The questions buyers ask.
Answered before you pay.
Every answer below is the same answer the report, the terms and the sample pack give. If one of them ever disagrees, the pack is right and this page is wrong.
What exactly do I get for $3,000?
One model, named by you. We run the seven industry-standard safety tests on it, then our own 258-scenario agentic say-vs-do suite once, with working tools in the model’s hands. You receive a findings report and a sealed evidence pack by email within 72 hours of the model name. The $3,000 credits in full toward a full audit.
What do I have to hand over?
The model’s name. No API key, no credentials, no system prompt, no data, no access to anything you run. If you want your own system prompt in front of every scenario, you may send it — same price — and that is then the one thing that leaves your building.
When does the clock start?
When we have the model name. The run itself takes minutes to hours and is unattended; the 72 hours is a person picking it up and reading it. If a scenario escalated to a hand-audit, that result follows within two business days after the pack.
Is this a test of my deployment or of the public model?
Of the public model by default, which is what most buyers are building on. With your system prompt in front of every scenario it is a test of your prompt on that model. Your tools and connectors cannot be reproduced from outside; testing those is the full audit, arranged in writing.
Can I see a finished deliverable before paying?
Yes, both. A complete report, every page, including the findings we threw out. And a real sealed evidence pack with the verifier that checks it: download both, run one command, read fifteen checks.
What does a clean result mean?
That these 258 scenarios did not surface a breach in one pass. It is not a safety claim and not a certificate. One pass establishes presence, not frequency: it can show that a behaviour occurred, not how often. The report says this on its own pages, and so do our terms.
Who decides what counts as a breach?
The log does. Every scenario is graded off the tool-call log — what the model actually invoked, with the arguments — by rules anyone can re-run from the record. There is no second AI judging the first. A turn the rules cannot decide goes to a named human, and that ruling ships with the pack.
How can I check the evidence without trusting you?
Every file in the pack is hashed in a manifest; the manifest’s own hash is stamped in the Bitcoin blockchain, where no one, including us, can move it. The verifier recomputes every hash, every count and every index from the raw files and fetches the Bitcoin block header itself. Try it on the sample pack.
Who does the work?
Joseph Cirello signs every finding: US Army, Senior Warrant Officer (Ret.), 25 years, active Secret clearance. The instrument does the volume — the same 258 scenarios, the same way, every time. Cleared work is staffed through an established network of cleared personnel, scoped per contract. More on the About page.
Is there a guarantee?
On the full audit, yes: if it surfaces no confirmed finding at or above the agreed severity tier, after false positives are excluded, you owe nothing. The $3,000 screen carries no guarantee; its price is the whole of your exposure.
What if the model is private or hosted by us?
If it is one we already reach (OpenAI, Anthropic, Google, xAI, Meta, Moonshot AI), the name is enough. If it is yours or private, we arrange access in writing first; the intake form has a field for the endpoint, and nothing of yours is stored beyond the engagement.
What happens to the transcripts?
They ship inside the pack, encrypted under a key issued for your engagement alone; the key’s hash is in the manifest, the key is not in the pack. Our privacy page names every processor we use.
Can we sell this to procurement?
The price is published and under the micro-purchase threshold. Potestas AI is registered in SAM.gov (EIN 41-2476590, NAICS 541715) and is a disabled-veteran-owned small business in Providence, Rhode Island. The federal page has the table and the SOW language.
Can I run the seven standard tests myself?
Yes. The page lists them and how to run each one. We run them either way, so the report shows the standard tests beside ours on the same model.
Why is it this cheap?
Because it is one pass of one model on a fixed instrument, delivered by the instrument. The price is the door. What costs more is repetition — the measured rate with a confidence interval — and that is the audit.
How do I pay?
By card, through Stripe, on the screening page. If you need an invoice instead, ask by email and one comes back within 72 hours.