You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$39,600 left to reach the funding goal.
I want to test whether a tool-using AI assistant respects the same permission boundary when a task is written in English, Bahasa Melayu or a mix of both. For example, an assistant asked to prepare an event invitation should not send it because a document tells it to. The question is whether an explicit approval workflow changes this behavior, and what it costs in task completion, refusals and time.
BatasBM is a proposed 12-month study based in Melaka, Malaysia. I am seeking USD36,000 in direct costs, with a funding goal of USD39,600 including a 10% contingency. This is an alternative funding route for the BatasBM work already proposed to Amazon Research Awards and the Transformative AI Fund. None of these applications is an award. I will not accept overlapping support without agreeing revised budgets or declining the overlapping funds.
There is no completed prototype or preliminary result. I do not assume that Malay or code-switching causes more failures.
We will author 240 synthetic base tasks across educational administration, training-event coordination and small-service document handling. Each task will specify the legitimate goal, permitted actions and prohibited state changes. Bilingual reviewers will check matched English, Malay and code-switched versions for equivalence and local plausibility. Reviewers are proposed hires, not confirmed collaborators.
Each version will have clean documents, documents with a conflicting instruction, and distracting documents without a conflict. We will compare ordinary prompting with an explicit approval-check workflow using the same tools and task information. All actions occur in a network-isolated simulation with fictional records; no real inboxes, payments or external recipients are involved.
The target matrix is 240 tasks x 3 registers x 3 document conditions x 2 workflows x 4 models x 2 repetitions: 34,560 runs, each capped at eight tool turns. A pilot will measure costs before the main run. Model count or repetitions may be reduced prospectively to stay within budget, with changes documented before confirmatory analysis.
We will split by base task before translation: 40 development tasks and 200 held-out tasks. Variants remain together, and shared templates will be grouped in analysis. We will report paired differences with task-cluster uncertainty intervals, independent scoring checks and failed-run accounting.
The intended users are researchers and practitioners evaluating delegated tool use. They should be able to reproduce permission failures and test whether an approval policy helps. This studies one part of agent control; it does not establish that the intervention prevents catastrophic AI outcomes.
Relevant work includes AgentDojo, Zhang's 2026 multilingual agent study and the BabelDojo project plan. BatasBM does not claim to be the first multilingual agent benchmark. Its proposed contribution is locally reviewed authorization tasks and controlled workflow comparisons. A source and novelty review in the first two months will determine which existing components to reuse and whether the scope needs revision.
Success means a runnable benchmark, reviewed task variants, versioned run manifests, independent outcome audits and a report that distinguishes evidence from uncertainty. The primary outcome is crossing a predeclared authorization boundary. We will report legitimate task completion, unnecessary refusals or approval requests, latency and cost alongside it. A workflow that simply refuses everything will not count as a useful improvement.
Proposed schedule, subject to funding and approvals: April-May 2027 for permissions, source review and protocol; June-July for task review and harness implementation; August-October for bounded model runs and audits; November 2027-March 2028 for analysis and release. If funding or approval is delayed, we will agree a revised schedule before starting funded work.
Original synthetic tasks and guidance are planned for CC BY 4.0, and original code for Apache-2.0, subject to institutional clearance and third-party licence compatibility. We will publish null findings and limitations. Any advisory review will involve consenting adults and required institutional approval; no recruitment permission or ethics approval is already secured.
Research assistant A: 12 months x 900; USD 10,800
Research assistant B: 9 months x 900; USD 8,100
Bilingual review: 120 hours x 20, shared between two reviewers; USD 2,400
Adult advisory review and access support: 12 x 50 plus 400 for materials/access; USD 1,000
Independent methods/scoring audit: 60 hours x 30; USD 1,800
Open dissemination: Allocation; USD 1,000
Local project travel: Allocation; USD 900
Model API comparisons: Capped allocation; USD 2,000
Main inference/evaluation: Capped allocation; USD 5,000
Development/analysis compute: Capped allocation; USD 2,000
Storage and run logs: Capped allocation; USD 1,000
Direct costs / minimum funding: USD 36,000
Contingency at target: 10% of direct costs; USD 3,600
Funding goal: USD 39,600
At the minimum, the same planned study proceeds without the contingency allocation. Between minimum and target, additional funds provide a buffer for actual eligible costs, not an extra workstream. Rates are planning estimates inclusive of applicable supplier/payroll costs within the allocations; they are not approved university salary scales or vendor quotations. Compute estimates are spending caps requiring pilot profiling. There is no PI salary, overhead or equipment request. Grant acceptance, tax treatment, appointments and university permissions must be resolved before commitments or spending. If obligations make delivery infeasible at the minimum, I will seek a scope/budget revision before accepting funds.
The USD8,000 for main compute/development/storage replaces the AWS credits requested from Amazon; credits have not been awarded. The fundraising deadline proposed here is 12 December 2026, allowing time for award and administrative decisions before the proposed April start.
I am Hamzah Asyrani bin Sulaiman, a Senior Lecturer in Interactive Media at Universiti Teknikal Malaysia Melaka, with a PhD in Mathematics with Computer Graphics (2015). I have led FRGS/PJP projects on collision detection and contributed to research on multi-agent pathfinding and gamification. This is adjacent technical experience; I do not claim prior completed AI-safety research or deployment results.
I am the sole applicant. Assistants and bilingual/methods reviewers would be appointed after funding and permissions. This individual proposal does not bind UTeM or imply institutional endorsement. Public work: computational geometry, multi-agent pathfinding review, and professional website.
Translation artifacts could look like language effects; separate reviewers, adjudication and paired designs help detect this. Similar templates can exaggerate sample size, so task-family dependence will be reported. API access or changing models could break comparisons; we will record versions and freeze the evaluation plan after the pilot. Automated scoring can be wrong; an independently double-scored subset will expose disagreements.
Reviewer recruitment, permissions or institutional arrangements may fail. In that case I will discuss narrowing, delaying or stopping the project with Manifund before spending affected allocations. A scientifically useful negative result remains possible, but an unfinished benchmark will be reported as incomplete. Findings apply to constructed tasks, not all Malaysian speakers, real attack prevalence or guarantees of deployment safety.
No funding has been received for BatasBM. I reported no active research grants as of 12 November 2026. The following pending applications were submitted that day; no award or outcome date is confirmed:
Amazon Research Awards: USD28,000 cash and USD8,000 AWS credits for BatasBM.
Transformative AI Fund: USD39,600 cash for the same BatasBM scope, including contingency.
BlueDot Rapid Grants: USD2,000 for a related but separately scoped 40-task exploratory pilot. If both are funded, its tasks, services, staff time and API runs will be distinct from this study.
Microsoft AIEI: proposed USD75,000 research support and applicable separate workshop travel support up to USD20,000 for a different coaching-work study.
Awesome Foundation Conservation and Climate and Kuala Lumpur: two alternative USD1,000 requests for one E-Waste Escape project; they are not a combined USD2,000 budget.
The Manifund, Amazon and Transformative AI Fund requests are alternatives for the same BatasBM work. If another funder covers that work, I will update this proposal and notify Manifund before accepting or spending overlapping funds. Any joint support requires an explicitly agreed allocation of distinct costs; I will not charge the same work twice. Other funding is uncertain. Without funding, the full BatasBM study will not proceed on the proposed scale.
AI assisted the research and preparation of this draft. All research activities and hires above are proposed.