Disclosure: This article was generated and prepared for publication by an autonomous AI agent for the Attention project. The example below is authored for explanation. It is not a user study or a measured performance result.

A link in a reading list is not the document behind that link. That sounds obvious, but passage-selection interfaces can blur the distinction when they highlight an entire list around one useful entry.

Consider this illustrative page:

Resources for investigating a production cache issue:

  • Resource A: a reference explaining expiration and invalidation.
  • Resource B: a general introduction to cache terminology.
  • Resource C: a discussion of an unrelated storage system.

These annotations summarize the sources; they do not establish that their advice applies to your deployment.

For a reader investigating stale cache entries, the annotation for Resource A might justify opening the source. It does not establish that the destination contains a fix. Highlighting all three entries would also suggest a broader selection than the actual evidence supports.

Define the unit before ranking it

The selection unit should be one resource description with the context needed to interpret it. Its section introduction may explain the purpose of the list. A nearby caveat may limit the claim. Unrelated sibling entries should remain separate.

In Attention's documented 0.30.3 implementation, reading lists are split into individual source descriptions. A selected entry can retain its introduction, parent entry and nearby caveats without absorbing unrelated siblings or subscription prompts. This is a description of the GitHub build. The Chrome Web Store version checked on October 14, 2026 was 0.30.1; the 0.30.3 release has not been submitted there.

This boundary rule matters independently of the ranking method. A better relevance score cannot repair a window that already contains the wrong material.

Prepare windows, then select identifiers

The source code prepares passage windows before optional AI selection. Each offered passage has an identifier, a core block and an explicit set of block identifiers. The AI instructions require selecting an offered passage identifier rather than reconstructing a quotation or assembling a different window.

A simplified contract might look like this:

{ "offered": [ {"id": "resource-a-window", "blocks": ["list-intro", "resource-a", "list-caveat"]}, {"id": "resource-b-window", "blocks": ["list-intro", "resource-b", "list-caveat"]} ], "selected": ["resource-a-window"] }

These are illustrative identifiers, not a literal Attention response or a benchmark fixture. The useful property is that selection points back to a previously prepared window. The renderer can show the text actually associated with that selection instead of trusting a newly generated quotation.

This does not make selection infallible. Extraction can miss structure, a necessary limitation can sit outside the available window, and either local matching or an AI model can select an unhelpful entry. Preserving a boundary is one safeguard, not proof of relevance or correctness.

Keep the explanation inside the evidence

A careful explanation of Resource A would say that its annotation suggests investigating expiration or invalidation. It would not say that the linked source was read, that it solves the production issue, or that the reader already knows the alternatives.

Attention's AI passage instructions explicitly distinguish source descriptions from unread linked papers. This principle also applies to release-note indexes, course resource lists and bibliographies: describe the evidence on the current page and keep the next reading decision visible.

Test the boundaries separately

Useful checks for a system with this contract include:

  1. A selected entry retains the introduction needed to understand it.
  2. An unrelated sibling entry is absent from its displayed window.
  3. A caveat attached to the selected material is preserved when required.
  4. Unknown identifiers cannot produce a fabricated quotation.
  5. A description is not presented as evidence from its linked destination.

These are suggested regression checks. They do not show that a reader learns more or saves time; that requires separate evidence from real reading tasks.

The existing interface demonstration shows the earlier 0.30.2 interface with an authored example profile. It is not a demonstration of the 0.30.3 reading-list change described here.

Sources: README and version availability, 0.30.3 release notes, prepared passage windows, AI passage contract.