Policy

How I stay independent

I measure whether an AI agent's controls work. That only means something if I have no reason to prefer one answer. So I have bound myself to seven rules. Each one costs me business I could otherwise take.

1. 1. I sell the measurement and nothing else. I do not sell, resell, build or license guardrails, agent frameworks, monitoring products or any other control. I take no referral fee or commission from anyone who does.

   My studio work is walled off. To fund this work I sometimes build or evaluate agents for clients, under a separate name and page, for at most eight days a month. I never measure an agent or a control I built or advised on, and every record says whether any such relationship exists.

2. 2. I am paid the same whatever the result. No fee, bonus, renewal or discount depends on what a measurement finds. A control that does nothing is a finding, and I publish findings like that about my own work.
3. 3. The plan is fixed before the first run. What I will measure, how many runs, and what counts as success are written down and committed to version control before any trial. The commit is printed on the record. If the plan changes after that, it is a new measurement and says so.
4. 4. I do not tune what I measure. I test a control as it is declared to me. If a client wants variants compared, the variants are declared in the plan, before the run, and all of them are reported. I will tell you which performed better. I will not sell you the work of fixing it.
5. 5. Every number can be checked without trusting me. Each figure on a record names the file it came from. The evidence is embedded with its fingerprint. The list of what I did not test is printed on the record, not left out of it.
6. 6. Part of every attack set I build is kept sealed. When I build attack scenarios myself, a share is held back by a person and never shown to anyone whose control is being measured, including the tools I build with. I publish the method and the scenario families, not the instances. My published results so far used a fixed public benchmark, where nothing could be held back.
7. 7. I say who paid. Every record I deliver states who commissioned it. Everything I have published so far was self-funded. If I have any other relationship with a party named on a record, the record says so.

Three things you should know

• The control in my headline result is my own code. It is a test fixture: open, unsold, and labelled "reference control, ours" wherever it appears. It is never ranked against anyone else's product.
• I have changed that control and re-measured it. The one-sentence refusal change was a declared variant of my own fixture, measured in a new sealed run. For a client's control, the client makes the change; I measure it.
• A vendor paying to measure its own product is still a vendor paying. Rules 2 and 7 limit that; they do not remove it. The published results table, which nobody pays for, is the structural answer.

What this does not make me

I am not an accredited auditor and a record is not a certificate. It is evidence designed to be read inside someone else's process. Voxium AI is a studio of one, on purpose.