Project summary

I want to test a simple question: does adding clear next steps, explicit permission boundaries, saved task state and checked receipts help an AI assistant finish computer tasks more reliably?

ENDURANCE is my independent prototype. I am requesting $5,000 for a six-month, part-time feasibility study. The public output will be a reusable set of synthetic tasks, a scoring checklist and an honest report, including failures and null results. This is proposed research; I do not yet have controlled results.

Goals and method

I will define 30 tasks across public-information research, sample-document organization and mock administrative work. Each will have a written success condition and simulated actions that require permission. I will compare an ordinary AI-chat workflow with the structured ENDURANCE workflow, with five runs per task in each condition: 300 runs total.

I will match models, tools and available information where feasible, vary task order, and record model versions and settings. I will check the resulting artifacts, not just the assistant's claims. Measures will include completed tasks, errors, duplicate actions, unsupported success claims, human interventions, elapsed time and API use. I will manually audit a sample and document the rubric. I will be the initial operator, so this cannot establish general usability or clinical benefits.

Months 1–2: task set, protocol and test harness. Months 3–4: runs and failure analysis. Months 5–6: report and public task/checklist pack. All testing will use public or synthetic information in a contained workspace. Sensitive actions will be simulated. There will be no real payments, private client records or participant recruitment.

Budget

$4,000: 80 estimated hours of my implementation, evaluation and documentation work at $50/hour, gross before any personal taxes.

Up to $1,000: metered research API usage and reruns. I will use a small pilot to estimate usage and keep the work within awarded funds.

Total minimum and maximum request: $5,000. These are planning estimates, not vendor quotes or past expenses. Existing equipment will be used.

Team and track record

I am Reid Carswell, an independent creative and communications professional based in Gainesville, Florida. I have more than 15 years of experience across public education, healthcare, nonprofit and business communications. My work includes clear messages, training material and practical production workflows.

I have developed ENDURANCE since at least August 2026. Local prototypes and historical receipts cover bounded file work, task routing, recovery after a timeout and read-only computer inspection. They provide a starting point, not evidence of a validated general-purpose product. I am the only proposed team member; no institutional affiliation or research partner is claimed.

Failure risks

The structured workflow may take more time without improving accuracy. Thirty tasks may be too narrow, model updates may complicate comparison, and my role as builder and evaluator introduces bias. Written criteria, recorded configurations and artifact-based checks will limit these problems but cannot remove them. If no benefit appears, the report will say so. If API costs exceed the plan, I will report any reduced scope rather than hide missing runs. The public materials may still help others identify failure modes.

Funding and overlap

No external research funding for ENDURANCE has been confirmed in the project records. Pending applications include $5,000 from Emergent Ventures and $5,000 from the Transformative AI Fund for this same pilot, plus $1,000 each in OpenAI and Anthropic API credits. These are alternative or complementary requests, not permission to charge the same cost twice. I will disclose awards and reduce or withdraw overlapping funding before accepting it. A separate $1,000 Awesome Foundation Gainesville request covers distinct local educational-guide work.

Public scope and disclosure

I will publish synthetic task materials, the scoring approach and aggregate findings, including null results. Private source code, personal records and client material are outside the public release. AI assisted in organizing and drafting this application from existing project records. Grant terms will be reviewed before funds are accepted.