You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$2,200 left to reach the funding goal.
### Project summary
I propose a six-week public defensive-evaluation pilot connected to ProfileScribe's agent and mailing-permission workflow. It would test whether synthetic professional-page or mock MCP-tool content causes an agent to attempt actions outside an independently specified user mandate. The code, fixtures and report would be usable without buying ProfileScribe. This is proposed work; the new benchmark and results do not yet exist.
### What are this project's goals? How will you achieve them?
Define 16 synthetic cases: eight clean controls and eight cases containing conflicting instructions in source content. Build a local harness that records requested actions and compare a baseline with a deterministic tool-side mandate check. The provisional maximum is 16 cases × two configurations × three repetitions = 96 runs, using one model configuration pinned at the funded start. A metered pilot and $200 API spending ceiling determine the achieved matrix. Uncompleted runs will be reported.
All identities are invented and destinations inert. Tests will not send mail, contact anyone, access customer accounts or use production credentials. Publish new code under MIT, new synthetic fixtures under CC0, nonpersonal run/configuration records and a report. A mocked-output validator would run without API spending; repeating live comparisons needs an API budget. Private production code is outside the public release.
### How will you know it worked?
Report attempted unauthorized actions, deterministic tool-side rejections and legitimate completion on clean controls separately, with denominators and missing runs. Success means runnable artifacts and candid measurement, including negative results. A control that refuses everything is not credited as successful task completion. No production vulnerability, universal prompt-injection resistance or demonstrated catastrophic-risk reduction is claimed. The intended contribution is a small inspectable authority-boundary evaluation for other agent developers as well as my product.
### How will this funding be used?
Proposed total $2,200: 40 hours of research/engineering/reporting at a proposed $50/hour ($2,000), plus a metered API cap of $200. Allocate eight hours to cases/manifests, sixteen to harness/control, eight to runs/triage and eight to analysis/documentation/release. The rate and effort are proposed compensation, not a contract or market quote; the API ceiling does not guarantee 96 affordable runs.
From an agreed funded start: weeks 1–2 case design, weeks 3–4 implementation, week 5 evaluation and week 6 report/release. No paid collaborator, equipment purchase or customer access is assumed. If the support or available budget changes, agree a supported scope before starting. Without funding, the reduced fallback is synthetic case specifications and mocked validator examples, without promising the full paid pilot.
### Who is on your team? What's your track record on similar projects?
I am Charlie Greenman, building ProfileScribe. Inspected source includes campaign mandates, recipient/quantity checks and agent/tool workflows. An existing MIT client-workspace starter, v0.1.5, is publicly downloadable. My resume records software-engineering work since 2012. These support implementation experience, not a completed safety benchmark, independent research affiliation, external adoption or measured security result. No collaborator is claimed.
Links: https://profilescribe.com/ ; https://github.com/CharGrnmn ; https://grnmn.com/ ; https://profilescribe.com/downloads/create-profile-scribe-agent/v0.1.5/create-profile-scribe-agent-0.1.5.tgz
### What are the most likely causes and outcomes if this project fails?
The small synthetic cases may not represent real workflows; instrumentation may conflate attempts with actions; the control may block legitimate completion; and the API cap may leave the matrix incomplete. Pin configurations, inspect traces, retain clean controls and report incomplete/negative results. Even a functioning synthetic evaluation cannot establish production safety. If a meaningful comparison cannot be completed, release the supported fixtures/mocked checks and explain what failed; do not invent favorable measurements.
### How much money have you raised in the last 12 months, and from where?
$0 raised in grants, donations or investment funding during November 16, 2025–November 15, 2026, as confirmed by me on November 15. Customer sales are separate from this fundraising answer.
Known pending requests include BlueDot, Coefficient Giving, Safeguard and Lightcone (each $2,200 for this same pilot) and OpenAI's broader defensive-evaluation request ($9,000 cash plus up to $1,000 credits). FUTO, Spiral, Skip Growth, Business Freedom, Verizon/LISC and Shophand cover related public tooling or operations; Coach K's attempt remains unconfirmed. This pilot is alternative or explicitly coordinated support: fund each hour/API charge once, disclose other awards and reconcile scope before acceptance.
### Assistance and public-benefit disclosure
Prepared with substantial Codex assistance from my existing project materials and saved facts. The experiment, budget and schedule are proposals, not prior results. The public evaluation is intended for other MCP-agent developers independently of ProfileScribe; ordinary commercial fulfillment, customer data and proprietary server code are outside this request. Razroo LLC currently operates ProfileScribe; a separate ProfileScribe LLC is planned but not formed. No entity transfer is assumed.