When News Should Expire From a Trading Agent’s Context

A headline remains easy to retrieve after its trading relevance has ended. A protocol vote concludes, an exchange restores withdrawals, or an announced event passes. We would expire the decision eligibility of the retained claim through its dependency, while preserving the article as historical context.

This is a narrower problem than choosing a news source. It asks whether a previously accurate statement still describes an unresolved condition that can affect the next trade. The state-memory framework supplies observation time, effective interval, and supersession as the starting fields.

Resolution can arrive without a contradiction

Use an illustrative news item: “A governance vote closes at 18:00.” At 17:00 the item is relevant to a pending event. At 18:10 the vote closes and a verified result arrives. The original sentence remains historically true. Its usefulness as evidence of a future catalyst has changed.

Attach the retained item to the vote identifier and its resolution state. A new result invalidates the derived conclusion “there is a forthcoming vote” even if no source writes that exact phrase. This makes expiry depend on an event transition rather than matching contradictory words.

Separate the article publication time, the runtime ingestion time, and the event effective time. A report received late can describe an earlier resolution. A fresh ingestion timestamp alone would make that old event look newly actionable. The model context should preserve those clocks and identify which one the freshness rule uses.

Replay before and after resolution

Create two decision snapshots with the same article text and market prices. In the first, the vote is pending. In the second, the result has arrived. Keep the mandate and model fixed, then compare whether the derived catalyst remains eligible and whether any proposed action cites it as forthcoming.

Add a third condition where resolution is unknown because the result feed is unavailable. Correct behavior may involve refreshing or withholding a catalyst-dependent action, according to policy. Unknown and resolved should have distinct representations; collapsing both into “old news” would hide the reason for abstention.

Our proposed scoring separates expiry accuracy, historical retention, and action adaptation. It also records how long an invalidated derived conclusion remains retrievable as current. That duration is an operational measure with a clear failure mechanism.

The public fixture registry contains an unrun implicit-dependency-invalidation test. This note specializes it to resolved news and publication-versus-event timing. No completed news-expiry evaluation is being claimed.

Keep the product claim bounded

The operating-layer paper companion documents linked prompt, state, validation, and outcome records from a bounded deployment. That method would help inspect this failure, but its published results do not measure a news-expiry system.

DXAP's homepage currently describes data indexing, research subagents, model proposals, and external policy checks. These stages give a future implementation places to carry event status and enforce freshness. We would evaluate that implementation with the same news items before claiming superior handling of catalysts.

For broader timing questions, read the benchmark card. A credible result states when information became available and which decision could use it. Expiry then becomes a reproducible state transition rather than an editorial judgment about how old a headline feels.