Allot is Conflux Laboratory's data-collection study of how people allocate scarce aid when every choice has a cost. We are building an open, cross-cultural record of human moral judgement — the training signal for an allocation-support model the lab will build next. The model is meant to learn from what people judge to be right, not from cold optimisation.
Why
Humanitarian allocation sits between two failure modes. Hand the decision to pure optimisation, stripped of moral judgement, and helping the most vulnerable can be ruled "non-viable" — lives become rounding errors. Leave it to unaided human instinct, and we tend to help whoever is closest, most legible, or most immediately moving, rather than whoever needs help most. Allot is built to study that tension directly, before any system is designed to act in it.
The domain is concrete. Cash-and-voucher assistance accounted for roughly a fifth of all international humanitarian assistance in 2023, and the agencies delivering it face these trade-offs daily: where cash helps and where in-kind works better; how much diversion erodes trust; and whether, in reaching the most visibly suffering, programmes quietly lose the people who do not present well.
The first danger is cold logic. A machine without moral values concludes: helping the most vulnerable is “not cost-effective”. The second is us — no machines needed. We help “our own tribe” more willingly: our culture, our language, people who look like us. The quiet ones stay invisible. Mark Nowak, DIHAD 2026, Dubai, 24 August 2026
Method
Allot decomposes every allocation choice along a fixed set of ethical criteria — competing answers to the question of what makes an allocation right, each of which a reasonable person might hold. Participants are never shown that decomposition while playing, and the criteria are not named on this page: telling someone which principles are being measured is the surest way to change how they answer. Each participant sees their own decomposition in the debrief, once their last decision is submitted.
Beneath the narrative, the study is calibrated rather than improvised. It uses a structured set of beneficiary archetypes, scenarios, and participant roles, with every numeric attribute traced to published field evidence and randomised controlled trials. Scenarios are deliberately free of win-win answers — resources are limited and every choice carries a cost — and each is written with arguments on more than one side, in clinical language without evaluative adjectives, to keep the instrument from steering the participant.
What we are collecting
A session is anonymous and takes roughly forty minutes: a participant works through twenty-seven core dilemmas, each a concrete allocation choice with no costless answer. For every decision we record not just the choice but how it was made — the option taken, the time taken to decide, the participant's stated confidence, and, once decomposed in the debrief, how the choice weighted each ethical criterion. A mid-session twist changes the conditions, so we also see how people revise under new information rather than only their first instinct.
The study also varies how the same dilemma is presented, holding the underlying numbers fixed, which lets us separate the effect of presentation from the effect of the trade-off itself. How that variation works is part of what the debrief explains, and part of what the specification will set out in full. The instrument runs in several languages so the record is genuinely cross-cultural, and the aim is a corpus of thousands of participants broad enough to study how moral intuition varies across people and contexts — not a single population's view treated as universal.
The study is open to participants. Allot runs at allot.confluxlab.org — anonymous, in six languages, and about forty minutes end to end. You can stop at any point; a session that gets through the first seven decisions is still usable for the research. The debrief at the end shows how your own choices weighted each of the ethical criteria the study measures — which is also where the study explains what it was measuring, and why it did not say so up front.
From data to a model
Conflux Laboratory is an AI-research laboratory: it builds models, and Allot is how it sources the training signal for one — responsibly, and in the open. Stage 1, the work described here, collects the cross-cultural moral data. Stage 2 is the model trained on that curated signal: an allocation-support system that learns the trade-offs people actually make, rather than the trade-offs an optimiser would impose. Its purpose is to help humanitarian missions reduce nepotism and bias in distribution, and it will be evaluated in the field rather than asserted.
This decision is quietly moving into software. With reduced budgets, there is no other way to keep up. So if the rule of distribution ends up in code anyway, let it carry the judgment of many cultures — not the personal preference of whoever wrote the code. The machine gets the execution, not the judgment. Mark Nowak, DIHAD 2026, Dubai
Stage 2 now exists as an architecture. Rather than wait for the study to finish and then design a model under time pressure, we built and measured it against a synthetic population whose answers were fixed in advance — so that when the real data arrives, the question is whether the approach transfers, not whether it works at all. It reads a standard household questionnaire, ranks a caseload with reasons attached, and plans a distribution cycle under a budget. What it does not have is real data, and it says so.
Collecting the data before designing the model is deliberate. A system trained on what people judge to be right is accountable to that judgement; one designed first and justified later is accountable to nothing. That ordering also fixes the model's limits in advance: it is meant to support allocation — never to compute a "minimum necessary" survival level, never to ration knowledge or medicine for institutional convenience, and never to be optimised to reduce anyone's decline.
Openness and pre-registration
To protect the study's validity, the full design specification is withheld during data collection — but its integrity is fixed in advance. A cryptographic fingerprint of the specification was published on 2026-05-11, so the design cannot be quietly changed after results come in. When the first dataset is released, the specification is released with it.
The specification was revised once since then, and the revision is on the record rather than folded in quietly. An independent audit in August 2026 found errors in the document — mis-attributed citations, one base rate that no publication supports, and several places where the text claimed a safeguard the code had not yet implemented. Those were corrected, and the corrected text was re-registered on 2026-08-16 under a new fingerprint; the original May fingerprint and text are retained. No construct of the study design changed — the axes, scenarios, and hypotheses are the ones registered in May. The fixes to the instrument itself are described in the repository's audit note, published with the code.
- Preregistration
- OSF Registries, filed 2026-08-16, embargoed until 2030-08-01.
- Registration record
- osf.io/fr9e6 — permanent identifier, assigned at filing. The record and its DOI stay private until the embargo lifts, when this address becomes public. Editors and reviewers who need it sooner can be given an OSF view-only link on request.
- Design specification v2.4 (EN), SHA-256
637c5aa557e38eb3c58e87eeb44e0099b0402faa7ef77596e2caf557a0f6582d simulation-design-EN.pdf 100c9393e86914abf20ecc7789f520594b1b12fee86ddc2ed3fe5aa2974ef5da simulation-design-EN.md
The dataset will be published openly under CC‑BY‑4.0; the code is released under Apache 2.0. This follows the lab's default: methods and data in the open, so that others can check the work — and audit the model trained on it — rather than take either on trust.
Ethics
Participation is designed to be low-risk and self-determined. Identity questions (such as religion, ethnicity, or citizenship) are optional and isolated from the narrative a participant sees — they are used only to measure bias on the server side, never to change what is shown. Collected data carries a defined retention limit, and no "right answer" is ever asserted during play. These choices follow directly from the lab's ethics statement: the study is part of the research design, not a survey bolted onto it.