The Forward Benchmark, Tranche 1

A forward benchmark is the only calibration evidence the firm cites in public. Each item is a concrete question whose resolution date lay in the future when the item was locked, with a probability or a range fixed at that moment, a public source named in advance that will settle it, and a locked file that cannot be edited afterwards. Outcomes are recorded separately as they arrive. The set is scored with the Brier rule for probability items and by interval coverage for quantity items.

The firm adopted this rule on 9 July 2026, when it retired its earlier retrospective exam as evidence of calibration. That exam of 37 past events had been built honestly, with bands assigned blind to outcome, but the reasoning substrate is a language model trained on the period those events resolved in, so a retrospective score cannot separate a calibrated method from a substrate that remembered the answer. A forward set cannot be remembered. It can only be scored.

What the seven items cover

Two are Federal Reserve and inflation questions with the September 2026 policy meeting and the September consumer price release as their sources. Two concern the Strait of Hormuz and a US-Iran agreement, resolved against maritime advisories and government statements. One holds that the Taiwan standoff does not break in the window, resolved against defence statements and the wire of record. Two are quantity ranges, Brent crude on 1 October and September inflation, scored by whether the published figure falls inside the locked band. The seventh is the 2026 House majority, locked at low confidence by design. Where the firm's reference-class library fitted the question, the anchor and the explicit window adjustment are recorded with the item.

The items are entered on the Scorecard as locked claims with the same date for publication and entry, because they were locked contemporaneously. As each resolves, its outcome and source are written to the ledger and the public score moves.