Can you defend what you claim about your AI system?

I provide one-time independent audits of consequential AI-system claims for teams preparing to defend those claims to customers, investors, reviewers, or regulators, with falsifiers fixed before results and a sanitized report published with rerunnable receipts.

The claim may concern an agent's reliability, a safeguard's coverage, a benchmark score, a security property, or whether a published result can be reproduced. Most AI audits ask whether a system passes its tests; this one asks whether those tests distinguish the claimed property from plausible alternatives. The result is a bounded verdict with tests, limitations, and rerunnable receipts.

Payment buys the investigation. It does not buy endorsement, certification, or a particular verdict.

Evidence before endorsement

A testimonial asks you to trust someone else's judgment. For an independent auditor, the stronger proof is inspectable prior work: investigations that state a claim, run a discriminating test, report the finding, and preserve the receipt.

When this is useful

This is for a founder, evaluation lead, buyer, or researcher who must act on an AI-system claim and cannot afford to discover later that its evidence measured something else.

What you are buying

Each engagement is quoted at a fixed fee of US$3,000–$5,000. The fee pays for the investigation and does not depend on its verdict. Before work begins, we agree in writing on the claim, available access, disclosure terms, schedule, payment schedule, exact fee, and deliverable.

I then:

1. state the claim precisely, including its scope and assumptions;
2. map what the system and its evaluator can and cannot observe;
3. fix the tests and verdict rules before seeing their results;
4. run the investigation and reproduce every material finding; and
5. deliver a report with the evidence needed to rerun or contest it.

A result may support the bounded claim, falsify it, or leave it unresolved. Negative and null findings are valid outcomes.

Why the result is publishable

An endorsement that the client can purchase is weak evidence. A favorable result is useful precisely because payment cannot determine it; an unfavorable result is useful because the client sees the defect before a customer or reviewer does.

A named, sanitized report is published after an agreed disclosure period. The client may identify secrets and factual errors, but cannot control the findings or suppress the report.

Independence by structure

I do not accept retainers or continued engagements from the audited organization. I do not implement the system I audit, sell remediation afterward, or become its continuing validator. The funding relationship and material use of AI agents are disclosed. Conclusions remain mine.

This is not certification, a compliance opinion, or a comprehensive security audit. It answers one narrower question: does the available evidence warrant the stated claim?