StewAI Blog · Decision Science · 14 min read
It Got Four Experts to Agree, Then Refused to Call It Consensus
Ask four experts how long a project takes and you get four numbers. This is the 1963 Delphi method built as a workflow: no agents, the model reasons but never decides what runs next, and the verdict is a threshold on a spread computed in code. On a real run it narrowed a 15-month disagreement to 8, then declined to call that a consensus.
Frequently asked questions
What is the Delphi method?
A structured forecasting method developed by Dalkey and Helmer at RAND in 1963: collect expert estimates privately, share anonymized feedback carrying arguments and statistics but never who said what, let experts revise, and measure convergence numerically instead of negotiating it in a room.
Why did the recipe refuse to declare consensus?
Convergence is defined as a threshold on a spread computed in code, 0.20 in this recipe. After one revision round the spread had fallen from 0.92 to 0.48, still above the threshold, so the gate routed to a dissent register instead of a consensus statement. The branching is a condition on a number, not a model's opinion.
What happened on the real run?
Four experts estimating an ERP migration started at 9, 14, 18, and 24 months. After a controlled feedback round, three revised and one held, landing at 12 to 20 months. The recipe reported the narrowed range plus a structured dissent register naming what still divided the panel and what evidence would resolve it.
Read the full article on StewAI