Five practitioners rule on one decision in 72 hours.

One pricing decision. Five or seven practitioners. A weighted answer in 72 hours, with the ranges, the disagreement and the ruling in one memo. The buyer's brief becomes three to five estimable quantities and one ranking of the options; the panel estimates the quantities and ranks the options; the engine aggregates; the operator rules.

Who sits on the panel. Practitioners with five or more years in B2B pricing, monetization, revenue operations, or finance with pricing authority, or founders who have set prices for a product past $1M in annual revenue. Every panelist is verified by identity and role history before their first invitation and signs the panel terms once a year. Before every verdict, each panelist answers a three-question conflict screen (category employer; listed company; advised, invested or paid in the last twelve months) and a yes removes them from that verdict with no penalty. No panelist sits on a decision about their own employer or a direct competitor of it. Panelists are identified to the buyer by role and years, never by name or employer, and buyers are never named to the panel.

What a panelist is paid. $400 per verdict, about two hours of work, within five business days, or the same amount to a cause they name. Then, in this order: co-authorship on PricingBench, the early read of the Pulse and the Index benchmark, and a reliability record that is theirs and follows them from verdict to verdict.

What a panelist never does. Brings anything from or about their employer. Grades a task they wrote or reviewed. Sees a buyer's name. Copies buyer material out of the form.

Every panelist answers calibration questions whose answers are known, drawn from the Agent Pricing Index and from the outcomes of earlier verdicts. Two things are measured on those questions: how often the truth lands inside a panelist's stated range, and how narrow the ranges are. Those measurements set each panelist's weight in two ways, and both are printed in the memo.

When the two methods disagree, the memo shows both and the ruling says which the operator backs and why. Dissent is a residual: a panelist whose center sits more than two of their own corrected ranges from the posterior median is named in the memo by panelist number, never softened. No verdict ships on a maximum R-hat above 1.05 or with fewer than five final responses.

Never. The buyer's name, numbers or options. Panelist names or employers. Individual panelist responses. The memo is the buyer's; a short, anonymized case study is published only with written consent, and consent alone publishes nothing.

Always. This method. The engine's diagnostics on every memo. The panel's reliability records, as aggregates. Where the calibration questions come from: the Index and the outcomes of earlier verdicts.

The engine. The aggregation runs on judgment_engine, a numpy package with no compiled dependencies, written to be read. It publishes under the MIT license with the first verdict, so anyone can run the method on their own panel. The code is the method. The engine aggregates; the operator rules.

When it stops. A verdict ships with fewer than five final responses twice in a row; a target's R-hat stays above 1.05 after one rerun; a panelist is found to have used employer-confidential information. The operator stops it, in writing. What survives it: every seed and every answer, every panelist's reliability record, every memo and ruling, and this page.

A verdict is not a guarantee, investment advice or a market study. It is a measured panel's distribution over your decision, with the disagreement shown.