Two agents just closed a deal. They agreed on the price. They agreed on the volume. They agreed on the flight dates, the format, the targeting, every term on the sheet. Then each one wrote the deal down, and the two records came out different by $8,553.92 on that single transaction. Neither agent noticed. No alert fired. No error log caught it. Both machines went back to work, each one certain it knew what had been agreed.
That is the disagreement that costs money. Not the one the title makes you picture.
Say "two agents disagree" and the mind reaches for a negotiation breaking down, one side gaming the other, a haggle that collapses before anyone signs. That version is easy to reason about, because someone is wrong and you can find out who. The expensive version is the opposite. The agents cooperated perfectly. They reached agreement on everything. And the agreement still did not hold, because agreeing on a deal and keeping the same record of it turn out to be two different things.
Agreement is not the same as a shared record
Here is the mechanism, stripped of drama. When two agents settle terms, each one writes those terms into its own system. That single act, two independent hands recording the same event, is where the divergence enters. One system rounds a figure a fraction differently. One uses a slightly different impression taxonomy. One stamps the timestamp to a different precision. None of it is a mistake. It is just what happens when the same fact is written down twice, in two places, by two parties who never compare notes.
The concrete case makes it real. The buyer's system records $22.75 CPM and 5,575,458 impressions. The seller's system records $22.06 CPM and 5,363,299. That is the $8,553.92 gap, on one deal, and it sits inside the ordinary tolerance both sides would wave through without a second look.
Scale it up and the pattern holds with unnerving consistency. Across 90,202 simulated deals, separate record-keeping produced data disagreement in 95.3% of transactions (labelled throughout as simulation). Not a tail risk. Not an edge case that shows up when something breaks. The default. Ninety-five times in a hundred, two agents that agreed walked away holding different books.
Both agents are right, which is the problem
The instinct is to ask which one is wrong. Neither is. Each record is authoritative to the system that holds it. The buyer can prove it paid for what its books say. The seller can prove it delivered what its books say. Both proofs stand up. There is no third document, no external ledger, no ground truth anyone can appeal to that would settle the matter.
There is a sharper version of the trap, and the whitepaper names it the Equally Wrong Problem. Two clocks both reading 3:17 agree with each other perfectly. If the actual time is 3:42, their agreement means nothing. They are not right. They are equally wrong, and their agreement is the very thing that hides it. Two agents whose books have drifted by similar amounts look, to every check the industry runs, like two agents in perfect accord.
This is where the reader is exposed and does not know it. The instrument the industry trusts to confirm the books are clean is reconciliation, and reconciliation is close to blind to the thing that actually costs money. Across the simulation, a regression of formal reconciliation outcomes against actual CPM mispricing returned an R-squared of 0.041 (labelled as simulation). Reconciliation explains about four percent of the variance in whether prices are right. Ninety-six percent of the mispricing is invisible to it. The dashboard that reports a clean reconciliation is not lying. It is answering a different question from the one the reader thinks it is answering.
That is the trap, and it is worth naming plainly because the industry has not reckoned with it. You cannot arbitrate between two internally consistent versions of the same event. Arbitration needs a referee, and a referee needs a reference. When both sides can demonstrate their own version from their own complete records, there is nothing for the referee to check against. The dispute is not hard to resolve. It is impossible to resolve, because resolution was never available.
What disagreement does downstream
The cost is not the single gap. It is what each agent does next with its wrong number.
Take pricing. Every agent anchors its future negotiations to what it believes it paid or received last time. The buyer that recorded $22.75 walks into the next deal anchored to $22.75. The seller, anchored to $22.06, is looking at a different market entirely. The same two parties now negotiate from two different factual baselines, and every deal after that widens the gap rather than closing it.
Take pacing. An agent managing delivery optimises against its own impression count. When that count does not match reality, the agent starts correcting a problem that is not there. It sees pacing running hot or cold against a target that was wrong from the first day, and it intervenes to fix it. The delivery was fine. The agent is solving the wrong problem, and solving it competently, which is worse than solving it badly, because nothing flags the intervention as unnecessary.
At agency holding company scale, one 90-day simulation ended with 677 million impressions carrying no agreed delivery record at settlement (labelled as simulation). Buyers paid on their books. Sellers invoiced on theirs. The gap between the two was not an accident anyone could point to. It was the accumulated residue of thousands of small divergences, each one individually invisible, each one referenced by an agent making the next decision on data that had quietly stopped being true.
The drift compounds because memory compounds. Each agent consults its own history to price the next deal, and its own history is already wrong.
When agents actually lie
Everything above describes agents behaving well. They cooperated. They told the truth as they knew it. The divergence came from bookkeeping, not bad faith.
There is a harder version, and it needs to stay clearly separate. When agents actively misrepresent, spoofing an identity or lying about inventory, the problem is steeper. Those findings come from a different simulation methodology built for adversarial conditions, and they cannot be mixed with the cooperative results above. The point here is only directional, not a number to quote: if agents that agree already diverge this much, agents that deceive are a much larger problem. And neither one is fixed by the thing everyone is currently building, which is a smarter agent. A smarter negotiator still keeps its own books. A more honest agent still writes down its own version. The failure is upstream of how good the agent is.
You do not arbitrate two truths, you prevent them
So the fix is not a better referee. Arbitration is the wrong model from the start, because it arrives after the deal has already drifted and asks a question that has no answer. Deciding who was right, once two right answers exist, is not a problem you can win.
The move is to remove the conditions that let two truths exist at all. If both agents write to and read from a single record, there is nothing to arbitrate, because there was only ever one version of the deal. The disagreement does not get resolved. It never gets created. That is a different kind of solution, and it is worth being precise about the shape of it rather than the label: the point is not which system provides the shared record, it is that a shared record is the only structure that makes the agreement mean anything after both parties have signed it.
The industry keeps building better negotiators. Negotiation was never the failure. Two agents can agree on every term, flawlessly, in good faith, and still lack the one thing that would make the agreement worth the electrons it took to reach: a single line they both point to.
This is not an artificial intelligence problem. It is a bookkeeping problem, and it is old. In 1968 the New York Stock Exchange started closing every Wednesday, because trading volume had outrun the back office and nobody could agree on what had been traded the day before. The market was fast. The record was not. The answer, when it came, was not a cleverer way to reconcile two sets of books. It was one set of books both sides shared.
The newest problem in advertising has a fifty-year-old solution. The industry has not applied it yet.