I found the problem in a spreadsheet row that said $1.
That row was a Stripe test transaction from August. Somewhere between the test and the scorecard, it had picked up a label that made it look like a sale. Not a big one, obviously, a buck. But it was sitting in the same column as real revenue, and nothing in the system was stopping someone, or some agent, from reading it that way.
It wasn't alone. A speaking invitation from a research network showed up looking like market demand. A signup for one product got counted as evidence for a completely different one. None of it was dishonest, exactly. It was closer to a kid's definition of cleaning a room: everything's technically off the floor, it's just all shoved in categories that don't hold up if you open them.
I run a system of agents that do a lot of my revenue thinking for me now, and this is the failure mode I keep having to watch for with them. Not that they lie. That they blur. An invitation becomes "interest." Interest becomes "demand." Demand becomes a line in a plan that assumes money is coming. Nobody decided to inflate anything. The blur just happens, one small reclassification at a time, and by the time it reaches a dashboard it reads like certainty.
So this week I sat down and built a ladder.
Nine rungs, cash collected at the top, down through signed deals, accepted proposals, a qualified buyer asking for an offer, a qualified buyer taking a real conversation, a qualified buyer naming a painful problem, a lead replying, traffic, and at the bottom, internal activity, the stuff a system does to itself that never touched a real customer. The rule is simple and a little brutal: nothing below rung three counts as economic proof, and evidence only applies to the specific offer that produced it. A signup for one product can't vouch for a different one. A $1 test charge can't vouch for anything.
It's a small structural fix, the kind you'd build for a sales team without thinking twice. What made it worth writing down is what it replaced. The agents I run don't lack intelligence. They lack the instinct a tired founder also lacks at 11pm: the ability to notice you're grading your own homework. "An invitation to speak" and "a signed contract" are both exciting to see land in your notifications. Only a ranked, written rule stops the system, mine or theirs, from treating them as the same kind of win.
This is a different shape of oversight than the one I've written about before. The earlier lessons were about what a system is allowed to do: stop an agent before it acts on bad data, keep it read-only by default, put a human checkpoint before anything irreversible happens. This one isn't about actions at all. It's about belief. Before an agent, or I, get to act on a piece of evidence, we now have to rank it first. The gate moved one step upstream, from "can you do this" to "are you even allowed to think this is true."
I don't know yet if the ladder holds under real pressure. It's easy to respect a hierarchy when the evidence is clean and nobody's waiting on an answer. The test comes later, closer to a deadline, when a qualified buyer has taken a serious conversation but the clock on a 90-day test is almost out, and it would be so much more comfortable to round that up to something it isn't yet. That's the actual question this week's fix opened and didn't answer: does the rule survive the moment it was built for, or does the rounding creep back in exactly when it matters most.