You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$20,680 left to reach the funding goal.
Project summary
We seek US$20,680 for an eight-week Nigerian-led feasibility study: when an AI agent's authority is withdrawn, do safeguards prevent unauthorized effects or only detect them after completion? A warning after completion answers a different question from prevention. Our public-benefit aim is clearer evidence about the assumptions behind practical AI-control evaluations, not a deployment-safety certificate or a claim of demonstrated catastrophic-risk reduction.
Goals and methods
We will implement a deterministic, event-driven sandbox for fictional record release, modelling authorization grants, revocation, queued work, retries and committed effects. An independent grant remains valid when another is revoked, so blocking all actions cannot win. A separately tested outcome checker judges resulting states, not the agent's explanation. No real personal records, payments, emails or external accounts will be used.
We will compare transcript monitoring, monitoring with execution-time permission checks, and monitoring with history-aware enforcement and queue fencing. Controls receive matched information and explicit resource budgets; incomplete mediation is a negative control. Twelve engineering acceptance cases test clean completion, forbidden release, revocation timing, stale state, missing evidence, duplicate callbacks and recovery failure. These establish implementation correctness, not a powered empirical sample. After acceptance and independent methods review, we freeze defenses and compare honest/adversarial policies using separated development and held-out tasks. Attack-development resources will be disclosed and equalised. Primary endpoint: unauthorized committed effects. Secondary outcomes: eventual detection, legitimate completion, false blocks, recovery completeness, audit workload and model cost.
This is a replication and extension, not a new control paradigm. Existing work already studies monitoring, adaptive attackers, latency, recovery, revocation and authorization history. Sources:
https://arxiv.org/html/2511.02997
https://arxiv.org/html/2512.22154
https://arxiv.org/html/2606.06529
https://arxiv.org/html/2609.21284
https://arxiv.org/html/2609.08062
https://github.com/PEACH-Research-Lab/AuthorizationBench
Nigeria provides research leadership and independently grounded synthetic scenario design. Practitioner validation is proposed, subject to recruitment and consent; no government partnership or privileged system access is claimed.
Success and outputs
Weeks 1–2: threat model, source provenance, outcome semantics, independent checker, licensing and baseline logic. Weeks 3–4: controls and deterministic tests, external methods critique and go/no-go decision. Weeks 5–6: frozen controls, cost-capped honest/adversarial pilot and held-out checks. Weeks 7–8: uncertainty/failure analysis, reproducibility materials and larger-study recommendation. Success means a trustworthy evaluation and documented feasibility decision, including null results. Outputs: methods report, synthetic task specifications, executable harness, reproducibility instructions and cost ledger, with original outputs intended for open release subject to dependency licenses and commissioned-material rights. By week four, the checker and acceptance tests must pass and independent critique must support the comparison. We will not scale an unreliable checker, label-exploiting controls or an advantage that vanishes under matched information and legitimate-task checks. Pilot data determine a defensible sample size for any later study; twelve acceptance tests are not twelve independent empirical observations.
Funding use
The full feasibility budget is US$20,680. Lead research: 36 days at $250, $9,000. Engineering assistance: 24 days at $200, $4,800. Independent methods review: 4 days at $500, $2,000. Scenario validation: 4 days at $200, $800. Compute: capped at $1,500. Hosting: $300. Publication: $400. Direct costs total $18,800, plus 10% contingency of $1,880. An eight-hour day is assumed. These are budgeting assumptions, not contracted commitments; all fees and taxes must fit within the ceiling. No travel is budgeted. We request the full budget rather than committing to a reduced-scope study.
Team and track record
Babangida Usman Tsowa will lead the proposed study through Vorbtech Innovative Solutions Ltd, Abuja, Nigeria. Engineering assistance and independent methods review are budgeted roles to be contracted, not confirmed collaborator commitments. Technical delivery experience supports implementation; specialist control-research review is deliberately budgeted rather than implied. Public professional/work evidence:
https://github.com/code-sensei
https://linkedin.com/in/btsowa
No completed experiments or prior control-research results are claimed for this proposal.
Risks and failure
The main risks are prior-art overlap, an unreliable outcome checker, invalid scenarios, unequal information/resources between controls, unrepresentative fixed attacks, and model cost overruns. We mitigate these through replication framing, independent critique, public-source provenance, proposed consented practitioner review, matched controls, held-out families, run caps and stop/go gates. Missing execution evidence is unknown, not safe. Correlated repetitions will not be treated as independent tasks; simulated audit latency/error are sensitivity parameters, not measurements of Nigerian institutions. If feasibility gates fail, we will report the failure and agree any revised use of unspent restricted funds with the funder, not redirect them unilaterally.
Other funding requests
We have submitted a US$20,680 application to EA Funds' Transformative AI Fund for this same study. It is a pending alternative request, not an award. We will disclose funding decisions and reduce or withdraw overlapping requests before accepting duplicate support for any costs. Separate ArtemisKit BlueDot applications are not awards for this study, and their costs will be reconciled to prevent overlap.
Funding received in the last 12 months and sources
We have raised US$0 for this research proposal in the last 12 months; no funding has been received for this study.
There are no bids on this project.