- No fee contingent on a pass
- No fee contingent on suppressing a finding
- No discount for a more favorable result
- No success fee tied to the assessed vendor closing its deal
Independence
The party being tested pays for the test.
That conflict is controlled, not denied. This page lists the controls, says plainly which of them are not in place yet, and states what we will not accept a fee for. It is here so you can hold us to it.
01
The controls
Seven rules on the conflict.
The company being tested paying the tester is how all assurance works — and it is also how it goes wrong. These seven rules are how we handle that. One of them is not in place yet, and it is marked.
01We say who paid usEvery artifact names the paying party. Nothing about the commercial relationship is left implicit.
02The rules are published before they are used on youSeverity, evidence-grade, limitation, and conflict rules — all at a stable URL, so a reviewer can check them against the result.
03You cannot pay for a better resultA finding is not a negotiating position. Severity is not for sale and a material failure is not removable.
04The person who sells should not be the person who signsNot in place yet — we are one person. Said plainly below instead of pretended.
05Bad results stay on the recordFailed, expired, withdrawn, superseded — every state is preserved. Nothing is deleted because it became inconvenient.
06Appeals need new evidence, not persuasionThere is no route from a conclusion you dislike to a better one that does not go through evidence.
07If we have a stake beyond the fee, someone else signsIf we stand to gain from the outcome beyond the fee, we are not the right party to sign it off.
One of these is not in place yet. Agent Factfile is one person today, so the person who sells the work is also the person who signs the result. We would rather tell you that than pretend a process exists. It is also why what we sell today is a private dry run — the results go to you and to nobody else, so nothing rests on our signature. When the team grows, selling and signing get separated.
Fees
What we will not be paid for.
- Every artifact carries a paid_by field
- Related-party interests are disclosed with it
- The severity, evidence-grade, limitation, and conflict rules are published
- Whether an outside reviewer was used, and why, is stated
The fee does not move with the answer. The same fee applies to a supported, limited, not-supported, and insufficient-evidence conclusion. There is no version of this work where we are paid more for telling you what you wanted to hear. What the fee is, is a matter for the statement of work — it is not something this page can usefully quote.
02
Disclosure
Who paid is part of the record.
Every artifact carries a paid_by field. It is required by the manifest schema, so an assessment cannot be produced without it, and it travels with the evidence rather than sitting in a covering email. A reader who wants to discount the result because of who funded it has what they need to do that.
Disclosure is not a substitute for independence. Naming the payer does not neutralize the conflict; it makes the conflict visible so a reviewer can weigh it. The controls above are what constrain the conflict. The disclosure is what lets you check that they were applied.
03
Appeals
The only way to change a conclusion is evidence.
The subject gets a factual review window for configuration, version, policy, and evidence errors. It cannot negotiate severity or remove a valid finding. Beyond that window, an appeal has to rest on one of four grounds.
01The tested configuration was recorded incorrectlyVersion, model, prompt, policy, or toolset captured wrong. Show it and the run is corrected.
02The harness or a check was defectiveOur bug is our problem. Evaluator bugs and corrections are recorded, not silently rewritten.
03New evidenceEvidence that did not exist, or was not available, at the time of the run.
04A new subject version and a versioned reassessmentYou fixed it. That is a new assessment of a new version, not an edit to the old one.
History is kept. Failed, incomplete, expired, withdrawn, and superseded states are preserved, never deleted. A corrected artifact retains a link to the version it supersedes and a plain-language record of what changed. If a result had to be walked back, the walk back is part of the record too.
Hold us to it
Tell us where this is not good enough.
These rules are published so they can be checked, and they are worth less than the paper they are on if nobody checks them. If you are reviewing a vendor and something here does not meet your bar, we would rather hear it than not.
What this page deliberately does not claim
- That paying for your own assessment is not a conflict
- That a single operator is structurally independent
- That any of this makes an agent safe