Calibrated Dissent

Methods paper · Protocol · OSF

A verifiable seal-and-resolution protocol for auditing adversarial AI forecasts, with a public pilot on M&A deals (N = 10)

If an AI is designed to challenge decisions rather than confirm them — a «cognitive contradictor» — how do you know whether its dissent is accurate rather than merely noisier? The question is empirical, but the field has no honest way to answer it: judging a forecaster after outcomes are known invites hindsight. This paper does not claim the answer. It builds the instrument that would let one earn it honestly.

The central contribution is a verifiable, hindsight-proof protocol. Every forecast is sealed in a public, append-only ledger — the Observatory — before any outcome is observable, with a double anchor: a SHA-256 hash of the content and an OpenTimestamps anchor on the Bitcoin blockchain. Resolution follows a fully deterministic rule over canonical public sources, compared against pre-registered baselines; the entire model–prompt–configuration bundle is frozen by version. Anyone can independently verify that the published text matches the original hash and that the hash existed before a given date.

The protocol is instantiated on a fixed, publicly pre-sealed pilot of N = 10 forecasts on high-stakes European M&A deals, sealed on 20 June 2026. The scored quantities are the forecaster’s probabilities; the engine’s entire challenge record — the seven-field schema and the Δ-CSI — enters as provenance only and cannot alter them. The Δ-CSI is a computed proxy for the intensity of the challenge, not for the correctness of the decision: treating it as an accuracy signal is a category error.