OpenRewardBackSearch
Search the web as it was
BackSearch is point-in-time web search and page fetch over a frozen archive. Every request carries an as_of date: search returns only documents crawled on or before it, and fetch returns an article's text as it was archived at that time. The corpus never moves, so the same query with the same as_of returns the same results forever.
BackSearch is currently a narrow preview: news domains only, covering December 2025 to July 2026. We plan to widen the domain coverage and expand the backdated period following initial feedback.
What you can use it for
Forecasting
Score a model on questions whose answers you already know, with the evidence horizon pinned before resolution.
Quant finance
Backtest a research-and-trade loop. Step an agent through a past window where it reads the news as it breaks.
RL environments
A stable, replayable web to train agentic environments and evaluations against.
Why not a date filter?
A live search API with a date filter ranks today's index with hindsight, and hands you today's bytes for a page that may have been rewritten since the event resolved. Either one leaks the answer into a question you meant to ask about the past.
BackSearch gates on crawl_date — when we archived the page — not on the article's self-reported publish_date. A page first captured after your cutoff will not be returned even if it claims an earlier publish date. That is what makes a run reproducible and leakage-free.
Quickstart
The base URL is https://search.openreward.ai. Authenticate with your existing OpenReward or_... key in the x-api-key header — there is no separate credential to obtain.
Each hit comes back with the archived URL, title, snippet, host, and both dates:
/fetch is the point-in-time counterpart. Give it a URL and a cutoff and you get the extracted article text from the latest capture on or before that date, not today's version of the page.
Inside an environment
The same engine ships as a toolset in the OpenReward SDK, so an environment can hand its agent backdated web access without any of the plumbing above:
The agent gets web_search and web_fetch, both bounded to the cutoff. Expose web_as_of as a property instead of a constant and the window advances with your simulation clock — the world opens up one day at a time, as it did.
Pricing
Pay-as-you-go against your prepaid OpenReward balance, with no subscription. Only successful requests are billed — a search that errors, or a fetch with no capture on or before your cutoff, costs nothing. See full pricing.
Coverage
A practical note on the preview window: an as_of outside December 2025 to July 2026 returns no results rather than an error, so a run that silently finds no evidence is usually a date past the edge of the archive. Tell us which sources and periods your work needs at hello@gr.inc and we'll expand coverage based on your use.