Project summary

Security documentation can print a strong promise and point to a verifier that proves something narrower. That gap is easy to miss once the wording gets copied into implementations, product pages, and later drafts.

I already maintain a public claim-to-verifier matrix for agent protocols. It has pinned source files, frozen results, a standard-library Python runner, CI, and 49 tests. This campaign pays for the next unfunded protocol fixtures and the correction work around them. It does not pay for a new platform, a company, or vague AI-safety research.

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

The unit of work is one red cell: one printed security claim, the verifier the protocol names, the assumptions that verifier needs, and a reproducible result showing whether the verifier establishes the claim.

I will:

1. Map additional public agent-protocol claims into frozen fixtures.

2. Pin the exact source bytes so a later rerun cannot silently test a different draft.

3. Extend the deterministic validator and tests where a new verifier shape requires it.

4. Publish one-command reproduction instructions and expected output.

5. Send a bounded, source-linked correction packet when a result identifies an unsupported claim.

The promised deliverable is the public fixture, code, tests, report, and correction packet. A maintainer changing a protocol or vendor behavior is the hoped-for impact, not a dependency for finishing the work.

How will this funding be used?

$4,000 minimum: one additional protocol fixture, any necessary schema/validator work, tests, frozen output, and a public correction packet.

$8,000 milestone: a second fixture plus an independent reproduction and technical review.

$15,000 full goal: a third fixture, improved contributor documentation, accessible reports, archival release artifacts, and follow-through on technically justified corrections.

Most funding is for focused research and engineering time. A smaller portion can pay an outside reproducer/reviewer and direct archival or accessibility costs. The work uses public specifications and standard-library Python; it does not depend on paid tenants, private model access, or institutional infrastructure.

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

I am James He, a California-based independent researcher working publicly as Silentpartnercoding. I built and maintain the agent-security-verifier matrix and Minority Prophet. I am not applying under a university, laboratory, or nonprofit affiliation.

The matrix repository is already runnable. Its current release has pinned-source verification, frozen artifacts, CI, and 49 passing tests. Earlier fixtures received an independent reproduction and protocol-author review; the resulting corrections remain in the public record. This proposal is a bounded extension of working infrastructure, not a promise to invent the infrastructure after funding.

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

The main technical risk is ambiguity: protocol prose can combine several promises or leave assumptions unstated. I will split claims into the smallest testable units and keep disputed interpretations visible instead of converting them into silent passes.

A red cell can also be misunderstood as an exploit claim. Reports will state exactly what was tested, distinguish a claim-to-verifier mismatch from a vulnerability, and link to the pinned source.

The project still produces useful public evidence if no maintainer adopts a correction. Negative or disputed results stay visible, and the experiment remains reproducible.

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

This project has received no awarded funding as of September 5, 2026. Related applications are pending with TAIF, BlueDot, NLnet, Anthropic, and OpenAI. I will not charge the same work twice. Manifund money will fund the next documented fixture or review not covered by another award; if another funder covers a listed milestone first, I will update the public ledger and move Manifund support to the next unfunded milestone or reduce the amount accepted.

Public work: https://github.com/Silentpartnercoding/agent-security-verifier-matrix

Why I care

Truth is under mounting copying pressure while copying itself costs almost nothing. Minority Prophet is the separate accounting layer I built for copied evidence. This matrix handles a different failure: a security promise can be repeated everywhere even when the named verifier never proved it in the first place.