At one point in the Law of the Sea negotiations, India, “representing the Third World bloc,” put a number on the table. Any company that wanted to mine the deep seabed should pay an initial fee of $60 million per site. The United States answered with a number of its own: no initial fee at all.
On the floor of the Pacific lay nodules of nickel, copper and cobalt that the UN had declared “the common heritage of mankind.” No one had yet mined them commercially, so what a mine would earn was an estimate at best. “Both sides dug in,” Roger Fisher and William Ury wrote in Getting to Yes (1981); “the matter became a contest of will.”
A contest of will ends when someone blinks. Whoever moves first looks weak, so nobody moves.
Then, in their telling, someone discovered that the Massachusetts Institute of Technology had developed a model for the economics of deep-seabed mining.
When a haggle stalls, agree on a fair yardstick before you argue about the number.
A ruler neither side owned
The model came from a team led by J. D. Nyhart, of MIT’s Sloan School and its Department of Ocean Engineering, which had begun work on the costs of seabed mining in 1976. It belonged to no delegation. Fisher and Ury say it was “gradually accepted by the parties as objective,” and it could answer one question for anyone who asked: what would this fee do to the economics of a mine?
India’s representative asked. He was shown that a fee “payable five years before the mine would generate any revenue” would make it “virtually impossible for a company to mine.” He said he would reconsider. The Americans, whose information had come mostly from the mining companies, saw that some initial fee was economically feasible, and they changed position too. “No one backed down,” Fisher and Ury wrote; “no one appeared weak.”
That line is the mechanism. In positional haggling, changing your number is a concession, and concessions invite more pressure. A standard turns the same move into something else. You are not giving in to them; you are following the figures. Fisher and Ury’s rule was “Never yield to pressure, only to principle.”
Flatmates splitting rent can argue all evening about who deserves the big room. Or they can measure the rooms and divide by floor area. The tape measure doesn’t care who wins, and losing to it costs nobody face.
Not every yardstick will do. Fisher and Ury wanted objective criteria to be “independent of will,” and also “legitimate and practical.” Figures from one side’s mining companies would have been more data to argue with. A university model, shown at a seminar that delegates “of all political persuasions” packed, as James Sebenius recalls, and that the chair of the financial talks, Singapore’s Tommy Koh, visibly supported, was harder to wave away.
When the ruler wobbles
A model is only as sure as its inputs. Sebenius, who served on Koh’s informal staff and later wrote a book on these talks, found that delegates came to regard the uncertainty around seabed economics as inherent. The model “offered less assurance about seabed economic outcomes than had seemed the case” soon after it arrived.
A shared standard also binds only the people who keep sharing it. The Convention adopted in 1982 set a $500,000 application fee and a $1 million annual fee. Harvard University Press describes Sebenius’s book as covering, too, “the abrupt U.S. rejection” of those negotiated results. A 1994 agreement set the $1 million fee aside. The UN’s own history says the economic forecast behind the seabed regime “has not been realized.”
The criterion got both sides past one deadlock. It could not make the deal last, or the forecast come true.
Twelve thousand or six
Say a client asks what you would charge to redesign their website. You say $12,000; they say their limit is $6,000. Splitting the difference at $9,000 can feel fair, but it can also reward whoever opened further from a sensible price.
The other route is to ask what a fair price would rest on. Your hours on the last three similar jobs. A published rate survey for your city. What they paid the previous agency for comparable work. If they bring a standard you can’t answer, you may have to move; if you bring one they can’t, they may. Either way, whoever moves can say why. It costs an extra round of emails, and it can still fail if the only number the client trusts is the budget.
The test before the argument
Teams building with AI models meet the same deadlock. One engineer says the new prompt is better; another says it broke something.
Anthropic’s guide to evaluating AI agents names the problem early: “Two engineers reading the same initial spec could come away with different interpretations on how the AI should handle edge cases. An eval suite resolves this ambiguity.” Write down the tasks and what counts as a pass before anyone argues about a change, and the argument becomes a run. With evals in place, the post says, teams can test a new model and “upgrade in days”; teams without them face weeks of testing.
The seabed lesson about wobbling rulers carries over too. Anthropic reports that Claude Opus 4.5 first scored 42% on a benchmark called CORE-Bench. Then a researcher found rigid grading that penalized “96.12” when it expected “96.124991…,” along with ambiguous tasks and results that could not be reproduced exactly. After the bugs were fixed and the model was given a less constrained setup, it scored 95%. As a rule, the post says, “we do not take eval scores at face value until someone digs into the details of the eval and reads some transcripts.”
An agreed test can settle which version is better. Whether the test measures what you care about is a separate argument, and the time to have it is before the run.