Squared with clipped corners — a technical look that still reads fast.
10 · Rajdhani — Wide tech
Condensed and squared: nearest to Bebas's proportions, but technical.
11 · Space Grotesk — Neutral
A quiet modern sans, sibling to the Space Mono already used for body.
NYU CGA × Microsoft
Exploring technology to improve
the human condition.
A research collaboration between NYU's Center for Global Affairs and Microsoft — changing the conversation on how people are informed, and how emerging technology can be used for good.
Mission
Achieving full human potential through technology — changing how people are informed, and how tech is used for good.
What are the circumstances of the human being, and how can technology improve them? We see technology as a tool — part of the solution.
Cohorts · 2025-2026
Each graduate student cohort of the Ethical Tech CoLab takes on applied research and proof-of-concept projects on intervention opportunities at the intersection of emerging technologies and the human condition.
Four projects, one frontier.
The current cohort takes four research projects from question to fielded prototype.
Prototyping and partner pilots.
Technical spikes (multi-agent harnesses, verifiable credentials, geospatial pipelines) tested against real partner needs.
Full archive coming soon
First applied research projects.
The lab's first applied research projects, spanning online safety, sustainability, AI's footprint, and generative storytelling.
Full archive coming soon
Collaborate
We partner with institutions, agencies, and communities to prototype interventions that hold up outside the lab.
NYU CGA × Microsoft
Exploring technology to improve
the human condition.
A research collaboration between NYU's Center for Global Affairs and Microsoft — changing the conversation on how people are informed, and how emerging technology can be used for good.
Mission
Achieving full human potential through technology — changing how people are informed, and how tech is used for good.
What are the circumstances of the human being, and how can technology improve them? We see technology as a tool — part of the solution.
Cohorts · 2025-2026
Each graduate student cohort of the Ethical Tech CoLab takes on applied research and proof-of-concept projects on intervention opportunities at the intersection of emerging technologies and the human condition.
Four projects, one frontier.
The current cohort takes four research projects from question to fielded prototype.
Prototyping and partner pilots.
Technical spikes (multi-agent harnesses, verifiable credentials, geospatial pipelines) tested against real partner needs.
Full archive coming soon
First applied research projects.
The lab's first applied research projects, spanning online safety, sustainability, AI's footprint, and generative storytelling.
Full archive coming soon
Collaborate
We partner with institutions, agencies, and communities to prototype interventions that hold up outside the lab.
Media · Ethical Tech Summit
Each semester the Ethical Tech CoLab convenes students, technologists, and policy practitioners for a Summit and Hackathon on the live boundary between emerging technology and global affairs.
Recurring threads: agentic systems and the future of work; verifiable supply chains under UFLPA / NDAA; online grooming and platform safety; forced labor and human trafficking; and micropayment rails for vulnerable populations and creator economies.
A visual sampler from past Summits and Hackathons — panels, student demos, partner roundtables, and the convening floor. Click any image to open the full-size version or linked demo.
Avatar Storytelling demo dashboard
Speaker · Amanda Lindsey
Speaker · Deborah Berebichez
Speaker · Ruchira Gupta
AI Researcher track
Summit floor
Working session
Panel discussion
Student demo
Hackathon
Convening
Summit moment
Summit momentSpring 2026 · NYC
Student-led applied work on x402, agent-to-agent micropayments, and downstream effects on vulnerable populations and creator economies. Industry roundtable on responsible deployment.
Fall 2025 · NYC
UFLPA / NDAA practitioner panel; demos of attestation patterns over the Supplychain Graph; student briefs on traceability gaps in critical-mineral pathways.
Sept 2025 · NYC
Ethical Tech CoLab side event during NY UN Climate Week, connecting the CoLab's emerging-technology and human-rights work to the climate agenda.
June 2025 · NYC
Ethical Tech CoLab gathering during NY Tech Week, convening technologists, students, and partners across the CoLab's emerging-technology and ethical-sourcing work.
Spring 2025 · NYC
Joint session with multilateral and civil-society partners; student demos on detection patterns and trauma-informed design; closing panel on platform accountability.
Fall 2024 · NYC
Launch of student-led applied-research projects on forced-labor structural risk; introduction of the Forced Labor Structural Risk Index methodology.
Sept 2024 · NYC
Ethical Tech CoLab side event during NY UN Climate Week, bringing the CoLab's technologists and policy practitioners into the climate conversation.
June 2024 · NYC
Ethical Tech CoLab gathering during NY Tech Week, bringing students and partners together around the CoLab's emerging-technology and human-rights work.
Spring 2024 · NYC
First convening of the Ethical Tech CoLab; framing session on how engineering practice can inform sourcing policy, and how policy can inform engineering boundaries.
Media · Ethical Tech Summit
Each semester the Ethical Tech CoLab convenes students, technologists, and policy practitioners for a Summit and Hackathon on the live boundary between emerging technology and global affairs.
Recurring threads: agentic systems and the future of work; verifiable supply chains under UFLPA / NDAA; online grooming and platform safety; forced labor and human trafficking; and micropayment rails for vulnerable populations and creator economies.
A visual sampler from past Summits and Hackathons — panels, student demos, partner roundtables, and the convening floor. Click any image to open the full-size version or linked demo.
Avatar Storytelling demo dashboardSpeaker · Amanda LindseySpeaker · Deborah BerebichezSpeaker · Ruchira GuptaAI Researcher trackSummit floorWorking sessionPanel discussionStudent demoHackathonConveningSummit momentSummit momentSpring 2026 · NYC
Student-led applied work on x402, agent-to-agent micropayments, and downstream effects on vulnerable populations and creator economies. Industry roundtable on responsible deployment.
Fall 2025 · NYC
UFLPA / NDAA practitioner panel; demos of attestation patterns over the Supplychain Graph; student briefs on traceability gaps in critical-mineral pathways.
Sept 2025 · NYC
Ethical Tech CoLab side event during NY UN Climate Week, connecting the CoLab's emerging-technology and human-rights work to the climate agenda.
June 2025 · NYC
Ethical Tech CoLab gathering during NY Tech Week, convening technologists, students, and partners across the CoLab's emerging-technology and ethical-sourcing work.
Spring 2025 · NYC
Joint session with multilateral and civil-society partners; student demos on detection patterns and trauma-informed design; closing panel on platform accountability.
Fall 2024 · NYC
Launch of student-led applied-research projects on forced-labor structural risk; introduction of the Forced Labor Structural Risk Index methodology.
Sept 2024 · NYC
Ethical Tech CoLab side event during NY UN Climate Week, bringing the CoLab's technologists and policy practitioners into the climate conversation.
June 2024 · NYC
Ethical Tech CoLab gathering during NY Tech Week, bringing students and partners together around the CoLab's emerging-technology and human-rights work.
Spring 2024 · NYC
First convening of the Ethical Tech CoLab; framing session on how engineering practice can inform sourcing policy, and how policy can inform engineering boundaries.
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
Newsletter · Monthly Intelligence Brief
One email a month: breakthrough AI for good, humanitarian innovation, exciting startups, research in plain English, opportunities for students and early-career professionals, and tools that make us better researchers. No model-release hype.
Join the list to get each edition in your inbox. Free, one email a month, unsubscribe anytime.
Newsletter · Monthly Intelligence Brief
One email a month: breakthrough AI for good, humanitarian innovation, exciting startups, research in plain English, opportunities for students and early-career professionals, and tools that make us better researchers. No model-release hype.
Join the list to get each edition in your inbox. Free, one email a month, unsubscribe anytime.
Portfolio · Live demos · Open source
Every project ships as an open repository — and most run live in your browser. Each subject below opens with the research question it answers: pick a title to read what it does, then press play and run it here. The five research questions driving the current cohort follow underneath.
How can AI inform evacuation decisions?
How can technology support the ethical return of cultural artifacts?
How can ethical claims in supply chains be made verifiable?
How can a community be warned early enough to act on the warning?
How can AI help practitioners rehearse high-stakes diplomacy?
How do new researchers learn to build this work, and share it openly?
How can a story carry a technical risk to the people it affects?
How can open evidence make infrastructure noise accountable?
Current cohort
The research questions behind the catalogue above, across disaster response, cultural heritage, supply chains, and diplomacy. Open a question to see the projects exploring it, or filter by topic.
Archive
Projects from earlier cohorts, grouped by the year they were worked on, with their live demos, code, and reports alongside.
An interactive index mapping the structural conditions that enable forced labor across 184 countries on a 0–1 risk scale, with national and sub-national layers.
A survey and toolkit for how AI helps researchers formulate rigorous, original research questions — finding gaps, generating candidate questions, summarizing the state of the art, and flagging contradictions, with guardrails against hallucinated citations. Ships with a reusable Copilot 'Researcher' prompt and a journal-credibility rubric.
Repurposing existing detection technology to identify and interrupt grooming behavior in online spaces, protecting minors on platforms where predators operate.
Standardizing ESG certificates and labels into a comparable, trustworthy signal so consumers can steer their behavior toward genuinely sustainable products.
Measuring the energy and carbon cost of AI systems, with guidance for lower-impact, more accountable deployment.
Human-condition storytelling with culturally grounded digital-human avatars — short pieces produced with generative media.
Portfolio · Live demos · Open source
Every project ships as an open repository — and most run live in your browser. Each subject below opens with the research question it answers: pick a title to read what it does, then press play and run it here. The five research questions driving the current cohort follow underneath.
How can AI inform evacuation decisions?
How can technology support the ethical return of cultural artifacts?
How can ethical claims in supply chains be made verifiable?
How can a community be warned early enough to act on the warning?
How can AI help practitioners rehearse high-stakes diplomacy?
How do new researchers learn to build this work, and share it openly?
How can a story carry a technical risk to the people it affects?
How can open evidence make infrastructure noise accountable?
Current cohort
The research questions behind the catalogue above, across disaster response, cultural heritage, supply chains, and diplomacy. Open a question to see the projects exploring it, or filter by topic.
Archive
Projects from earlier cohorts, grouped by the year they were worked on, with their live demos, code, and reports alongside.
An interactive index mapping the structural conditions that enable forced labor across 184 countries on a 0–1 risk scale, with national and sub-national layers.
A survey and toolkit for how AI helps researchers formulate rigorous, original research questions — finding gaps, generating candidate questions, summarizing the state of the art, and flagging contradictions, with guardrails against hallucinated citations. Ships with a reusable Copilot 'Researcher' prompt and a journal-credibility rubric.
Repurposing existing detection technology to identify and interrupt grooming behavior in online spaces, protecting minors on platforms where predators operate.
Standardizing ESG certificates and labels into a comparable, trustworthy signal so consumers can steer their behavior toward genuinely sustainable products.
Measuring the energy and carbon cost of AI systems, with guidance for lower-impact, more accountable deployment.
Human-condition storytelling with culturally grounded digital-human avatars — short pieces produced with generative media.
Publications · Academic report
From AI-Informed Evacuation to Digital Public Goods for Refugee Economic Inclusion
How can AI inform evacuation decisions, and what happens after? This report examines the CoLab's five fielded evacuation prototypes, their limitations and research frontiers, and extends the arc to the camp: the measurement, financial-modeling, and rights infrastructure that can shorten protracted displacement and improve lives. Dzaleka Refugee Camp in Malawi is the case study ground.
6
built instruments: five fielded evacuation prototypes and a completed forced-labour risk index, all open source
53k to 60k
people now at Dzaleka, a camp built in 1994 for 10,000 to 12,000
90%
cut to UNHCR's Malawi budget, from about US$8 million to about US$1 million
0
biometric identity systems: the category the CoLab refuses to build, recorded as a versioned decision pattern
01
The Ethical Tech CoLab's evacuation portfolio was built to answer a protection question: how can and should AI tools be used to enhance the protection and evacuation of civilians during armed conflict, and in what ways do they support or challenge existing obligations under International Humanitarian Law?
Across five fielded prototypes, a crisis-information index, a unified risk platform, a calibrated cost-and-risk decision framework, an agent-based behavior model, and a readiness-under-uncertainty simulator, one finding recurred: the binding constraint on protection is rarely the absence of a route or a resource, but the quality, accessibility, and equity of the information people and responders act on. IHL frames evacuation as a last-resort duty; the technology's job is to make that duty implementable under uncertainty.
What happens after the evacuation? A successful evacuation typically ends in a camp, where the modal outcome is years of protracted displacement. Using Dzaleka Refugee Camp, Malawi, where humanitarian financing has collapsed while an encampment policy bars residents from work, we show that the same measurement discipline, financial modeling, and uncertainty methods that served evacuation convert directly into durable-solutions infrastructure: an open Livelihoods Information and Access Index, a resettlement-and-inclusion financial model built on the CoLab's operational evacuation cost engine, privacy-preserving cash and credential rails, and a civil-society legal-assistance map. We also state what we will not build, biometric identity for displaced people, and why that refusal is itself a public good.
Most of the world's refugees live in protracted situations, and the modal outcome of reaching safety is years, sometimes decades, inside a camp or settlement. The legal architecture that governs what comes next is the Geneva refugee-protection consensus: the 1951 Convention relating to the Status of Refugees and its 1967 Protocol, the doctrine of durable solutions that UNHCR administers, and the 2018 Global Compact on Refugees. That consensus recognizes three ways a displacement can end well, voluntary repatriation, local integration, and resettlement, and each of them is gated far more by information and legal status than by physical capacity.
The CoLab's built work already reaches beyond the evacuation corridor. The same measurement discipline produced the Forced Labor Structural Risk Index, a completed instrument that maps the structural risk of forced labour across countries and sub-national regions, the exploitation that displacement, encampment, and the loss of the right to work make more likely. Read together, the evacuation models and the risk index share one purpose: to improve the information on which protection depends, and through it to reduce forced labour and the vulnerabilities that displaced people carry.
This report keeps three registers distinct throughout: what is built and validated, what is built but not yet validated, and what is proposed. No claim in these pages exceeds the register of the tool it describes.
The ask, in one paragraph. The Ethical Tech CoLab seeks grant funding to develop this technology infrastructure work as open, portable public goods, to improve the lives of displaced communities. It would be validated first at Dzaleka with the field partner Fraternidade Sem Fronteiras, and we are open to other partners with the same mission-driven motivations. Our intent is developing tools and releasing them to the public so that any research or practitioner team can adopt them. One possibility is the Displaced Livelihoods Initiative run by Innovations for Poverty Action and J-PAL, in its Infrastructure and Public Goods track. The report is written to align with that submission and to stand on its own for any peer funder or academic reviewer.
02
The protection question and the accounting gap. The portfolio's organizing question comes from Teresa Cantero's doctoral research at the Instituto de Derechos Humanos Gregorio Peces-Barba, Universidad Carlos III de Madrid: how can and should artificial intelligence be used to enhance the protection and evacuation of civilians during armed conflict, and in what ways does it support or challenge the implementation of existing obligations under international humanitarian law.
Under that law, evacuation is not discretionary. When the security of the population or imperative military reasons demand it, moving civilians out of danger is a binding duty under Article 49 of the Fourth Geneva Convention, Articles 57 and 58 of Additional Protocol I, and Article 17 of Additional Protocol II. A movement into which people are coerced is forcible transfer, a war crime regardless of motive. The duty runs wider than the battlefield: the doctrine of the responsibility to protect holds that where a state cannot or will not shield its population from mass atrocity, that responsibility passes to the international community, which frames protection as an obligation rather than a matter of goodwill.
Two structural facts motivate the program. The first is a protection gap. States move quickly to adopt AI for targeting and operational advantage, yet rarely operationalize computational models to protect civilians or to make an evacuation decision legible under the incomplete, fast-changing information those decisions face. The second is an accounting gap. Displacement is tracked in detail, on the order of 117 million people forcibly displaced worldwide by 2025, yet there is no global count of evacuated civilians, because evacuation is fragmented across militaries, governments, humanitarian agencies, volunteers, and civilians acting on their own.
The legal and institutional architecture. The 1951 Convention and its 1967 Protocol define who is a refugee and fix the cornerstone duty of non-refoulement, that no one may be returned to a place where they face persecution. The 2018 Global Compact on Refugees adds a cooperative framework organized around four objectives: easing pressure on host countries, enhancing refugee self-reliance, expanding access to third-country solutions, and supporting conditions for return in safety and dignity. International humanitarian law governs conduct in armed conflict, and the responsibility to protect frames the state's duty toward its own population.
The institutions matter as much as the instruments. UNHCR leads protection and durable solutions for refugees; the International Organization for Migration is the United Nations migration agency; the Office for the Coordination of Humanitarian Affairs coordinates emergency response through the cluster approach; and the Office of the High Commissioner for Human Rights anchors the broader human-rights instruments that apply regardless of a person's status. One structural limitation shapes everything downstream: many host states are not parties to the 1951 Convention or maintain significant reservations, so the protection a displaced person actually receives depends heavily on where they land.
The three durable solutions. The arc does not end when the shooting stops. An evacuation succeeds when a person reaches relative safety, and safety, in practice, usually means a camp. There the analytical problem inverts. The question is no longer whether to move and how, but how to leave well now that one has arrived.
The three durable solutions recognized by the refugee-protection regime, voluntary repatriation to the country of origin, local integration in the country of asylum, and resettlement to a third country, are the recognized routes out. Each is gated by information and legal status as much as by physical or economic capacity, a point the regime's own scholarship makes plainly. Resettlement reaches only a small fraction of those who need it, and complementary pathways such as labour or study routes remain modest in scale, which is why local integration and the conditions for safe return carry most of the weight in practice. This report treats the durable-solutions phase as the natural continuation of the evacuation research question, and asks what open measurement and modeling infrastructure would help displaced people and the organizations serving them find the fastest defensible route out.
03
This section describes what the CoLab has already built, and is candid about how mature each piece is, because a reader deciding whether to fund the work is owed that precision. Three registers are kept distinct. The Evacuation Risk and Cost Framework is built and validated in its evacuation form. The Evacuation Information Index, the Exodus platform, and the two simulators are built and operational but not yet expert-validated or calibrated. The Forced Labor Structural Risk Index is built and completed, documented in a methodology page and an associated publication. The livelihoods instruments the report proposes, the Livelihoods Information and Access Index and the Resettlement and Inclusion Capacity Simulator, together with the value-transfer and legal-map builds, are proposed, and their validation is future work. All are open source.
| Prototype | Function | Core method |
|---|---|---|
| Evacuation Information Index (EII) | Scores quality, accessibility, and equity of crisis information by geography | Composite index on INFORM, ACLED, and FSI logic; live news and conflict feeds |
| Exodus, Civilian Evacuation Risk Platform | Unifies EII, the cost model, and endangerment and feasibility assessment | FastAPI backend; shared calculators and data layer |
| ERCF, Evacuation Risk and Cost Framework | Estimates human and financial cost of evacuating against staying | Seven-dimension scoring; calibrated cost model; break-even analysis |
| Evacuation Behavior Simulator | Models how information and demographics shape who leaves, and when | Agent-based model; probabilistic status lifecycle |
| Readiness and Uncertainty Simulator | Models how intelligence uncertainty degrades destination choices | Monte Carlo; gatekeeper factors; uncertainty slider |
The Evacuation Information Index. The index treats crisis information itself as the object of measurement. It scores three families of factors, impact, conditions, and complexity, on a scale of 1 to 5, using weights synthesized from the INFORM Severity Index, the Armed Conflict Location and Event Data project, and the Fragile States Index. A live backend adds current news and a conflict timeline for each crisis, and a map layer can overlay AI-derived building-damage and route-damage tiles from Microsoft HASTE and Planet imagery. The design is deliberately transparent: weights are labeled provisional pending expert validation, and where the hosted site cannot run the live backend, responses are baked into a clearly labeled static snapshot so that nothing is presented as live when it is not.
The index has real limits, and the report states them rather than hiding them.
Future research on the index. Run the three-phase validation the CoLab's methodology work already specifies, budget-allocation and Delphi elicitation with 8 to 12 domain experts, Fuzzy-AHP on contested weights, then historical case calibration, and add an equity layer measuring who receives information, disaggregated by gender, age, nationality, and residence inside or outside a camp. That equity layer is the move that turns the index into the livelihoods instrument described later in this report.
Exodus, the integration layer. Exodus unifies the crisis map, the cost-and-scenario model, and an endangerment-and-feasibility assessment behind a single backend, design language, and interface, including calculators for the cost of staying, the cost of evacuating, resource needs, and scenario management. Integration is not validation. Unifying three research prototypes behind a more authoritative interface inherits their uncalibrated assumptions, which raises rather than lowers the duty to label confidence honestly. The same interface is the seam where the durable-solutions module plugs in. Its stated limitation is that it is not automatically updated; its future research is to validate the approaches, weights, and measurements.
The Evacuation Risk and Cost Framework. Now at version 7.2, the framework is the CoLab's calibrated decision-support tool for the human and financial cost of civilian evacuation, and it is the one component of the portfolio that is validated rather than provisional, in its evacuation form. It scores scenarios across seven dimensions, kinetic threat, mobility, authorization, logistics, destination, urgency, and information, estimates evacuation resources and in-situ assistance costs, and produces a break-even analysis: the day after which a one-time evacuation becomes cheaper than the mounting daily cost of sustaining a population in place. Its cost parameters are sourced to humanitarian standards and individually tagged as Validated, Estimated, or Unvalidated, drawing on UNHCR water and sanitation guidance, the Sphere standards, and World Food Programme figures.
The mortality model is disclosed as indicative rather than predictive, with a reported log-log fit of 0.855, a leave-one-out cross-validation of 0.807, and a documented case where the structure fails. The financial estimates are substantially more reliable than the mortality figures, which is why the durable-solutions adaptation leads with cost. The framework's ethical architecture is explicit: it does not place a monetary value on a human life. Its figures are logistics estimates, its labels are descriptive rather than prescriptive, and it presents data to support a decision rather than issuing one. That framing carries verbatim into every extension in this report.
The framework's limitations are equally explicit.
The Evacuation Behavior Simulator. An agent-based model of how information spread and demographics shape who leaves and when. Agents move through a lifecycle from unaware to seeking, milling, evacuating, and done, with family clusters that wait for all members, confirmation loops whose speed depends on how clear the information is, neighbor influence, and calibrated delays for elders, young children, pregnant individuals, and unaccompanied minors. It is pedagogical rather than predictive, and its household-composition logic transfers directly to modeling who is left behind by information failures inside a camp. Its behavior may vary by society, culture, location, personality, income, and education, and the future research is to build intersectionality between behavior simulations.
The Readiness and Uncertainty Simulator. A seeded, fully reproducible instrument that models how uncertainty in field intelligence degrades an evacuation decision. Destinations are scored across seven factors, 500 Monte Carlo trials per group-and-destination pair convert readiness into a distribution of success probabilities, and an uncertainty control degrades every factor's confidence at once, so a user can watch predicted outcomes collapse while nothing on the ground has changed.
Two design choices carry into the rest of this report. First, three factors are non-substitutable gatekeepers, security, the consent of the local authority, and the willingness of the host community, any of which caps readiness: a destination that is dangerous, legally closed, or unwelcoming cannot be redeemed by good logistics. Second, the model insists on the difference between a factor that is unknown, a gap that better information can close, and a factor that is a confirmed refusal, a settled political exclusion. Treating the two as the same would make an unresolvable refusal look like a solvable data problem. Structurally, assigning groups to destinations under consent, willingness, security, and capacity constraints, in the presence of uncertainty, is resettlement matching, which is the transfer the durable-solutions model makes. The simulator has not yet been stress tested with different AI models.
The Forced Labor Structural Risk Index. The index applies the same composite-index method to a different harm. It is a completed, interactive index that maps the structural conditions enabling forced labour across 184 countries on a scale from 0 to 1, with national and sub-national layers, so that the risk can be seen and compared rather than assumed. It belongs in this report because displacement is a forced-labour risk multiplier: the loss of legal status and the right to work, the collapse of assistance, and weakened oversight are precisely the conditions under which exploitation grows, and forced labour is a defined harm in international law with an estimated 27.6 million people subject to it worldwide.
Its significance for the durable-solutions argument is conceptual: measuring structural forced-labour risk and measuring livelihoods-information access are two halves of one protective picture, one showing where exploitation is likely, the other showing whether the information that could prevent it is reaching people. As a composite index its weights remain open to expert re-elicitation, it is not automatically updated, and sponsorship mechanisms remain unsourced at country scale while an eight-country hand-coded kafala signal feeds two domains in the shipped build. Writing an ILO-endorsed kafala passage on top of that unresolved discrepancy would have overstated the index's grounding.
Cross-cutting synthesis. Read together, the built prototypes yield a short set of lessons. Information access is the binding constraint, and it is the thread that carries from the corridor into the camp. Uncertainty has to be modeled explicitly, with the unknown kept separate from the confirmed, or the analysis misleads by omission. We address gaps in data by generating synthetic data that mimics real-life data, and address the residual uncertainty with Monte Carlo simulation. A composite index built from the sequence of borrowing precedents, eliciting expert weights, and calibrating against cases is portable infrastructure: the same method that scores crisis information also scores structural forced-labour risk. And cost modeling in a context of human lives has to stay descriptive rather than prescriptive, following the framework's refusal to monetize life.
04
A successful evacuation deposits people into a slower emergency. The concept of a protracted refugee situation, conventionally a population of at least 25,000 of the same nationality in exile for five years or more, describes where most refugees actually live, and care-and-maintenance encampment becomes the default by inertia rather than by anyone's plan.
The duration is not a rounding error in a person's life. By UNHCR's own accounting, the average duration of a protracted refugee situation rose from about 9 years in 1993 to roughly 17 years by 2003, and while later analysis rightly cautions that a single average conceals wide variation, the order of magnitude is not in dispute: protracted exile routinely lasts a decade or more, and often a generation. The literature on encampment and self-settlement has long argued that the camp is often less a humanitarian necessity than an administrative and political choice, and that its costs to refugees, in mobility, work, and self-reliance, are systematically understated.
Dzaleka and the funding collapse. Dzaleka Refugee Camp in Dowa District, Malawi, is the sharp end of that pattern. Built in 1994 for 10,000 to 12,000 people, it now holds on the order of 53,000 to 60,000 people, predominantly from the Democratic Republic of the Congo, Burundi, and Rwanda, and it is caught in the funding collapse of 2025 and 2026.
The numbers describe a system in retreat.
Malawi's encampment policy, enforced since March 2023, bars most residents from working outside the camp, so the population cannot substitute self-reliance for collapsing aid. Local civil society, including Inua Advocacy, argues the crisis is inseparable from that policy and is pressing for the right to work; weakened oversight is meanwhile reported to embolden trafficking networks.
This is not a context for piloting speculative technology on vulnerable people. It is a context where the highest-value contribution is measurement, evidence, and legibility: tools that help residents and the organizations serving them see and argue for the fastest route out, and that document the true cost of the status quo. This is a global humanitarian emergency that needs collective collaborative effort, and building infrastructure, digital goods and services, information, training, and capacity might transform generational poverty and increase access to resources.
Encampment as a driver of forced labour. These are the textbook conditions for forced labour. When legal work is barred and assistance collapses, the informal and coerced economy fills the gap. Reducing that exposure is a direct objective of the CoLab's work: the Forced Labor Structural Risk Index names where the risk concentrates, and the livelihoods-information instrument measures whether the information that could protect people, about their rights, their options, and legitimate work, is actually reaching them. That link has no official validation yet, and the future research is to separate recruitment, exploitation, and monetization by intersectionality using social demography, and to deepen the demand side.
05
The financial-modeling engine already exists, validated for one problem and now re-pointed to another, and the distinction matters. The Evacuation Risk and Cost Framework is validated for evacuation costing; its encampment-versus-inclusion parameterization is new and unvalidated, and most of its durable-solutions cost parameters are tagged Estimated or Unvalidated. What transfers cleanly is the architecture and the break-even mathematics, a one-time investment set against a mounting recurring cost, not a set of validated cost figures for the new use.
The engine, re-pointed.
The Resettlement and Inclusion Capacity Simulator is that engine re-pointed, and re-earning its validity in the durable-solutions domain is the work, not a formality. The capacity-and-willingness layer transfers from the readiness simulator: it maps destinations, whether resettlement states or local-integration pathways, against the non-substitutable gatekeepers of legal consent, host willingness, and security, alongside the substitutable capacities of housing, jobs, and services, while preserving the distinction between an option that is unknown and one that is refused, so that an unlobbied state is never mistaken for a closed door. The uncertainty layer carries over as well, so that thin data on host capacity or a state's posture visibly widens the confidence band and prevents false precision in an advocacy number.
What the evidence predicts. The World Bank's recent work on refugee self-reliance in Sub-Saharan Africa indicates that if refugees had the right to move and work, they would contribute to the hosting country's economy, with real fiscal and economic gains. The honest caveat is built into that same literature. The World Bank finds that self-reliance remains elusive across much of the region, because camps often sit in marginal, high-poverty areas where even full local integration would lift incomes only a little, and the scholarship on self-reliance programming warns that the concept is often invoked to justify the withdrawal of assistance rather than to expand rights.
The simulator has to be able to return the answer that inclusion does not pay here, where host markets are too thin, or it is advocacy rather than analysis. Its output is a defensible, reproducible sentence of the form: sustaining this population at Dzaleka costs approximately this much per year; this pathway costs approximately that much up front and breaks even in a given year, subject to these gatekeepers and this uncertainty band. The simulator could also provide country comparisons across different legal models and measure the economic impact of different refugee policies, and listing the policies of each hosting country would better inform humanitarian organizations and populations.
What the literature says the model will find. The engine formalizes an argument the development-economics evidence already supports.
06
The toolkit is ordered from the safest build to the one that demands the most caution. The organizing principle is data minimization: build the thinnest system that solves the problem, and refuse the parts whose risks outweigh their benefit for this population. A digital public good, in the sense set out by the Digital Public Goods Alliance and endorsed within the United Nations system, is open-source software, open data, an open standard, or an open model that is relevant to the Sustainable Development Goals, does no harm by design, and, importantly for this population, minimizes the collection of personally identifiable information. The Principles for Digital Development point the same way, toward privacy, reuse, and designing with the user. Measured against those criteria, the CoLab's contributions are intended to qualify as digital public goods rather than to resemble them loosely.
Measurement: the Livelihoods Information and Access Index. The measurement instrument is the Evacuation Information Index re-specified for displaced livelihoods and legal rights: an open, versioned instrument measuring the quality, accessibility, and equity of information about work rights, employment pathways, financial services, and entrepreneurship support reaching displaced populations. The architecture transfers from the evacuation index, but the construct does not, and saying so is part of the method. Measuring information access to work rights is a different thing from scoring crisis severity, so the index has to establish its own construct validity from the ground up rather than inherit it.
In draft form the index is a matrix. Its three dimensions, quality, meaning whether information is accurate, current, and actionable; accessibility, meaning whether people can reach it in a language and channel they use; and equity, meaning whether it reaches sub-populations evenly, are scored across four domains: work rights and legal status, employment pathways, financial services, and entrepreneurship support. Each cell draws on a mix of sources: an audit of what official and organizational information exists, a structured record from field partners of what actually circulates, and, in the future primary study, what residents themselves report receiving.
The equity dimension takes seriously the finding in the forced-migration literature that displacement is gendered, that women, men, and sexual and gender minorities face different risks and different information barriers, so disaggregation is a design requirement rather than an afterthought. Because information inside a camp travels largely by word of mouth, community leaders, and channels that are not in English, the instrument has to weigh informal and in-language sources deliberately, or it will reproduce the very bias it is built to detect. The framework is provisional and is itself a deliverable, with the weights subject to the expert-elicitation pipeline before any score is treated as evidence.
The index does not enter an empty field, and the report says so plainly rather than claiming to be first. The humanitarian sector already assesses information reaching affected populations through accountability-to-affected-populations frameworks, communication-with-communities practice, and information-ecosystem assessments. What those approaches largely lack, and what the index adds, is a composite, versioned, openly standardized measure that renders information access as a comparable variable, disaggregated by sub-population, which a later impact evaluation can treat as an outcome or a mechanism. The contribution is the measure and its portability, not the discovery that information matters. Its deliverables are the index, an open data standard, a reference implementation, and methodology documentation, validated first at Dzaleka with Fraternidade Sem Fronteiras.
The open questions the index carries.
Future research on the index. Developing a work card and a system that would allow refugees to record their expertise and be matched with regions or organizations that need those services.
Value transfer: privacy-preserving cash, decoupled from biometrics. The nearest-term practical contribution is moving value more cheaply and reliably as traditional financing collapses. The reference case is the World Food Programme's Building Blocks, the largest blockchain-based humanitarian platform, which settled entitlements on a private, permissioned ledger and paid vendors in consolidated sums, saving roughly 98 percent of transaction costs on an early deployment, improving reconciliation across agencies, and establishing a co-governed network that UN Women later extended to cash-for-skills in refugee camps.
The same case is the cautionary tale. Building Blocks ran on biometric registration, and independent analysis warns that many such pilots functioned as donor showpieces under-invested in privacy, and that even a permissioned ledger can leak sensitive metadata when on-chain identifiers are correlated with off-chain records. The lesson is to decouple the payment rail from biometric identity.
A wallet designed for Dzaleka would therefore:
Portability: skills credentials, never identity. Portability of skills is the third contribution, and it must stop short of identity. The CoLab operates an adjacent architecture that demonstrates the cryptographic pattern, an agent that issues cryptographically signed, tamper-evident verifiable credentials for cultural objects, and Professor Rhodes's prior work with the ID2020 effort concerned decentralized, portable humanitarian credentials with UNHCR and the World Food Programme. The pattern is shared; the risk surface of issuing credentials to displaced people is not, and the report treats that gap as real rather than solved.
The responsible application is narrow: portable, self-held credentials for skills, qualifications, and completed training, so that a nurse or an electrician does not lose a decade of accreditation to displacement, issued to and controlled by the individual and presented selectively. The bright line is firm. A skills credential records what a person can do. It must never become a de facto identity card that gatekeeps aid or is legible to a persecuting authority. The moment it records who someone is, it inherits the risks discussed in the next section and should be stopped.
Rights: mapping civil-society legal assistance. Donor cuts stripped UNHCR's own paralegals out of Dzaleka, which makes a maintained, open directory of who actually provides legal help the safest and highest-leverage build. Organized by legal need, refugee-status determination, work authorization, documentation, protection from gender-based violence, protection from trafficking, and resettlement referral, the map records which civil-society and refugee-led organizations provide assistance, with contact details, language, and eligibility, under the same allowlist-and-citation discipline the provenance agent enforces, so that no unsourced claim ships.
It connects to the index directly. The index shows where rights information fails; the map routes a person to the organization that can close the gap. Local actors such as Inua Advocacy are the anchor nodes, rather than the government. This legal-and-jurisdiction layer, covering work rights, documentation rules, encampment policy, recognition of qualifications, and access to financial services, is a first-class component of the work, coded from public legal sources, and it reflects the long-standing finding in the refugee literature that legal status and the right to work are among the decisive determinants of whether displaced people can build a livelihood at all.
07
The Rohingya precedent. Refugees are an exceptionally vulnerable group whose personal data demands unusual rigor, and the technology that most tempts humanitarian builders, a comprehensive digital identity for displaced people, is the one this team refuses to build. The evidence is documented rather than hypothetical. Human Rights Watch found that biometric and biographic data collected from Rohingya refugees in Bangladesh was shared onward, ultimately reaching Myanmar, the government they had fled, for the purpose of assessing repatriation eligibility, covering some 830,000 people between 2018 and 2021, without the full data-protection impact assessment UNHCR's own policy requires and without meaningful informed consent.
Consent in that system was structurally compromised. Refugees were told that aid depended on their registration, which makes agreement coerced rather than free. Consent materials were commonly presented in English, a language many of those required to register did not read, and reporting indicates that families who refused enrollment were cut off from food and services. The defining property of the harm is its irreversibility. Biometrics are sticky; an iris cannot be reissued once it is compromised, and a database built to facilitate solutions suffered precisely the function creep that critics had predicted, as data gathered for one stated purpose was repurposed for another.
Mitigation is mostly refusal, then minimization, and it can be stated as a small number of firm commitments.
For a funder, this refusal is an asset. A responsible-AI lab that can name the one thing it will not build, and cite exactly why, is more credible than one that promises everything.
08
Taken together, these instruments become an evidence engine that attacks the core information asymmetry facing displaced people, the fact that the rights-and-options case that could change an individual's situation today requires specialized legal and economic knowledge that the individual does not have. The index shows, with disaggregation, which sub-populations are systematically failed by rights-and-pathways information, which is evidence rather than anecdote. The financial model reframes the ask from a plea for help into a costed comparison, that prolonged encampment costs a given amount per year while a given pathway costs less by a given year, which is the argument that donors and states respond to. The legal map converts the evidence into action, routing people to organizations that can file, refer, and litigate.
One limit has to be stated plainly, because the evidence engine cannot dissolve it. Durable solutions are rationed by state consent, resettlement quotas, and political will, and the readiness model's own logic holds that a confirmed refusal is a settled exclusion that no measurement moves. Better information and a sharper cost argument can widen the set of options that are seriously considered and can equip advocates who would otherwise argue from anecdote, but where the binding constraint is political rather than informational, rigorous evidence may change nothing. The impact claims in this report are scoped to that reality: the instruments improve the quality and legibility of the case, which is necessary but not sufficient for a different outcome.
The outputs are designed to be wielded by refugee-led organizations and civil-society advocates, and, where they will use them, by durable-solutions units within UNHCR, rather than to sit in a dashboard. This design choice reflects a clear direction in the field, the growing insistence, visible across the refugee-studies and practitioner literature on knowledge, voice, and power and on refugee self-representation, that displaced people should hold and deploy the evidence about their own situations rather than be the objects of it.
Two guardrails hold throughout. Every advocacy artifact carries its uncertainty band honestly, and the aggregation layer never becomes a registry of identifiable individuals. It argues about populations and gaps, not named people.
09
Two things are absent by design: any biometric or identity system, and any bespoke ledger. Everything proposed is either a measurement or knowledge artifact that the CoLab is suited to own, or a thin integration on audited external rails. That is the correct risk posture for a small lab working with a highly vulnerable population, and it follows directly from the digital-public-goods and data-minimization commitments set out above.
| Layer | Component | Reuses | Posture |
|---|---|---|---|
| Measurement | LIAI index, open data standard, reference implementation, methodology | EII and the composite-index method | Build, the funding ask |
| Economics | RICS: encampment-versus-inclusion break-even; capacity and willingness map | ERCF engine (operational) and readiness gatekeepers | Adapt, engine in place |
| Value transfer | Privacy-preserving cash wallet, biometric-free | External audited network; CoLab adds threat model and measurement | Integrate, do not rebuild |
| Portability | Self-held skills and qualification credentials | Provenance verifiable-credential architecture; ID2020 lineage | Build narrowly, skills only |
| Rights | Open civil-society legal-assistance map | Provenance sourcing discipline | Build, safest and high leverage |
| Governance | DDC decision-pattern schema; DPIA and ethics layer | DDC-0001 | Build, the credibility spine |
10
A funder should not have to take this program on faith, so success is defined in advance in three tiers: outputs, whether the instruments shipped and are rigorous; outcomes, whether they work and are used; and impact, whether they changed decisions. The measurement instruments themselves double as the evidence base. One line governs the whole framework and resolves an ambiguity a careful reader would otherwise catch: under this grant, every activity involving residents is consultation and co-design of the instruments, together with a desk audit and the field partner's structured knowledge, and no personal data is collected from residents. The measurement of residents' own information access, which would require primary data collection, is a separate future study.
Outputs: did we ship, and is it rigorous? Over the first year, success means a first open-source release of the index, comprising the index itself, an open data standard, a reference implementation, and methodology documentation, each versioned and citable. It means completed and published validation of the instrument's design, with expert elicitation across 8 to 12 experts over two rounds, reported weights, consistency ratios within accepted thresholds, and divergence analysis, including negative results. It means the reference-context validation executed with Fraternidade Sem Fronteiras, drawing on an audit of existing information and the partner's structured knowledge of what circulates, using open and synthetic data. It means the financial model parameterized for Dzaleka on the operational cost engine, with every parameter tagged and the full parameter table public. And it means the civil-society legal map live with the large majority of entries verified against primary sources, and the biometric-refusal decision pattern published with its governance schema.
Full psychometric validation of the instrument on responses from the displaced population itself, covering test-retest reliability, construct validity, and measurement invariance across subgroups, requires primary data collection from that population and is not part of this phase. It is written up here as a published protocol and pre-analysis plan for a future primary study, so that the instrument and its validation plan are themselves a public good, and so that a later study can execute the plan without re-deriving it. The longitudinal read on whether information-access scores for the worst-served sub-populations improve over time belongs to that same future study. This is the honest boundary of the present work, and stating it plainly is part of the method.
Outcomes: does it work, and is it used? Over the first two years, success means concordance between the instrument and the field partner's operational knowledge of where information gaps sit at Dzaleka, and equity disaggregation that reproduces independently observed disparities. It means adoption, with at least one external research team taking up the index as an instrument and at least one practitioner organization using its scores to retarget outreach. It means the financial model cited in at least two advocacy or policy documents by refugee-led or civil-society organizations, with the uncertainty bands reproduced intact, since fidelity of use matters and not only volume. It means community legitimacy, with refugee community members engaged as co-designers and a community review session signing off on how findings are represented before publication. And it means open-source health, with public contribution metrics, quarterly releases, and a documented path to a second site.
Impact: did decisions change? Over one to three years, success means that a livelihoods or durable-solutions actor at Dzaleka changes a concrete allocation, outreach, or referral decision that can be attributed to the evidence, documented through decision memos or partner attestation. It means that the measurement gap itself closes, with information access appearing as a specified variable in at least one subsequent funded evaluation design that cites the standard, which is the definition of success for infrastructure. These impact claims are deliberately modest about attribution, for the gatekeeper reason set out above: the instruments can change what is known and argued, while the political constraints on durable solutions lie outside their reach.
Failure and accountability. Failure deserves equal candor, stated now: weights that fail expert consistency checks but ship anyway; adoption claims that rest on download counts rather than documented decisions; a map that decays unverified; or any drift from measurement toward identity. The project commits to reporting against both lists, to publishing a pre-registered analysis plan for the validation study, and to releasing all instruments, de-identified aggregate data, and code under open access at the end of the grant. A standing do-no-harm review screens any proposed feature that touches personal data against the commitments above before development, and the screening record is public.
11
This report backs the CoLab's submission to the Displaced Livelihoods Initiative in its Infrastructure and Public Goods track: the Livelihoods Information and Access Index as an open measurement public good, validated at Dzaleka. The fit is structural. That track invites a new measure that enables future rigorous evaluations rather than the evaluation of a single program, and the initiative's own call names the lack of information on legal status as a barrier to displaced livelihoods and invites AI-supported measurement in hard-to-reach contexts. The forced-labour framing is not a detour from that agenda but its sharpest edge: strengthening information about work rights and legitimate pathways is a protective factor against the exploitation the Forced Labor Structural Risk Index measures.
The proposal is scoped to research and measurement costs only, with no intervention-delivery lines, and it treats the operational cost engine, the provenance-credential architecture, and the field partner's community access as in-kind contributions rather than billed items. The request sits within the initiative's cap for this track, and the design honors the initiative's guidance that researcher-only designs tend to underperform on policy influence, which is why the field partnership with Fraternidade Sem Fronteiras is written into the plan rather than added to it.
Team and partnership. Sylvia Maier, of the NYU SPS Center for Global Affairs, serves as Principal Investigator and eligibility anchor, since the initiative requires a Principal Investigator with a primary university affiliation and demonstrated experience with rigorous evaluation. Teresa Cantero, a doctoral candidate at Universidad Carlos III de Madrid and a visiting scholar at the NYU Center for Global Affairs, is Primary Researcher; the Summer 2026 cohort carries implementation; and Fraternidade Sem Fronteiras is the field partner.
The partnership is established rather than cold: Yago Rocha volunteered with Fraternidade Sem Fronteiras at Dzaleka on refugee livelihoods and financial inclusion, and Carolina Moron visited Dzaleka while living in Malawi and building a multi-sector data coalition, which is first-hand familiarity that lowers the risk around community access. The same architecture serves adjacent funders, because the measurement-and-governance framing travels. The honest assessment is that the initiative's funded portfolio favors teams with strong evaluation track records, so competitiveness rests on the rigor of the measurement work and the depth of the team's field experience rather than on the breadth of the claims.
12
The work is designed for a term of roughly twenty-four months, beginning within the funder's required window after an award. The sequence moves from re-specifying the indicator framework for livelihoods and legal-rights information, through the expert-elicitation and calibration pipeline for the index, to the reference-context validation at Dzaleka with the field partner, and then to open release of the index, the data standard, the reference implementation, and the methodology, followed by dissemination to the research and practitioner community, including remote and partner-led presentation of results in Malawi.
Because this is an infrastructure-and-public-goods effort, the technology is the deliverable rather than overhead, and a substantial share of the budget maps to the reference implementation and the model adaptation. The cost engine itself is an in-kind contribution, already operational and validated for evacuation, which is why its re-pointing to durable solutions is an adaptation rather than a new build.
13
The limitations are stated together so that they are hard to miss.
The next steps follow from those limits.
The through-line from the evacuation corridor to the camp is a single claim, tested in more than one domain: that the binding constraint on protection, on durable solutions, and on freedom from forced labour alike is the quality, accessibility, and equity of the information that people can act on, and that open, honestly labeled measurement infrastructure is the public good most worth building for it.
References
The Ethical Tech CoLab is a research initiative of the NYU School of Professional Studies Center for Global Affairs, conducted in collaboration with Microsoft. Views and findings are those of the researchers and do not represent the official positions of New York University, Microsoft, UNHCR, WFP, or any partner institution. External programs cited are referenced as evidence, not as CoLab partnerships.
Publications · Academic report
From AI-Informed Evacuation to Digital Public Goods for Refugee Economic Inclusion
How can AI inform evacuation decisions, and what happens after? This report examines the CoLab's five fielded evacuation prototypes, their limitations and research frontiers, and extends the arc to the camp: the measurement, financial-modeling, and rights infrastructure that can shorten protracted displacement and improve lives. Dzaleka Refugee Camp in Malawi is the case study ground.
6
built instruments: five fielded evacuation prototypes and a completed forced-labour risk index, all open source
53k to 60k
people now at Dzaleka, a camp built in 1994 for 10,000 to 12,000
90%
cut to UNHCR's Malawi budget, from about US$8 million to about US$1 million
0
biometric identity systems: the category the CoLab refuses to build, recorded as a versioned decision pattern
01
The Ethical Tech CoLab's evacuation portfolio was built to answer a protection question: how can and should AI tools be used to enhance the protection and evacuation of civilians during armed conflict, and in what ways do they support or challenge existing obligations under International Humanitarian Law?
Across five fielded prototypes, a crisis-information index, a unified risk platform, a calibrated cost-and-risk decision framework, an agent-based behavior model, and a readiness-under-uncertainty simulator, one finding recurred: the binding constraint on protection is rarely the absence of a route or a resource, but the quality, accessibility, and equity of the information people and responders act on. IHL frames evacuation as a last-resort duty; the technology's job is to make that duty implementable under uncertainty.
What happens after the evacuation? A successful evacuation typically ends in a camp, where the modal outcome is years of protracted displacement. Using Dzaleka Refugee Camp, Malawi, where humanitarian financing has collapsed while an encampment policy bars residents from work, we show that the same measurement discipline, financial modeling, and uncertainty methods that served evacuation convert directly into durable-solutions infrastructure: an open Livelihoods Information and Access Index, a resettlement-and-inclusion financial model built on the CoLab's operational evacuation cost engine, privacy-preserving cash and credential rails, and a civil-society legal-assistance map. We also state what we will not build, biometric identity for displaced people, and why that refusal is itself a public good.
Most of the world's refugees live in protracted situations, and the modal outcome of reaching safety is years, sometimes decades, inside a camp or settlement. The legal architecture that governs what comes next is the Geneva refugee-protection consensus: the 1951 Convention relating to the Status of Refugees and its 1967 Protocol, the doctrine of durable solutions that UNHCR administers, and the 2018 Global Compact on Refugees. That consensus recognizes three ways a displacement can end well, voluntary repatriation, local integration, and resettlement, and each of them is gated far more by information and legal status than by physical capacity.
The CoLab's built work already reaches beyond the evacuation corridor. The same measurement discipline produced the Forced Labor Structural Risk Index, a completed instrument that maps the structural risk of forced labour across countries and sub-national regions, the exploitation that displacement, encampment, and the loss of the right to work make more likely. Read together, the evacuation models and the risk index share one purpose: to improve the information on which protection depends, and through it to reduce forced labour and the vulnerabilities that displaced people carry.
This report keeps three registers distinct throughout: what is built and validated, what is built but not yet validated, and what is proposed. No claim in these pages exceeds the register of the tool it describes.
The ask, in one paragraph. The Ethical Tech CoLab seeks grant funding to develop this technology infrastructure work as open, portable public goods, to improve the lives of displaced communities. It would be validated first at Dzaleka with the field partner Fraternidade Sem Fronteiras, and we are open to other partners with the same mission-driven motivations. Our intent is developing tools and releasing them to the public so that any research or practitioner team can adopt them. One possibility is the Displaced Livelihoods Initiative run by Innovations for Poverty Action and J-PAL, in its Infrastructure and Public Goods track. The report is written to align with that submission and to stand on its own for any peer funder or academic reviewer.
02
The protection question and the accounting gap. The portfolio's organizing question comes from Teresa Cantero's doctoral research at the Instituto de Derechos Humanos Gregorio Peces-Barba, Universidad Carlos III de Madrid: how can and should artificial intelligence be used to enhance the protection and evacuation of civilians during armed conflict, and in what ways does it support or challenge the implementation of existing obligations under international humanitarian law.
Under that law, evacuation is not discretionary. When the security of the population or imperative military reasons demand it, moving civilians out of danger is a binding duty under Article 49 of the Fourth Geneva Convention, Articles 57 and 58 of Additional Protocol I, and Article 17 of Additional Protocol II. A movement into which people are coerced is forcible transfer, a war crime regardless of motive. The duty runs wider than the battlefield: the doctrine of the responsibility to protect holds that where a state cannot or will not shield its population from mass atrocity, that responsibility passes to the international community, which frames protection as an obligation rather than a matter of goodwill.
Two structural facts motivate the program. The first is a protection gap. States move quickly to adopt AI for targeting and operational advantage, yet rarely operationalize computational models to protect civilians or to make an evacuation decision legible under the incomplete, fast-changing information those decisions face. The second is an accounting gap. Displacement is tracked in detail, on the order of 117 million people forcibly displaced worldwide by 2025, yet there is no global count of evacuated civilians, because evacuation is fragmented across militaries, governments, humanitarian agencies, volunteers, and civilians acting on their own.
The legal and institutional architecture. The 1951 Convention and its 1967 Protocol define who is a refugee and fix the cornerstone duty of non-refoulement, that no one may be returned to a place where they face persecution. The 2018 Global Compact on Refugees adds a cooperative framework organized around four objectives: easing pressure on host countries, enhancing refugee self-reliance, expanding access to third-country solutions, and supporting conditions for return in safety and dignity. International humanitarian law governs conduct in armed conflict, and the responsibility to protect frames the state's duty toward its own population.
The institutions matter as much as the instruments. UNHCR leads protection and durable solutions for refugees; the International Organization for Migration is the United Nations migration agency; the Office for the Coordination of Humanitarian Affairs coordinates emergency response through the cluster approach; and the Office of the High Commissioner for Human Rights anchors the broader human-rights instruments that apply regardless of a person's status. One structural limitation shapes everything downstream: many host states are not parties to the 1951 Convention or maintain significant reservations, so the protection a displaced person actually receives depends heavily on where they land.
The three durable solutions. The arc does not end when the shooting stops. An evacuation succeeds when a person reaches relative safety, and safety, in practice, usually means a camp. There the analytical problem inverts. The question is no longer whether to move and how, but how to leave well now that one has arrived.
The three durable solutions recognized by the refugee-protection regime, voluntary repatriation to the country of origin, local integration in the country of asylum, and resettlement to a third country, are the recognized routes out. Each is gated by information and legal status as much as by physical or economic capacity, a point the regime's own scholarship makes plainly. Resettlement reaches only a small fraction of those who need it, and complementary pathways such as labour or study routes remain modest in scale, which is why local integration and the conditions for safe return carry most of the weight in practice. This report treats the durable-solutions phase as the natural continuation of the evacuation research question, and asks what open measurement and modeling infrastructure would help displaced people and the organizations serving them find the fastest defensible route out.
03
This section describes what the CoLab has already built, and is candid about how mature each piece is, because a reader deciding whether to fund the work is owed that precision. Three registers are kept distinct. The Evacuation Risk and Cost Framework is built and validated in its evacuation form. The Evacuation Information Index, the Exodus platform, and the two simulators are built and operational but not yet expert-validated or calibrated. The Forced Labor Structural Risk Index is built and completed, documented in a methodology page and an associated publication. The livelihoods instruments the report proposes, the Livelihoods Information and Access Index and the Resettlement and Inclusion Capacity Simulator, together with the value-transfer and legal-map builds, are proposed, and their validation is future work. All are open source.
| Prototype | Function | Core method |
|---|---|---|
| Evacuation Information Index (EII) | Scores quality, accessibility, and equity of crisis information by geography | Composite index on INFORM, ACLED, and FSI logic; live news and conflict feeds |
| Exodus, Civilian Evacuation Risk Platform | Unifies EII, the cost model, and endangerment and feasibility assessment | FastAPI backend; shared calculators and data layer |
| ERCF, Evacuation Risk and Cost Framework | Estimates human and financial cost of evacuating against staying | Seven-dimension scoring; calibrated cost model; break-even analysis |
| Evacuation Behavior Simulator | Models how information and demographics shape who leaves, and when | Agent-based model; probabilistic status lifecycle |
| Readiness and Uncertainty Simulator | Models how intelligence uncertainty degrades destination choices | Monte Carlo; gatekeeper factors; uncertainty slider |
The Evacuation Information Index. The index treats crisis information itself as the object of measurement. It scores three families of factors, impact, conditions, and complexity, on a scale of 1 to 5, using weights synthesized from the INFORM Severity Index, the Armed Conflict Location and Event Data project, and the Fragile States Index. A live backend adds current news and a conflict timeline for each crisis, and a map layer can overlay AI-derived building-damage and route-damage tiles from Microsoft HASTE and Planet imagery. The design is deliberately transparent: weights are labeled provisional pending expert validation, and where the hosted site cannot run the live backend, responses are baked into a clearly labeled static snapshot so that nothing is presented as live when it is not.
The index has real limits, and the report states them rather than hiding them.
Future research on the index. Run the three-phase validation the CoLab's methodology work already specifies, budget-allocation and Delphi elicitation with 8 to 12 domain experts, Fuzzy-AHP on contested weights, then historical case calibration, and add an equity layer measuring who receives information, disaggregated by gender, age, nationality, and residence inside or outside a camp. That equity layer is the move that turns the index into the livelihoods instrument described later in this report.
Exodus, the integration layer. Exodus unifies the crisis map, the cost-and-scenario model, and an endangerment-and-feasibility assessment behind a single backend, design language, and interface, including calculators for the cost of staying, the cost of evacuating, resource needs, and scenario management. Integration is not validation. Unifying three research prototypes behind a more authoritative interface inherits their uncalibrated assumptions, which raises rather than lowers the duty to label confidence honestly. The same interface is the seam where the durable-solutions module plugs in. Its stated limitation is that it is not automatically updated; its future research is to validate the approaches, weights, and measurements.
The Evacuation Risk and Cost Framework. Now at version 7.2, the framework is the CoLab's calibrated decision-support tool for the human and financial cost of civilian evacuation, and it is the one component of the portfolio that is validated rather than provisional, in its evacuation form. It scores scenarios across seven dimensions, kinetic threat, mobility, authorization, logistics, destination, urgency, and information, estimates evacuation resources and in-situ assistance costs, and produces a break-even analysis: the day after which a one-time evacuation becomes cheaper than the mounting daily cost of sustaining a population in place. Its cost parameters are sourced to humanitarian standards and individually tagged as Validated, Estimated, or Unvalidated, drawing on UNHCR water and sanitation guidance, the Sphere standards, and World Food Programme figures.
The mortality model is disclosed as indicative rather than predictive, with a reported log-log fit of 0.855, a leave-one-out cross-validation of 0.807, and a documented case where the structure fails. The financial estimates are substantially more reliable than the mortality figures, which is why the durable-solutions adaptation leads with cost. The framework's ethical architecture is explicit: it does not place a monetary value on a human life. Its figures are logistics estimates, its labels are descriptive rather than prescriptive, and it presents data to support a decision rather than issuing one. That framing carries verbatim into every extension in this report.
The framework's limitations are equally explicit.
The Evacuation Behavior Simulator. An agent-based model of how information spread and demographics shape who leaves and when. Agents move through a lifecycle from unaware to seeking, milling, evacuating, and done, with family clusters that wait for all members, confirmation loops whose speed depends on how clear the information is, neighbor influence, and calibrated delays for elders, young children, pregnant individuals, and unaccompanied minors. It is pedagogical rather than predictive, and its household-composition logic transfers directly to modeling who is left behind by information failures inside a camp. Its behavior may vary by society, culture, location, personality, income, and education, and the future research is to build intersectionality between behavior simulations.
The Readiness and Uncertainty Simulator. A seeded, fully reproducible instrument that models how uncertainty in field intelligence degrades an evacuation decision. Destinations are scored across seven factors, 500 Monte Carlo trials per group-and-destination pair convert readiness into a distribution of success probabilities, and an uncertainty control degrades every factor's confidence at once, so a user can watch predicted outcomes collapse while nothing on the ground has changed.
Two design choices carry into the rest of this report. First, three factors are non-substitutable gatekeepers, security, the consent of the local authority, and the willingness of the host community, any of which caps readiness: a destination that is dangerous, legally closed, or unwelcoming cannot be redeemed by good logistics. Second, the model insists on the difference between a factor that is unknown, a gap that better information can close, and a factor that is a confirmed refusal, a settled political exclusion. Treating the two as the same would make an unresolvable refusal look like a solvable data problem. Structurally, assigning groups to destinations under consent, willingness, security, and capacity constraints, in the presence of uncertainty, is resettlement matching, which is the transfer the durable-solutions model makes. The simulator has not yet been stress tested with different AI models.
The Forced Labor Structural Risk Index. The index applies the same composite-index method to a different harm. It is a completed, interactive index that maps the structural conditions enabling forced labour across 184 countries on a scale from 0 to 1, with national and sub-national layers, so that the risk can be seen and compared rather than assumed. It belongs in this report because displacement is a forced-labour risk multiplier: the loss of legal status and the right to work, the collapse of assistance, and weakened oversight are precisely the conditions under which exploitation grows, and forced labour is a defined harm in international law with an estimated 27.6 million people subject to it worldwide.
Its significance for the durable-solutions argument is conceptual: measuring structural forced-labour risk and measuring livelihoods-information access are two halves of one protective picture, one showing where exploitation is likely, the other showing whether the information that could prevent it is reaching people. As a composite index its weights remain open to expert re-elicitation, it is not automatically updated, and sponsorship mechanisms remain unsourced at country scale while an eight-country hand-coded kafala signal feeds two domains in the shipped build. Writing an ILO-endorsed kafala passage on top of that unresolved discrepancy would have overstated the index's grounding.
Cross-cutting synthesis. Read together, the built prototypes yield a short set of lessons. Information access is the binding constraint, and it is the thread that carries from the corridor into the camp. Uncertainty has to be modeled explicitly, with the unknown kept separate from the confirmed, or the analysis misleads by omission. We address gaps in data by generating synthetic data that mimics real-life data, and address the residual uncertainty with Monte Carlo simulation. A composite index built from the sequence of borrowing precedents, eliciting expert weights, and calibrating against cases is portable infrastructure: the same method that scores crisis information also scores structural forced-labour risk. And cost modeling in a context of human lives has to stay descriptive rather than prescriptive, following the framework's refusal to monetize life.
04
A successful evacuation deposits people into a slower emergency. The concept of a protracted refugee situation, conventionally a population of at least 25,000 of the same nationality in exile for five years or more, describes where most refugees actually live, and care-and-maintenance encampment becomes the default by inertia rather than by anyone's plan.
The duration is not a rounding error in a person's life. By UNHCR's own accounting, the average duration of a protracted refugee situation rose from about 9 years in 1993 to roughly 17 years by 2003, and while later analysis rightly cautions that a single average conceals wide variation, the order of magnitude is not in dispute: protracted exile routinely lasts a decade or more, and often a generation. The literature on encampment and self-settlement has long argued that the camp is often less a humanitarian necessity than an administrative and political choice, and that its costs to refugees, in mobility, work, and self-reliance, are systematically understated.
Dzaleka and the funding collapse. Dzaleka Refugee Camp in Dowa District, Malawi, is the sharp end of that pattern. Built in 1994 for 10,000 to 12,000 people, it now holds on the order of 53,000 to 60,000 people, predominantly from the Democratic Republic of the Congo, Burundi, and Rwanda, and it is caught in the funding collapse of 2025 and 2026.
The numbers describe a system in retreat.
Malawi's encampment policy, enforced since March 2023, bars most residents from working outside the camp, so the population cannot substitute self-reliance for collapsing aid. Local civil society, including Inua Advocacy, argues the crisis is inseparable from that policy and is pressing for the right to work; weakened oversight is meanwhile reported to embolden trafficking networks.
This is not a context for piloting speculative technology on vulnerable people. It is a context where the highest-value contribution is measurement, evidence, and legibility: tools that help residents and the organizations serving them see and argue for the fastest route out, and that document the true cost of the status quo. This is a global humanitarian emergency that needs collective collaborative effort, and building infrastructure, digital goods and services, information, training, and capacity might transform generational poverty and increase access to resources.
Encampment as a driver of forced labour. These are the textbook conditions for forced labour. When legal work is barred and assistance collapses, the informal and coerced economy fills the gap. Reducing that exposure is a direct objective of the CoLab's work: the Forced Labor Structural Risk Index names where the risk concentrates, and the livelihoods-information instrument measures whether the information that could protect people, about their rights, their options, and legitimate work, is actually reaching them. That link has no official validation yet, and the future research is to separate recruitment, exploitation, and monetization by intersectionality using social demography, and to deepen the demand side.
05
The financial-modeling engine already exists, validated for one problem and now re-pointed to another, and the distinction matters. The Evacuation Risk and Cost Framework is validated for evacuation costing; its encampment-versus-inclusion parameterization is new and unvalidated, and most of its durable-solutions cost parameters are tagged Estimated or Unvalidated. What transfers cleanly is the architecture and the break-even mathematics, a one-time investment set against a mounting recurring cost, not a set of validated cost figures for the new use.
The engine, re-pointed.
The Resettlement and Inclusion Capacity Simulator is that engine re-pointed, and re-earning its validity in the durable-solutions domain is the work, not a formality. The capacity-and-willingness layer transfers from the readiness simulator: it maps destinations, whether resettlement states or local-integration pathways, against the non-substitutable gatekeepers of legal consent, host willingness, and security, alongside the substitutable capacities of housing, jobs, and services, while preserving the distinction between an option that is unknown and one that is refused, so that an unlobbied state is never mistaken for a closed door. The uncertainty layer carries over as well, so that thin data on host capacity or a state's posture visibly widens the confidence band and prevents false precision in an advocacy number.
What the evidence predicts. The World Bank's recent work on refugee self-reliance in Sub-Saharan Africa indicates that if refugees had the right to move and work, they would contribute to the hosting country's economy, with real fiscal and economic gains. The honest caveat is built into that same literature. The World Bank finds that self-reliance remains elusive across much of the region, because camps often sit in marginal, high-poverty areas where even full local integration would lift incomes only a little, and the scholarship on self-reliance programming warns that the concept is often invoked to justify the withdrawal of assistance rather than to expand rights.
The simulator has to be able to return the answer that inclusion does not pay here, where host markets are too thin, or it is advocacy rather than analysis. Its output is a defensible, reproducible sentence of the form: sustaining this population at Dzaleka costs approximately this much per year; this pathway costs approximately that much up front and breaks even in a given year, subject to these gatekeepers and this uncertainty band. The simulator could also provide country comparisons across different legal models and measure the economic impact of different refugee policies, and listing the policies of each hosting country would better inform humanitarian organizations and populations.
What the literature says the model will find. The engine formalizes an argument the development-economics evidence already supports.
06
The toolkit is ordered from the safest build to the one that demands the most caution. The organizing principle is data minimization: build the thinnest system that solves the problem, and refuse the parts whose risks outweigh their benefit for this population. A digital public good, in the sense set out by the Digital Public Goods Alliance and endorsed within the United Nations system, is open-source software, open data, an open standard, or an open model that is relevant to the Sustainable Development Goals, does no harm by design, and, importantly for this population, minimizes the collection of personally identifiable information. The Principles for Digital Development point the same way, toward privacy, reuse, and designing with the user. Measured against those criteria, the CoLab's contributions are intended to qualify as digital public goods rather than to resemble them loosely.
Measurement: the Livelihoods Information and Access Index. The measurement instrument is the Evacuation Information Index re-specified for displaced livelihoods and legal rights: an open, versioned instrument measuring the quality, accessibility, and equity of information about work rights, employment pathways, financial services, and entrepreneurship support reaching displaced populations. The architecture transfers from the evacuation index, but the construct does not, and saying so is part of the method. Measuring information access to work rights is a different thing from scoring crisis severity, so the index has to establish its own construct validity from the ground up rather than inherit it.
In draft form the index is a matrix. Its three dimensions, quality, meaning whether information is accurate, current, and actionable; accessibility, meaning whether people can reach it in a language and channel they use; and equity, meaning whether it reaches sub-populations evenly, are scored across four domains: work rights and legal status, employment pathways, financial services, and entrepreneurship support. Each cell draws on a mix of sources: an audit of what official and organizational information exists, a structured record from field partners of what actually circulates, and, in the future primary study, what residents themselves report receiving.
The equity dimension takes seriously the finding in the forced-migration literature that displacement is gendered, that women, men, and sexual and gender minorities face different risks and different information barriers, so disaggregation is a design requirement rather than an afterthought. Because information inside a camp travels largely by word of mouth, community leaders, and channels that are not in English, the instrument has to weigh informal and in-language sources deliberately, or it will reproduce the very bias it is built to detect. The framework is provisional and is itself a deliverable, with the weights subject to the expert-elicitation pipeline before any score is treated as evidence.
The index does not enter an empty field, and the report says so plainly rather than claiming to be first. The humanitarian sector already assesses information reaching affected populations through accountability-to-affected-populations frameworks, communication-with-communities practice, and information-ecosystem assessments. What those approaches largely lack, and what the index adds, is a composite, versioned, openly standardized measure that renders information access as a comparable variable, disaggregated by sub-population, which a later impact evaluation can treat as an outcome or a mechanism. The contribution is the measure and its portability, not the discovery that information matters. Its deliverables are the index, an open data standard, a reference implementation, and methodology documentation, validated first at Dzaleka with Fraternidade Sem Fronteiras.
The open questions the index carries.
Future research on the index. Developing a work card and a system that would allow refugees to record their expertise and be matched with regions or organizations that need those services.
Value transfer: privacy-preserving cash, decoupled from biometrics. The nearest-term practical contribution is moving value more cheaply and reliably as traditional financing collapses. The reference case is the World Food Programme's Building Blocks, the largest blockchain-based humanitarian platform, which settled entitlements on a private, permissioned ledger and paid vendors in consolidated sums, saving roughly 98 percent of transaction costs on an early deployment, improving reconciliation across agencies, and establishing a co-governed network that UN Women later extended to cash-for-skills in refugee camps.
The same case is the cautionary tale. Building Blocks ran on biometric registration, and independent analysis warns that many such pilots functioned as donor showpieces under-invested in privacy, and that even a permissioned ledger can leak sensitive metadata when on-chain identifiers are correlated with off-chain records. The lesson is to decouple the payment rail from biometric identity.
A wallet designed for Dzaleka would therefore:
Portability: skills credentials, never identity. Portability of skills is the third contribution, and it must stop short of identity. The CoLab operates an adjacent architecture that demonstrates the cryptographic pattern, an agent that issues cryptographically signed, tamper-evident verifiable credentials for cultural objects, and Professor Rhodes's prior work with the ID2020 effort concerned decentralized, portable humanitarian credentials with UNHCR and the World Food Programme. The pattern is shared; the risk surface of issuing credentials to displaced people is not, and the report treats that gap as real rather than solved.
The responsible application is narrow: portable, self-held credentials for skills, qualifications, and completed training, so that a nurse or an electrician does not lose a decade of accreditation to displacement, issued to and controlled by the individual and presented selectively. The bright line is firm. A skills credential records what a person can do. It must never become a de facto identity card that gatekeeps aid or is legible to a persecuting authority. The moment it records who someone is, it inherits the risks discussed in the next section and should be stopped.
Rights: mapping civil-society legal assistance. Donor cuts stripped UNHCR's own paralegals out of Dzaleka, which makes a maintained, open directory of who actually provides legal help the safest and highest-leverage build. Organized by legal need, refugee-status determination, work authorization, documentation, protection from gender-based violence, protection from trafficking, and resettlement referral, the map records which civil-society and refugee-led organizations provide assistance, with contact details, language, and eligibility, under the same allowlist-and-citation discipline the provenance agent enforces, so that no unsourced claim ships.
It connects to the index directly. The index shows where rights information fails; the map routes a person to the organization that can close the gap. Local actors such as Inua Advocacy are the anchor nodes, rather than the government. This legal-and-jurisdiction layer, covering work rights, documentation rules, encampment policy, recognition of qualifications, and access to financial services, is a first-class component of the work, coded from public legal sources, and it reflects the long-standing finding in the refugee literature that legal status and the right to work are among the decisive determinants of whether displaced people can build a livelihood at all.
07
The Rohingya precedent. Refugees are an exceptionally vulnerable group whose personal data demands unusual rigor, and the technology that most tempts humanitarian builders, a comprehensive digital identity for displaced people, is the one this team refuses to build. The evidence is documented rather than hypothetical. Human Rights Watch found that biometric and biographic data collected from Rohingya refugees in Bangladesh was shared onward, ultimately reaching Myanmar, the government they had fled, for the purpose of assessing repatriation eligibility, covering some 830,000 people between 2018 and 2021, without the full data-protection impact assessment UNHCR's own policy requires and without meaningful informed consent.
Consent in that system was structurally compromised. Refugees were told that aid depended on their registration, which makes agreement coerced rather than free. Consent materials were commonly presented in English, a language many of those required to register did not read, and reporting indicates that families who refused enrollment were cut off from food and services. The defining property of the harm is its irreversibility. Biometrics are sticky; an iris cannot be reissued once it is compromised, and a database built to facilitate solutions suffered precisely the function creep that critics had predicted, as data gathered for one stated purpose was repurposed for another.
Mitigation is mostly refusal, then minimization, and it can be stated as a small number of firm commitments.
For a funder, this refusal is an asset. A responsible-AI lab that can name the one thing it will not build, and cite exactly why, is more credible than one that promises everything.
08
Taken together, these instruments become an evidence engine that attacks the core information asymmetry facing displaced people, the fact that the rights-and-options case that could change an individual's situation today requires specialized legal and economic knowledge that the individual does not have. The index shows, with disaggregation, which sub-populations are systematically failed by rights-and-pathways information, which is evidence rather than anecdote. The financial model reframes the ask from a plea for help into a costed comparison, that prolonged encampment costs a given amount per year while a given pathway costs less by a given year, which is the argument that donors and states respond to. The legal map converts the evidence into action, routing people to organizations that can file, refer, and litigate.
One limit has to be stated plainly, because the evidence engine cannot dissolve it. Durable solutions are rationed by state consent, resettlement quotas, and political will, and the readiness model's own logic holds that a confirmed refusal is a settled exclusion that no measurement moves. Better information and a sharper cost argument can widen the set of options that are seriously considered and can equip advocates who would otherwise argue from anecdote, but where the binding constraint is political rather than informational, rigorous evidence may change nothing. The impact claims in this report are scoped to that reality: the instruments improve the quality and legibility of the case, which is necessary but not sufficient for a different outcome.
The outputs are designed to be wielded by refugee-led organizations and civil-society advocates, and, where they will use them, by durable-solutions units within UNHCR, rather than to sit in a dashboard. This design choice reflects a clear direction in the field, the growing insistence, visible across the refugee-studies and practitioner literature on knowledge, voice, and power and on refugee self-representation, that displaced people should hold and deploy the evidence about their own situations rather than be the objects of it.
Two guardrails hold throughout. Every advocacy artifact carries its uncertainty band honestly, and the aggregation layer never becomes a registry of identifiable individuals. It argues about populations and gaps, not named people.
09
Two things are absent by design: any biometric or identity system, and any bespoke ledger. Everything proposed is either a measurement or knowledge artifact that the CoLab is suited to own, or a thin integration on audited external rails. That is the correct risk posture for a small lab working with a highly vulnerable population, and it follows directly from the digital-public-goods and data-minimization commitments set out above.
| Layer | Component | Reuses | Posture |
|---|---|---|---|
| Measurement | LIAI index, open data standard, reference implementation, methodology | EII and the composite-index method | Build, the funding ask |
| Economics | RICS: encampment-versus-inclusion break-even; capacity and willingness map | ERCF engine (operational) and readiness gatekeepers | Adapt, engine in place |
| Value transfer | Privacy-preserving cash wallet, biometric-free | External audited network; CoLab adds threat model and measurement | Integrate, do not rebuild |
| Portability | Self-held skills and qualification credentials | Provenance verifiable-credential architecture; ID2020 lineage | Build narrowly, skills only |
| Rights | Open civil-society legal-assistance map | Provenance sourcing discipline | Build, safest and high leverage |
| Governance | DDC decision-pattern schema; DPIA and ethics layer | DDC-0001 | Build, the credibility spine |
10
A funder should not have to take this program on faith, so success is defined in advance in three tiers: outputs, whether the instruments shipped and are rigorous; outcomes, whether they work and are used; and impact, whether they changed decisions. The measurement instruments themselves double as the evidence base. One line governs the whole framework and resolves an ambiguity a careful reader would otherwise catch: under this grant, every activity involving residents is consultation and co-design of the instruments, together with a desk audit and the field partner's structured knowledge, and no personal data is collected from residents. The measurement of residents' own information access, which would require primary data collection, is a separate future study.
Outputs: did we ship, and is it rigorous? Over the first year, success means a first open-source release of the index, comprising the index itself, an open data standard, a reference implementation, and methodology documentation, each versioned and citable. It means completed and published validation of the instrument's design, with expert elicitation across 8 to 12 experts over two rounds, reported weights, consistency ratios within accepted thresholds, and divergence analysis, including negative results. It means the reference-context validation executed with Fraternidade Sem Fronteiras, drawing on an audit of existing information and the partner's structured knowledge of what circulates, using open and synthetic data. It means the financial model parameterized for Dzaleka on the operational cost engine, with every parameter tagged and the full parameter table public. And it means the civil-society legal map live with the large majority of entries verified against primary sources, and the biometric-refusal decision pattern published with its governance schema.
Full psychometric validation of the instrument on responses from the displaced population itself, covering test-retest reliability, construct validity, and measurement invariance across subgroups, requires primary data collection from that population and is not part of this phase. It is written up here as a published protocol and pre-analysis plan for a future primary study, so that the instrument and its validation plan are themselves a public good, and so that a later study can execute the plan without re-deriving it. The longitudinal read on whether information-access scores for the worst-served sub-populations improve over time belongs to that same future study. This is the honest boundary of the present work, and stating it plainly is part of the method.
Outcomes: does it work, and is it used? Over the first two years, success means concordance between the instrument and the field partner's operational knowledge of where information gaps sit at Dzaleka, and equity disaggregation that reproduces independently observed disparities. It means adoption, with at least one external research team taking up the index as an instrument and at least one practitioner organization using its scores to retarget outreach. It means the financial model cited in at least two advocacy or policy documents by refugee-led or civil-society organizations, with the uncertainty bands reproduced intact, since fidelity of use matters and not only volume. It means community legitimacy, with refugee community members engaged as co-designers and a community review session signing off on how findings are represented before publication. And it means open-source health, with public contribution metrics, quarterly releases, and a documented path to a second site.
Impact: did decisions change? Over one to three years, success means that a livelihoods or durable-solutions actor at Dzaleka changes a concrete allocation, outreach, or referral decision that can be attributed to the evidence, documented through decision memos or partner attestation. It means that the measurement gap itself closes, with information access appearing as a specified variable in at least one subsequent funded evaluation design that cites the standard, which is the definition of success for infrastructure. These impact claims are deliberately modest about attribution, for the gatekeeper reason set out above: the instruments can change what is known and argued, while the political constraints on durable solutions lie outside their reach.
Failure and accountability. Failure deserves equal candor, stated now: weights that fail expert consistency checks but ship anyway; adoption claims that rest on download counts rather than documented decisions; a map that decays unverified; or any drift from measurement toward identity. The project commits to reporting against both lists, to publishing a pre-registered analysis plan for the validation study, and to releasing all instruments, de-identified aggregate data, and code under open access at the end of the grant. A standing do-no-harm review screens any proposed feature that touches personal data against the commitments above before development, and the screening record is public.
11
This report backs the CoLab's submission to the Displaced Livelihoods Initiative in its Infrastructure and Public Goods track: the Livelihoods Information and Access Index as an open measurement public good, validated at Dzaleka. The fit is structural. That track invites a new measure that enables future rigorous evaluations rather than the evaluation of a single program, and the initiative's own call names the lack of information on legal status as a barrier to displaced livelihoods and invites AI-supported measurement in hard-to-reach contexts. The forced-labour framing is not a detour from that agenda but its sharpest edge: strengthening information about work rights and legitimate pathways is a protective factor against the exploitation the Forced Labor Structural Risk Index measures.
The proposal is scoped to research and measurement costs only, with no intervention-delivery lines, and it treats the operational cost engine, the provenance-credential architecture, and the field partner's community access as in-kind contributions rather than billed items. The request sits within the initiative's cap for this track, and the design honors the initiative's guidance that researcher-only designs tend to underperform on policy influence, which is why the field partnership with Fraternidade Sem Fronteiras is written into the plan rather than added to it.
Team and partnership. Sylvia Maier, of the NYU SPS Center for Global Affairs, serves as Principal Investigator and eligibility anchor, since the initiative requires a Principal Investigator with a primary university affiliation and demonstrated experience with rigorous evaluation. Teresa Cantero, a doctoral candidate at Universidad Carlos III de Madrid and a visiting scholar at the NYU Center for Global Affairs, is Primary Researcher; the Summer 2026 cohort carries implementation; and Fraternidade Sem Fronteiras is the field partner.
The partnership is established rather than cold: Yago Rocha volunteered with Fraternidade Sem Fronteiras at Dzaleka on refugee livelihoods and financial inclusion, and Carolina Moron visited Dzaleka while living in Malawi and building a multi-sector data coalition, which is first-hand familiarity that lowers the risk around community access. The same architecture serves adjacent funders, because the measurement-and-governance framing travels. The honest assessment is that the initiative's funded portfolio favors teams with strong evaluation track records, so competitiveness rests on the rigor of the measurement work and the depth of the team's field experience rather than on the breadth of the claims.
12
The work is designed for a term of roughly twenty-four months, beginning within the funder's required window after an award. The sequence moves from re-specifying the indicator framework for livelihoods and legal-rights information, through the expert-elicitation and calibration pipeline for the index, to the reference-context validation at Dzaleka with the field partner, and then to open release of the index, the data standard, the reference implementation, and the methodology, followed by dissemination to the research and practitioner community, including remote and partner-led presentation of results in Malawi.
Because this is an infrastructure-and-public-goods effort, the technology is the deliverable rather than overhead, and a substantial share of the budget maps to the reference implementation and the model adaptation. The cost engine itself is an in-kind contribution, already operational and validated for evacuation, which is why its re-pointing to durable solutions is an adaptation rather than a new build.
13
The limitations are stated together so that they are hard to miss.
The next steps follow from those limits.
The through-line from the evacuation corridor to the camp is a single claim, tested in more than one domain: that the binding constraint on protection, on durable solutions, and on freedom from forced labour alike is the quality, accessibility, and equity of the information that people can act on, and that open, honestly labeled measurement infrastructure is the public good most worth building for it.
References
The Ethical Tech CoLab is a research initiative of the NYU School of Professional Studies Center for Global Affairs, conducted in collaboration with Microsoft. Views and findings are those of the researchers and do not represent the official positions of New York University, Microsoft, UNHCR, WFP, or any partner institution. External programs cited are referenced as evidence, not as CoLab partnerships.
Publications · Instrument and findings
Reading What a Repository Models, and What It Leaves Out
A repository that simulates a population makes a claim about who that population contains, and the claim is legible in its source. This report describes an instrument that reads it, and what the instrument found across 22 repositories: work that models age and disability far more often than income or literacy, an evaluation layer that is strongest exactly where the subject matter is weakest, and eight repositories whose behavioral findings rest on a language model they never name.
22
repositories analyzed, 15 of them the CoLab's own and 7 reference frameworks
62
signals across six axes, each traceable to the file and line it fired on
3 of 22
repositories that model income; 19 model children
8
repositories doing model-driven behavioral work that name no model version
01
When a system generates or evaluates a synthetic population at scale, it is making a claim about who that population contains. The claim is rarely written down. It is implied by which attributes the code carries, which behaviors the agents are allowed to have, and which differences between people the model treats as differences at all. A population dimension nobody models is a population nobody simulates, and an unmodeled dimension is an implicit assertion that it does not matter to the outcome.
That claim is legible in the source. The Agentic Behavior Observatory reads it. Paste a GitHub repository URL and it returns an evidence-linked account of how that repository models agentic behavior, scored on six axes, with the demographic dimensions and the model versions it rests on pulled out. Every point of every score links to the file and the line where its signal fired.
What a score is not. A score is signal coverage, not a quality judgment. A small, sharp repository can and should score lower than a sprawling framework, and the tool says so on its own front page. The number answers one question only: how much of the taxonomy's vocabulary does this repository exhibit. Whether the work is good is a question for a reader, not a regular expression.
02
The analyzer fetches a repository's metadata and file tree from the GitHub API, then reads up to 120 file bodies from the raw content host, chosen by a heuristic that favors manifests, prose, and modeling-core filenames. Text source in roughly 25 extensions counts, including the single-file HTML applications much of the CoLab's work ships as. Generated bundles and lockfiles are skipped. Nothing is cloned and nothing is executed.
Each file is matched against 62 signals grouped into six axes. A signal is a dependency declaration, a file path, or a pattern in source and prose, and it carries a weight and a human-readable label. An axis score is the share of that axis's weighted signal set the repository covers, expressed on a 0 to 100 scale.
| Axis | Signals | What it detects |
|---|---|---|
| Agent-based simulation | 16 | Mesa, AgentPy, PettingZoo, NetLogo, Repast, SimPy; agent classes, schedulers, step loops, spatial environments |
| Synthetic data generation | 10 | SDV, CTGAN, synthcity, Faker, Gretel; population synthesis and IPF, census seeds, differential privacy, generation at scale |
| Model-based behavioral modeling | 10 | Anthropic and OpenAI SDKs, LangGraph, AutoGen, CrewAI, DSPy, local runtimes; personas, memory streams, generative-agent architectures, silicon sampling |
| Reinforcement learning | 8 | Gymnasium, Stable-Baselines3, RLlib, TorchRL; named algorithms, reward machinery, RLHF and DPO, the reset and step contract |
| Evaluation and validation | 8 | Tests, fidelity metrics such as KS, Wasserstein and TSTR, sensitivity analysis and ablations, seeded runs, experiment tracking, bias and representativeness audits |
| Context isolation | 10 | Private and privileged instructions, BATNAs and red lines; visibility and disclosure rules; per-agent scoped context; fresh sessions and state resets; independent replications; staying in character and knowledge-cutoff handling; contamination and leakage checks; blind evaluation |
The headline relevance score uses only the three subject axes, because the other three describe how work is done rather than what it is about.
relevance = 0.6 x max(subject axes) + 0.4 x mean(subject axes)
Subject axes are agent-based simulation, synthetic data generation, and model-based behavioral modeling. The max term keeps a repository that does one of them thoroughly in scope; the mean term rewards work that spans them.
One taxonomy, two runtimes. The scoring logic exists once, in signals.py, and is exported to the browser as JSON. The command-line analyzer that writes the committed corpus and the in-browser analyzer that scores a repository you paste therefore score identically. A visitor can reproduce any published number without installing anything.
03
The committed corpus is 22 repositories: 7 reference frameworks that define the field's vocabulary, and 15 Ethical Tech CoLab projects, mostly evacuation and negotiation simulators, plus the CoLab's agent work and its evaluation tooling. The observatory analyzes itself as well, which is discussed as a limitation below rather than claimed as a validation.
the tool scoring itself
| Repository | Relevance |
|---|---|
| Ethical-Tech-CoLab/agentic-behavior-observatory | 60 |
| microsoft/autogen | 56 |
| camel-ai/camel | 50 |
| mesa/mesa | 48 |
| joonspk-research/generative_agents | 32 |
| Ethical-Tech-CoLab/race-condition-mod | 31 |
| AgentTorch/AgentTorch | 29 |
| Farama-Foundation/PettingZoo | 29 |
| Ethical-Tech-CoLab/ercf | 25 |
| sdv-dev/SDV | 25 |
Eleven of the 22 have model-based behavioral modeling as their strongest subject axis, 8 agent-based simulation, and 3 synthetic data generation. That distribution is a fact about which repositories were chosen, not about the field, and the corpus is small enough that a single addition moves it.
04
Across all 22 repositories the analyzer found 33 distinct demographic dimensions in code and prose. They are not evenly distributed, and the shape of the imbalance is the most substantive finding in this report.
| Dimension | Repositories |
|---|---|
| Children | 19 |
| Age | 15 |
| Disability | 15 |
| Region | 12 |
| Gender | 8 |
| Migration | 7 |
| Education | 6 |
| Race | 6 |
| Income | 3 |
| Religion | 1 |
| Employment | 1 |
| Literacy | 1 |
Populations with bodies but no economics. Nineteen of 22 repositories model children and 15 model disability, both of which change how a person moves in an evacuation. Three model income, one models employment, and one models literacy. In a body of work substantially about who gets out of a disaster and who does not, the attributes that determine whether a household has a car, can afford to leave early, or can read the warning are the ones least often represented. The instrument cannot say whether that is an oversight or a defensible modeling choice. It can say the choice is being made silently in most of these repositories.
Eight repositories name no model. The analyzer extracts every model identifier it finds. Across the corpus it found 42 distinct versions, led by gpt-4o-mini in 7 repositories. Eight repositories doing model-driven behavioral work name no version anywhere the analyzer could read. Behavioral findings drift between model versions, so a result whose model is unrecorded cannot be reproduced later even by the people who produced it.
The corpus is better at isolating contexts than at modelling economies. Context isolation was added to the taxonomy after the findings above, to ask whether a repository has any vocabulary for keeping information where it belongs: one agent out of another's private brief, one run out of the next, the model's own training out of the persona it was given, the test item out of the prompt meant to test it. Median across the corpus is 40. The clearest case is the Diplomatic Simulator at 73, its highest axis by a wide margin and not an accident of vocabulary: every delegation holds private instructions, a BATNA and red lines, and the table carries explicit rules for what crosses between parties. Its own paper states the principle outright. At the other end the Stanford generative agents repository scores 0, which is worth sitting with, because the memory stream is the architecture and nothing in it names a boundary between one agent's stream and another's.
Evaluation is strongest where the subject is weakest. Median evaluation across the corpus is 48, higher than any subject axis: the medians for agent-based simulation, synthetic data generation, and model-based behavioral modeling are 11.5, 7 and 17. AgentTorch scores 84 on evaluation against 29 relevance. The pattern is partly an artifact of the taxonomy, since test files and CI configuration are easy vocabulary to detect, and partly real: it is easier to add a test suite than to model an economy.
The landmark exception. The Stanford generative agents repository, the most cited artifact in this literature, scores 0 on evaluation. It is research code released to accompany a paper rather than a maintained framework, and the paper carried the validation. The score is accurate and would be misread as a verdict by anyone who took it for one, which is the clearest illustration in the corpus of why coverage is not merit.
05
The instrument is a vocabulary detector. Reading it as anything more would reproduce exactly the error it was built to expose.
06
The dashboard runs the analysis in the browser. Two calls to the GitHub API for metadata and the file tree, then file bodies from the raw content host, which is CORS-open and not rate limited: roughly thirty repositories an hour with no token at all. Results stay in the browser and are marked live. Downloading the JSON is how one joins the shared corpus.
The command-line analyzer writes the committed corpus. It has no dependencies beyond the Python standard library, clones nothing, and picks up a GitHub token from the environment or from the gh CLI if one is available.
Teaching it a framework. The taxonomy is one file. Adding a simulation framework is one tuple naming its axis, key, kind, pattern, weight and label. The build step exports the same taxonomy to the browser, so a single edit updates both runtimes and no score can differ between them.
Analyzing the CoLab's own corpus exposed three defects in the taxonomy, all since fixed: HTML and JSX source was not being read at all, the file size cap excluded single-file applications, and the agent-detection patterns assumed agents are called agents rather than members, families, or residents. Any repository analyzed before those fixes scored misleadingly low, which is the argument for versioning a taxonomy the way one versions a model.
References
Publications · Instrument and findings
Reading What a Repository Models, and What It Leaves Out
A repository that simulates a population makes a claim about who that population contains, and the claim is legible in its source. This report describes an instrument that reads it, and what the instrument found across 22 repositories: work that models age and disability far more often than income or literacy, an evaluation layer that is strongest exactly where the subject matter is weakest, and eight repositories whose behavioral findings rest on a language model they never name.
22
repositories analyzed, 15 of them the CoLab's own and 7 reference frameworks
62
signals across six axes, each traceable to the file and line it fired on
3 of 22
repositories that model income; 19 model children
8
repositories doing model-driven behavioral work that name no model version
01
When a system generates or evaluates a synthetic population at scale, it is making a claim about who that population contains. The claim is rarely written down. It is implied by which attributes the code carries, which behaviors the agents are allowed to have, and which differences between people the model treats as differences at all. A population dimension nobody models is a population nobody simulates, and an unmodeled dimension is an implicit assertion that it does not matter to the outcome.
That claim is legible in the source. The Agentic Behavior Observatory reads it. Paste a GitHub repository URL and it returns an evidence-linked account of how that repository models agentic behavior, scored on six axes, with the demographic dimensions and the model versions it rests on pulled out. Every point of every score links to the file and the line where its signal fired.
What a score is not. A score is signal coverage, not a quality judgment. A small, sharp repository can and should score lower than a sprawling framework, and the tool says so on its own front page. The number answers one question only: how much of the taxonomy's vocabulary does this repository exhibit. Whether the work is good is a question for a reader, not a regular expression.
02
The analyzer fetches a repository's metadata and file tree from the GitHub API, then reads up to 120 file bodies from the raw content host, chosen by a heuristic that favors manifests, prose, and modeling-core filenames. Text source in roughly 25 extensions counts, including the single-file HTML applications much of the CoLab's work ships as. Generated bundles and lockfiles are skipped. Nothing is cloned and nothing is executed.
Each file is matched against 62 signals grouped into six axes. A signal is a dependency declaration, a file path, or a pattern in source and prose, and it carries a weight and a human-readable label. An axis score is the share of that axis's weighted signal set the repository covers, expressed on a 0 to 100 scale.
| Axis | Signals | What it detects |
|---|---|---|
| Agent-based simulation | 16 | Mesa, AgentPy, PettingZoo, NetLogo, Repast, SimPy; agent classes, schedulers, step loops, spatial environments |
| Synthetic data generation | 10 | SDV, CTGAN, synthcity, Faker, Gretel; population synthesis and IPF, census seeds, differential privacy, generation at scale |
| Model-based behavioral modeling | 10 | Anthropic and OpenAI SDKs, LangGraph, AutoGen, CrewAI, DSPy, local runtimes; personas, memory streams, generative-agent architectures, silicon sampling |
| Reinforcement learning | 8 | Gymnasium, Stable-Baselines3, RLlib, TorchRL; named algorithms, reward machinery, RLHF and DPO, the reset and step contract |
| Evaluation and validation | 8 | Tests, fidelity metrics such as KS, Wasserstein and TSTR, sensitivity analysis and ablations, seeded runs, experiment tracking, bias and representativeness audits |
| Context isolation | 10 | Private and privileged instructions, BATNAs and red lines; visibility and disclosure rules; per-agent scoped context; fresh sessions and state resets; independent replications; staying in character and knowledge-cutoff handling; contamination and leakage checks; blind evaluation |
The headline relevance score uses only the three subject axes, because the other three describe how work is done rather than what it is about.
relevance = 0.6 x max(subject axes) + 0.4 x mean(subject axes)
Subject axes are agent-based simulation, synthetic data generation, and model-based behavioral modeling. The max term keeps a repository that does one of them thoroughly in scope; the mean term rewards work that spans them.
One taxonomy, two runtimes. The scoring logic exists once, in signals.py, and is exported to the browser as JSON. The command-line analyzer that writes the committed corpus and the in-browser analyzer that scores a repository you paste therefore score identically. A visitor can reproduce any published number without installing anything.
03
The committed corpus is 22 repositories: 7 reference frameworks that define the field's vocabulary, and 15 Ethical Tech CoLab projects, mostly evacuation and negotiation simulators, plus the CoLab's agent work and its evaluation tooling. The observatory analyzes itself as well, which is discussed as a limitation below rather than claimed as a validation.
the tool scoring itself
| Repository | Relevance |
|---|---|
| Ethical-Tech-CoLab/agentic-behavior-observatory | 60 |
| microsoft/autogen | 56 |
| camel-ai/camel | 50 |
| mesa/mesa | 48 |
| joonspk-research/generative_agents | 32 |
| Ethical-Tech-CoLab/race-condition-mod | 31 |
| AgentTorch/AgentTorch | 29 |
| Farama-Foundation/PettingZoo | 29 |
| Ethical-Tech-CoLab/ercf | 25 |
| sdv-dev/SDV | 25 |
Eleven of the 22 have model-based behavioral modeling as their strongest subject axis, 8 agent-based simulation, and 3 synthetic data generation. That distribution is a fact about which repositories were chosen, not about the field, and the corpus is small enough that a single addition moves it.
04
Across all 22 repositories the analyzer found 33 distinct demographic dimensions in code and prose. They are not evenly distributed, and the shape of the imbalance is the most substantive finding in this report.
| Dimension | Repositories |
|---|---|
| Children | 19 |
| Age | 15 |
| Disability | 15 |
| Region | 12 |
| Gender | 8 |
| Migration | 7 |
| Education | 6 |
| Race | 6 |
| Income | 3 |
| Religion | 1 |
| Employment | 1 |
| Literacy | 1 |
Populations with bodies but no economics. Nineteen of 22 repositories model children and 15 model disability, both of which change how a person moves in an evacuation. Three model income, one models employment, and one models literacy. In a body of work substantially about who gets out of a disaster and who does not, the attributes that determine whether a household has a car, can afford to leave early, or can read the warning are the ones least often represented. The instrument cannot say whether that is an oversight or a defensible modeling choice. It can say the choice is being made silently in most of these repositories.
Eight repositories name no model. The analyzer extracts every model identifier it finds. Across the corpus it found 42 distinct versions, led by gpt-4o-mini in 7 repositories. Eight repositories doing model-driven behavioral work name no version anywhere the analyzer could read. Behavioral findings drift between model versions, so a result whose model is unrecorded cannot be reproduced later even by the people who produced it.
The corpus is better at isolating contexts than at modelling economies. Context isolation was added to the taxonomy after the findings above, to ask whether a repository has any vocabulary for keeping information where it belongs: one agent out of another's private brief, one run out of the next, the model's own training out of the persona it was given, the test item out of the prompt meant to test it. Median across the corpus is 40. The clearest case is the Diplomatic Simulator at 73, its highest axis by a wide margin and not an accident of vocabulary: every delegation holds private instructions, a BATNA and red lines, and the table carries explicit rules for what crosses between parties. Its own paper states the principle outright. At the other end the Stanford generative agents repository scores 0, which is worth sitting with, because the memory stream is the architecture and nothing in it names a boundary between one agent's stream and another's.
Evaluation is strongest where the subject is weakest. Median evaluation across the corpus is 48, higher than any subject axis: the medians for agent-based simulation, synthetic data generation, and model-based behavioral modeling are 11.5, 7 and 17. AgentTorch scores 84 on evaluation against 29 relevance. The pattern is partly an artifact of the taxonomy, since test files and CI configuration are easy vocabulary to detect, and partly real: it is easier to add a test suite than to model an economy.
The landmark exception. The Stanford generative agents repository, the most cited artifact in this literature, scores 0 on evaluation. It is research code released to accompany a paper rather than a maintained framework, and the paper carried the validation. The score is accurate and would be misread as a verdict by anyone who took it for one, which is the clearest illustration in the corpus of why coverage is not merit.
05
The instrument is a vocabulary detector. Reading it as anything more would reproduce exactly the error it was built to expose.
06
The dashboard runs the analysis in the browser. Two calls to the GitHub API for metadata and the file tree, then file bodies from the raw content host, which is CORS-open and not rate limited: roughly thirty repositories an hour with no token at all. Results stay in the browser and are marked live. Downloading the JSON is how one joins the shared corpus.
The command-line analyzer writes the committed corpus. It has no dependencies beyond the Python standard library, clones nothing, and picks up a GitHub token from the environment or from the gh CLI if one is available.
Teaching it a framework. The taxonomy is one file. Adding a simulation framework is one tuple naming its axis, key, kind, pattern, weight and label. The build step exports the same taxonomy to the browser, so a single edit updates both runtimes and no score can differ between them.
Analyzing the CoLab's own corpus exposed three defects in the taxonomy, all since fixed: HTML and JSX source was not being read at all, the file size cap excluded single-file applications, and the agent-detection patterns assumed agents are called agents rather than members, families, or residents. Any repository analyzed before those fixes scored misleadingly low, which is the argument for versioning a taxonomy the way one versions a model.
References
