Write a decision as an indented tree — options, chances, payoffs, all as honest 90% ranges. Get the expected value, the recommended path, how often it actually wins, and exactly what would flip the decision.
Start typing — or load an example.
Every range you write is a 90% interval, sampled 10,000 times (log-normal for positive ranges, normal otherwise; probabilities are normalised so siblings sum to 1 in every simulation). The tree rolls back a policy: at each decision node the tool commits to the option with the highest mean expected value — you decide once, you don't get to re-decide with hindsight in each simulation. What you read at each node is the distribution of expected value given what you don't know.
The verdict includes how often the recommendation actually beats the alternative across simulations — an EV edge that wins only 55% of the time deserves different scrutiny than one that wins 95%. "What would flip this" is the sensitivity that matters in a meeting: the probability threshold at which the recommendation changes, found by bisection, and any payoff ranges wide enough to change the answer within themselves.
Sources: Ronald Howard's decision analysis tradition; Douglas Hubbard's How to Measure Anything on calibrated ranges; and the Fermi estimator, this tool's sibling, for the estimation habits underneath it.