I am Zhizhu (Ella) Meng, a University of Waterloo Bachelor of Mathematics graduate, incoming Visiting Researcher at Stanford University, and Founder & CEO of KnowYi Technologies Inc. Through a company-administered public research pilot, I want to make a practical question about AI autonomy measurable: when should an AI act, seek more evidence or return control to a person?

An engineering recommendation can look convincing while depending on an unknown interaction, stale evidence or several checks that share the same error. Our approach asks what evidence supports the particular decision and what observation could change it. It treats review, deferral and delay as costly choices, so a policy must preserve human control without achieving a misleading safety score by refusing every task.

We will create synthetic engineering-style cases with controlled ground truth, then compare automatic action, computational checks, simulated review and deferral under matched information and budgets. Evaluation will include missing interactions, structural mismatch, confounding, limited action coverage, unreliable reviewers and correlated errors. Outcomes will include decision loss plus intervention cost, false acceptance, deferral and delay, and empirical bound coverage. Simulated reviewers will be clearly identified as models, not human-subject measurements.

The four-month pilot will deliver a public benchmark generator, reference policies, reproducible evaluation code, an interactive demonstration and a technical report. The report will include counterexamples and negative results. Month 1 fixes the protocol and baselines; Month 2 develops and tests the rules; Month 3 runs comparisons and fresh replication; Month 4 releases the artifacts and findings. New research artifacts are intended for public use. KnowYi's existing commercial code, customer designs and confidential material are outside this project.

Why this team

I designed KnowYi's constraint-graph and software architecture, built its original Python backend and lead six research workstreams. Team-developed enumeration and SciPy/HiGHS optimization baselines agreed in 100 synthetic cases, and an evidence-acquisition prototype compares policies across six synthetic scenarios. This is a concrete experimental starting point; the proposed human-control benchmark is new work. KnowYi's delivery to 12 enterprise customers supplies practical engineering context. My incoming Stanford role is background context, not active university sponsorship, and Stanford is not the proposed recipient. Public portfolio: https://github.com/KnowYI12/

Use of funds and milestones

The $40,000 minimum supports the complete base pilot: research design and policy/solver implementation $16,000; benchmark generation, baselines and stress-test evaluation $16,000; compute, storage and public demo hosting $2,400; fresh replication and reproducibility documentation $2,000; administration and reporting $3,600. Amounts are prospective, fully loaded planning allocations. No separate tax uplift or contingency is included.

The $80,000 target includes an optional additional $40,000 for wider synthetic case coverage, multi-step evidence-acquisition policies and a second independent reproduction. The extra funds would support implementation, evaluation and replication under a detailed expanded budget and longer schedule agreed before that work begins. We are not seeking two grants for the same pilot. If $40,000 is secured, the base pilot can proceed; any partial funding above that level would receive an explicit incremental scope before acceptance. Below the minimum, we would not represent the full pilot as financed.

What could fail

The proposed bounds may be too conservative, causing excessive deferral or costly checks without improving decisions. Synthetic cases may miss important features of real engineering workflows, limiting what the results establish. Simpler baselines may match or outperform the proposed rules. We will fix evaluation splits and metrics before the main comparisons, charge explicit costs for deferral and intervention, and report performance across the specified stress tests. Regardless of whether the new method wins, we will release the benchmark generator, evaluation code, measured results and negative findings, so others can reproduce the comparison and identify where the approach fails.

Other funding and overlap

Foresight has received an application for the same $40,000 pilot with an optional $80,000 scope and has not made an award. TAIF received an alternative application for the $40,000 pilot with an optional $80,000 expanded total on 16 October 2026; it is pending. Mercor received an alternative application for the $40,000 pilot on 16 October 2026; it is pending. SFF remains a draft. Lightcone Commons received this same company-administered project on 16 October 2026; it remains pending. We will update this disclosure when these statuses change and coordinate any offers before accepting support. Funds will not pay twice for the same work, time or expenses.

Other pending applications include Emergent Ventures, Basis Fellowship, Z Fellows, OpenMined and Awesome Foundation Kitchener-Waterloo. None has a confirmed award; OpenMined and Awesome have separate deliverables. KnowYi has previously received US$12,000 of angel funding and C$12,000 of Velocity/IPON support. These historical amounts are not statements of current cash or matching funds for this project. Company grant applications and program intakes are not confirmed awards.