Five Steps to Governed AI

Chapter Seven  /  Step Four

01
02
03
04
05

Prove


Before any output reaches your team or your clients, it is checked against your firm’s rules.

Pass or fail. With the rule named.

Why it matters

A confidence score is not proof.

Unverified, the firm carries liability for every output. When something goes wrong the cost is not the error itself. It is every hour spent establishing whether the error happened, plus every client who now checks your work before paying. A confidence score is a property of the model, not of the answer. What survives the question is the rule, the input, the output and the link between them.

What good looks like

A separate verification layer, not the same AI grading its own homework, working from your written-down policies and returning a definitive pass or fail. On a failure you see the rule that broke and the data that caused it.

In PitCrew Factory, this layer is powered by AWS Automated Reasoning, a formal verification system that checks the rules as written. It is not a second model checking the first.

Proof

R 3.4 was applied to this input and produced this output. Pass.

Not proof

The model reported 94% confidence.

The same output, two answers

VERIFIED
R 3.4 applied to this input.
Output consistent.
Pass.
UNVERIFIED
94% confidence.
Source not cited.
No rule named.

Figure 7.1

The rules check. A different system does the checking.

PRODUCED BY THE AGENT
June invoice, vendor 40-118843
$8,412
Rules checkYOUR POLICIES, WRITTEN DOWN
R 2.1Invoices under $5K auto-approve when the PO matches exactly.
R 3.4Invoices from $5K to $10K need department manager sign-off.
R 5.0Variance above 2% from the PO amount routes to the controller.
Checked, not estimated. Not the agent checking itself.
PASS
Released with proof attached
Rule, input, output and timestamp travel with the result.
FAIL
Held, with the rule named
R 5.0 broke on this input. Nothing reaches the vendor.

The gate is separate from the thing it is checking. That separation is what makes the result proof rather than an opinion.

What travels with a pass

THE RULE
R 3.4, as written and approved
THE INPUT
Invoice amount at close, timestamped
THE LINK
Why the two are consistent

In your firm

The agent reports that the wire matches standing instructions. The routing number is one digit off from the file.

Pattern matching found a near-fit and called it a match. A rules check compares the field exactly, and catches it before $200K moves to the wrong account.

What the AI saw
A routing number that looked right.
What the rule saw
A field that did not match the record, character for character.
What it saved
The $200,000 that would have left, plus the recall, the client call and the incident report.

The number was there. Nothing recorded which rule had produced it.

Skip this step
You cannot prove the output is correct. A confidence score does not survive that conversation.