Project summary

I'm Michael Lip, a Polish solo developer. I built 108 Chrome extensions used by 150,000+ people. Now I need compute to build faster.

I use Claude Code and automated Playwright checks in the Zovo build pipeline. For this grant, I want to test a specific question: when an agent reaches a budget limit or fails a quality check, does the surrounding software actually stop it?

I propose a public ml0x evaluation package for coding-agent budget limits and quality gates. I would publish the test code under the MIT licence, synthetic examples, setup instructions and a report that includes failures. This is a proposed project, not a claim that these results already exist. Grant spending would support this public work, not extension sales or private product features.

What are this project's goals? How will you achieve them?

The goal is to give other developers tests they can run when deciding how to constrain coding agents. I would test budget exhaustion, repeated retries, failed quality checks and attempts to continue after a stop condition. Each case would state the expected behavior, observed result and cost.

At the full $5,000 goal, I would work part-time for eight weeks after funding is available. In weeks 1 and 2, I would define synthetic tasks and expected stop behavior. In weeks 3 to 5, I would build the runner and compare runs with and without the controls. In weeks 6 to 8, I would rerun failed cases and publish the code, instructions and report.

Success means another developer can run the examples and inspect the results. I do not assume an improvement percentage. These controls may support safer agent use, but they do not establish alignment or prove a reduction in catastrophic AI risk.

How will this funding be used?

The full budget is $5,000: $3,000 for 100 hours of my engineering and documentation time at $30 per hour; $1,500 for model API and test compute; and $500 for test hosting and release checks.

The minimum is $2,000 for a four-week subset covering budget exhaustion and retry limits. That budget is $1,200 for 40 hours of work, $600 for API compute and $200 for hosting and release checks. The code, examples and limits would still be public. Between the minimum and full goal, I would add quality-gate cases and repeat runs as the budget allows.

These are estimates. I would track actual spending. Any usable credits would reduce the cash charged for the same compute. I would discuss unused funds or material changes with Manifund. No hiring is planned.

Who is on your team? What's your track record on similar projects?

I would do the work alone. My track record is shipping software and operating an agent-assisted build and QA pipeline. I do not claim AI safety research credentials or an independent validation of the guardrails.

Zovo is my software product studio. I am applying as an individual for this public project.

Website: https://zovo.one
GitHub: https://github.com/theluckystrike
LinkedIn: https://www.linkedin.com/in/chromeextensions/
Contact: support@zovo.one

What are the most likely causes and outcomes if this project fails?

The tests could be too narrow to reflect real agent behavior. Model updates could make results hard to reproduce. The controls could fail to stop retries or work after a failed check. Setup could also be too difficult for another developer to use.

I would reduce these risks with synthetic tasks, recorded model versions and settings, repeated runs, and clear setup instructions. If the controls fail, the useful result would be a public account of those failures and the conditions that caused them. I would not present passing toy tests as proof that an agent is safe. If the minimum is not funded, this grant-funded scope would not start.

How much money have you raised in the last 12 months, and from where?

Zovo has raised $0 from investors. My existing development work is self-funded. I am not presenting pending grant requests as secured money for this package.

I have submitted a $5,000 request to Lightcone Commons for this same eight-week scope. The funding hub also records a submitted $15,000 request to EA Funds for its AI fund for a broader six-month ml0x guardrail project. This package overlaps that broader work. I would disclose any award before accepting funds and reduce or withdraw overlapping requests. The same costs would be funded once.

The proposal seeks a non-dilutive grant. The fundraising deadline is November 20, 2026. The work schedule starts after funding is available.