To determine if a model predicts well, it is tested on a future that has already passed.

General Reasoning's BackSearch tool freezes the web at a specific date. A solution for evaluating the reliability of analytical models' forecasts.

Models are increasingly being asked to predict the future, and the simplest way to check if they can do so is to question them about a future that has already occurred. However, this requires freezing the world as it was at the time. General Reasoning is tackling this with BackSearch, two functions that search and read the web in the state of a frozen archive, through its OpenReward service.

The problem the tool solves is information leakage. Traditional search APIs, even when equipped with a date filter, rank today's index with the hindsight of the present and return the current version of a page, which is sometimes rewritten after the event has unfolded. BackSearch attaches a reference date to each query: the search only returns documents collected on or before that date, and the reading function restores the text of the article as it was archived at that time. The same query at the same date will always return the same result.

One technical point governs everything else. The date filters based on when the page was collected, not on the publication date it displays. A page archived after the deadline is discarded even if it claims to have been published before—a deliberate precaution: the self-declared date is precisely the field that a retrospective archive cannot trust.

The targeted use cases revolve around evaluation: scoring a model on questions for which the answer is already known by blocking the horizon of evidence before the outcome, replaying a financial research and decision loop over a past window, or training agents against a stable and replayable web. The open version remains a narrow preview, limited to news sites and a window spanning from December 2025 to July 2026, which is set to expand based on feedback. Billing is usage-based, at ten dollars per thousand searches and two dollars per thousand reads, with failed queries not being counted.