Assume the buyer has a model open
A verification standard written around one picture: a corporate buyer with an AI auditing every page.
The question on the table was what bar the engagement documents had to clear before reaching a marketplace where corporate buyers screen consultants. The answer: picture a frontier model reading every page for logic and arithmetic, and build to survive that reader, not to sound smooth.
Our existing review only checked whether a claim was permitted. Nothing asked whether the analysis itself held. Those are two different failures, and the second hurts more. A buyer's model finds a total that fails to add up in seconds, then marks down everything else shipped with it. So the cheap protection is to run the hostile check ourselves before anything goes out, and to publish what it found.
We added a reasoning pass beside the compliance pass. The arithmetic must recompute. A term has to mean one thing from first page to last. Every assumption carrying weight needs its basis written down. The conclusion has to follow from the work, and the first skeptical question gets answered inside the document. My own merchandise plan became the house standard. It reaches its headline number three independent ways, prints the variance, and labels each input documented, benchmarked or placeholder.
Putting your uncertainty in print flips an audit from a threat into a selling point. Most consultants deliver a model that looks sure of itself. A model that shows precisely where it is solid is the rarer thing, and worth more.
I set the standard by describing a reader and not by writing a rule. A buyer with a model open is a concrete person, and writing for a person is easier than writing to a policy.
Story details
| Type | Strong opinion |
|---|---|
| Year | 2026 |
| Firm telling | on the Common Ground wiki |