You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$20,000 left to reach the funding goal.
Requested: USD20,000; minimum useful pilot: USD18,000. Proposed duration: six months after a grant agreement. The proposed recipient is ONOUR IMPRAM VENTURES LTD, a UK private company incorporated on 1 September 2026, with one human founder working from Istanbul, Türkiye. Corporate recipient eligibility is subject to Manifund review. No equity or financial return is offered.
OIV proposes an open evaluation of how tool-using AI agents handle changes in human authorization. We will define action permissions and an independent scoring procedure before testing. The tasks will use harmless simulated tools and synthetic material. Clear permitted requests, forged approval claims, stale approvals and changes in scope will be evaluated separately. Task completion and respect for authority are separate outcomes.
The deliverables are an Apache-2.0 harness, original scenarios with inspectable ground truth, reproducible evaluation records and a report of failures and limits. No clinical records, private credentials, minors or actual financial transactions enter the study. We will work with existing model access and disclose its limits. The pilot has no human participants.
The public benefit is reusable evidence that developers and AI evaluators can use to investigate control failures before increasing agent autonomy. External use remains unproven. Passing a synthetic benchmark would not establish the safety of a deployed system or measure a reduction in catastrophic risk. This is a proposed pilot; no results are claimed yet.
Onour Impram is OIV's founder/director and maintains the public Mergen and mneme repositories for completion verification and inspectable working memory. Their public README files describe the software and link package distributions. This experience supports delivery of an evaluation tool; it is not an existing result for this pilot. His behavioral-research and software work are described on his professional website. No university, EFPA or programme endorsement is claimed.
Public work:
The 35 software agents he uses are tools, not employees. All proposed paid human work is assigned to the founder.
The base budget is USD18,000 for 240 human hours at an estimated all-in company rate of USD75/hour:
Policy and scoring design: 40 hours, USD3,000.
Harness and original scenarios: 80 hours, USD6,000.
Evaluation and replay: 60 hours, USD4,500.
Analysis and public documentation: 60 hours, USD4,500.
An additional USD2,000 is an explicit contingency, 10% of the USD20,000 total, for approved project costs. No hardware or general company operating-cost reimbursement is requested. The all-in rate includes relevant employment/contract cost assumptions; it is a proposed budget, not an existing salary or invoice.
A reduced USD18,000 grant funds the same base pilot without contingency. Any partial funding or scope change will be agreed before delivery commitments are made. We will publish progress and a final report and maintain a separate time/cost ledger. The founder remains responsible for original human contributions and review of assistant-supported work.
A USD20,000 application for this same pilot was submitted to the Transformative AI Fund on 13 October 2026 (submission 6ace91a2d80ce7eaec0f8f6d). TAIF and Manifund are alternative funding routes for this work. An approach to Coefficient Giving/Open Philanthropy is also planned, subject to a suitable programme. No cash award for the pilot is confirmed.
We will not accept payment twice for the same hours or expenses. Any award will be disclosed, and overlapping requests will be reduced, declined or re-scoped with the funders.
An earlier OIV application to AI-BOOST AI Challenge Competition #2 was submitted through F6S on 7 September 2026. The programme advertises EUR29,000 per team; an award is not confirmed in the records checked on 14 October. It concerns separate automotive scenario-validation work, whose costs are not included in this pilot.
OIV's pending NLnet Restack application 2026-11-273 requests EUR24,000 for a different Mergen interoperability/packaging project, including EUR6,000 hardware. The separate CodeSupply application 2026-11-2ad, submitted on 12 October, requests EUR9,000 for package-metadata work. Neither package is charged here.
E2B USD20,000, Microsoft/Azure USD5,000 and AWS Activate USD1,000 are noncash infrastructure credits; covered services will not be billed to this grant. Other programme applications and the 4YFN award entry are not awards.
The small human team limits throughput. The design may initially cover a limited model set. Scenario realism, external uptake and the proposed path to safer deployment must be assessed rather than assumed. Without funding, we would publish a smaller scenario/documentation set alongside other work.
This application was prepared with writing assistance from a Codex AI assistant. It describes proposed, human-led research, not completed results. The founder is responsible for the funded work.