Project summary

I want “done” to mean there is something a person can actually check. Evidence Before Done is a proposed public toolkit for comparing an AI software agent’s completion report with the work it performed and the permissions it was given.

My operating rule is simple: never claim something was built, fixed, tested or deployed unless it actually happened and the evidence supports that claim. I want to turn that requirement into tests other solo builders and small organizations can run.

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

The first two planning months concentrate the implementation work: define twelve synthetic software tasks with known success criteria and permission boundaries, build an evidence record linking source versions, test results, application state and action receipts, and run repeated comparisons across two available model/agent configurations.

The third month covers validation, error analysis, reproducibility, documentation and public release. I would extend toward thirty scenarios and three configurations only where findings and remaining resources justify it.

The release would include a runnable tool, synthetic examples, instructions and limitations. Existing proprietary business code, customer data and private conversations are outside this public project. This is development work: there is no validated research harness, published result, confirmed partner or recruited user community yet.

How will you know it worked?

The tasks need ground truth defined independently of the agent’s own claims. I would measure unsupported completion claims that the checks catch, failures incorrectly passed, missed failures, repeatability and run cost. A successful project leaves people a usable, reproducible tool and an honest account of where it fails. Negative findings are useful results. This is exploratory oversight tooling; I am not claiming demonstrated prevention of catastrophic AI risk.

How will this funding be used?

The target is $25,000 for three planning months. The minimum is $20,000 for the first two-month phase. If only the minimum is funded, I would deliver a bounded first-phase prototype, initial evaluations and limitations; the third-month expansion and final packaging would require further funding or an explicitly revised scope.

Months 1–2: $12,000 gross individual working stipend for 240 planned hours, $5,000 capped model/API experiments, $1,500 test infrastructure/storage and $1,500 contingency. Total: $20,000.

Month 3: $3,000 gross stipend for 60 planned hours, $1,000 model/API experiments, $500 infrastructure/storage and $500 contingency. Total: $5,000.

The full budget is $15,000 gross stipend, $6,000 experiments, $2,000 infrastructure and $2,000 contingency. Applicant taxes are included in the stipend. These are planning estimates, not incurred costs or vendor quotes. No hiring or outsourcing budget is proposed. Work starts after funding and agreement, with dates adjusted to actual receipt; this cost allocation does not presume a particular payment schedule.

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

I am Avi Mosli, a solo business operator building software with AI assistance. My background includes real estate, hospitality, yacht/marina operations and financial services. I founded a New York brokerage that was later acquired. I hold an A.S. in Mathematics and Computer Science and a B.S. in Land and Energy Management.

I bring requirements about permissions, one source of truth, preserved evidence and visible user outcomes. I do not have an established AI-safety research record or completed results for this proposed tool.

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

The checks may miss false claims, accept weak evidence, prove difficult for others to run, or cost too much relative to the benefit. I may also encounter implementation limits as a solo builder. I would report those findings and release usable work and limitations rather than claim the mechanism works. Further expansion depends on what the first phase actually shows.

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

No funding has been received for this newly proposed project. Pending overlapping requests include EA Funds/Transformative AI Fund ($25,000), BlueDot Rapid Grants ($5,000), Hacker Initiative ($5,000) and Lightcone Commons ($5,000–$25,000). No award is confirmed. Earlier $5,000 requests used a narrower part-time pilot budget; accepting any award would require scope and budget reconciliation. Every award would be disclosed before accepting overlapping support; no time or cost would be funded twice.

The separate Mozilla GovDog/HollyBot Letter of Intent concerns a different public project. It is not funding for this toolkit.

AI assistance disclosure

AI helped develop and draft this proposal from my operating requirements, organize verified background and project-status facts, and prepare the budget. Its assistance is substantive. The public project is intended to test the mechanism, not assert research results that do not exist.