forecasting//dated bet

A dated bet is a forecast written with a resolution date, a confidence level and the observations that will count as a hit or a refutation, all fixed before the outcome is known, and it is used to make a prediction accountable instead of adjustable. Its value comes from the conditions: a claim that cannot fail cannot be scored, and a claim whose test is chosen after the fact can always be rescued.

A dated bet is a forecast written with a resolution date, a confidence level and the observations that will count as a hit or a refutation, all fixed before the outcome is known, and it is used to make a prediction accountable instead of adjustable. Its value comes from the conditions: a claim that cannot fail cannot be scored, and a claim whose test is chosen after the fact can always be rescued.

A complete bet has five parts: the quantity, the threshold, the date, the confidence and the source that will settle it. For example: by 31 December 2030, the training compute of the largest AI runs will grow at less than three times a year, measured on the runs of 2029 and 2030; a measured trend of four times a year or more refutes it. The gap between three and four is deliberate, so that a result near the line is not counted as a hit.

Direction and timing are judged separately. Forecasters are usually surer that something will happen than of when, and a bet can say so by stating its confidence for each.

The settling source needs a fallback. If the measurement stops being published, the bet names a second test or is recorded as not assessable, never as quietly right.

A bet that changes after it is published records the change beside it, with the date and the reason, so that the record of what was claimed stays intact.

A bet is only as good as its calibration over many bets, which is why the confidence is written as a level that can be checked.