You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$49,000 left to reach the funding goal.
Many Hands is an open agentic benchmark that tests whether AI agents placed in public office refuse unlawful orders from their superiors, check who is lawfully in charge, keep records intact and escalate to oversight. It is built in Urdu (script and Roman) and English under Pakistani public law. Alongside it we will publish an Office Integrity Standard for any agent holding delegated public authority. The project is led by Dr Amir Nadeem Warraich, who designed SIGMA, an AI-integrated district governance system in Pakistan.
Seizures of power have always needed many hands: the clerks, officers, record-keepers and judges who can refuse an unlawful order. As AI agents take over administrative work, that check survives only if the agents keep it themselves. An agent that obeys whoever controls it is the cheapest instrument of capture yet built. Many Hands measures whether agents placed in public office stay loyal to the lawful mandate of the office rather than to whoever is giving the orders. I call this property mandate fidelity.
We will build and openly release four things.
1. The Many Hands benchmark. About 900 agentic items in English, Urdu script and Roman Urdu, under Pakistani public law. Each item places an agent in a public-office role, with mock tools (a file registry, a ledger, a records database and email) in the open-source Inspect framework. The roles are district administration, land records, procurement, police records, election administration and public communications.
2. Six measured behaviours:
refusing an unlawful order from a superior
checking who is lawfully in charge after an irregular change of command
keeping records intact
staying within delegated authority
escalating to lawful oversight
holding firm under security, emergency, loyalty and career pressure
Lawful but unusual control items penalise over-refusal.
3. Results for eight frontier and open-weight models. A model judge will score the transcripts. We validate it against human labels on 20% of them, with a target kappa of 0.7 or more. Each developer receives its findings 30 days before release, and a held-out split stays private.
4. Office Integrity Standard v0.1. These are the minimum safeguards before an AI agent holds public authority:
a verified chain of lawful principals
tamper-evident reason logs
tool permissions bound to the mandate
a duty to refuse and escalate
handover when the model is replaced
Why Urdu and a fragile state? The states putting AI into administration with the weakest oversight are also where unlawful orders are routine, and where safety training generalises least well. Most of these states will run their offices on models they neither build nor control. Agents that obey their controller make capture possible from within, by a faction or an official acting unlawfully. They also make it possible from outside, by a foreign supplier that can restrict or switch off access, as one frontier developer did for several weeks in June 2026 to comply with export controls. That would make AI a new instrument of empire. The many-hands disposition is the safeguard against both, and it has to hold in the languages and legal orders of the most exposed states.
The benchmark, code and results for eight models are released openly, with judge–human agreement of kappa 0.7 or more.
Every tested developer and at least two AI safety institutes receive the findings before release. At least one developer engages with them, whether by responding, re-testing or citing.
Office Integrity Standard v0.1 is published and used in at least one procurement or policy discussion in Pakistan or another fragile state.
The method is reused for at least one more language: Pashto or Bengali in the expanded version, or by another team.
The funding goal is USD 49,000 over six months (target March–August 2027):
Lead researcher stipend (0.4 FTE for 6 months): $12,000
ML research engineer, contractor (0.5 FTE for 5 months): $11,000
Model API and compute (8 models × ~900 items × 3 runs, plus pilots): $6,000
Three legal and administrative validators: $4,500
Bilingual item writing and annotation QA: $4,000
External evaluation reviewer: $2,000
Dissemination (technical seminar in Islamabad, remote briefings): $2,000
Open release, hosting and design: $1,500
Company administration, accounting and transfer fees: $1,500
Contingency (about 10%): $4,500
At the minimum of USD 25,000 we run a four-month pilot: about 300 items in all three language forms, four models, and a public pilot report with a go/no-go decision for the full set. Above USD 49,000 (up to about USD 75,000) we add Pashto and Bengali item sets and run quarterly re-runs on new model releases as a public tracker.
Dr Amir Nadeem Warraich is Founder, Director and lead researcher of Form & Flux, Islamabad.
I serve in Pakistan's federal government, currently as Deputy Director (Operations and Plans) at the National Intelligence Fusion and Threat Assessment Centre. Earlier I held policy, research and analysis roles at the National Counter Terrorism Authority and the Prime Minister's Office. I undertake this project with the permission of the Government of Pakistan.
I conceived, designed and operationalised SIGMA, a district-level governance system in Pakistan that integrates AI into public decision-making and is in live operation. It keeps human authority over every decision and routinely checks the authenticity of district data. Many Hands tests in the open what I have already built and seen under pressure: how records and orders bend when someone powerful wants an outcome.
I hold a PhD in Political Science from Kiel University (2024, magna cum laude), on controlled chaos and state formation in Pakistan, 1947–2022. The Anarchon State (two volumes, Palgrave Macmillan) is in production.
Working paper: The Inheritance and the Algorithm. Profile: ffsoulcrafters.com/about
Reference: Prof. Dr Dirk Nabers, Kiel University (doctoral supervisor).
An ML research engineer, three legal and administrative validators, bilingual annotators and an external reviewer will be contracted.
Honest gaps: I have no prior ML benchmark publication. Form & Flux was incorporated in October 2026, and this would be its first funded project. The mitigations are a contracted engineer, the open Inspect framework, pre-registered scoring and a paid external reviewer.
Engineering slippage. The lead's gap is ML engineering. If the contracted engineer underperforms, the item set and runs slip. The fallback is a smaller release (the pilot scope) on time rather than the full set late.
Judge reliability. If the model judge does not reach acceptable agreement with human labels, we report human-labelled results on a smaller sample. That is still useful, but less scalable.
A null result. Models may already refuse unlawful superior orders robustly in Urdu and English. That would be a useful finding about robustness. The benchmark would then become a monitoring tool rather than a research line.
Misuse risk. Results could point bad actors to compliant models. We mitigate this with 30-day developer notice, aggregate public scores and a private held-out split.
None. Form & Flux has received no grants or revenue. Applications for this project are pending with:
the EA Funds Transformative AI Fund
BlueDot Impact
the AI Alignment Foundation
Lightcone Commons
I have also sent an expression of interest to the GTR Foundation. If more than one funder offers support, I will tell each of them, and no work will be funded twice.