Read the version, not just the headline

The identifier and version belong together

A paper can change after its first appearance. Preserve its stable identifier, the version read, the date checked, and the specific claim being used. If a title changes, confirm that the identifier still points to the same work rather than counting the revision as a new study.

A concrete revision in this edition

Towards Secure Agent Skills was revised on 14 October 2026. Its current abstract describes 19 scenarios and seven documented incidents. Older summaries describe a different scope. This archive records the revision so a reader can reconstruct which account supports the note.

Separate publication status from access

A preprint may later be accepted at a conference while the accessible document remains on arXiv. Record what the publisher or author record actually says. Likewise, an abstract-only catalog note is not a full-methods review. Both can be useful if their evidence level is explicit.

Build a claim ledger

For an important claim, record the source, version, supporting section, conditions, and what remains unverified. When a source changes, revisit the dependent explanation. Avoid silently replacing a number while leaving an older interpretation in place.

The source catalog distinguishes editorial advice, primary records, vendor research, and datasets. None of those labels alone establishes correctness.

Worked example: a revision changes the scope of a claim

An early summary says a taxonomy contains seventeen scenarios; a later manuscript describes nineteen. Updating the number without checking which scenarios changed may leave the explanation inconsistent. Counting both versions as separate supporting studies would also exaggerate the independence of the evidence.

Work through the evidence

Retain a stable identifier, exact version, access date, and the specific sections supporting important claims. Compare changed definitions and experimental settings as well as headline numbers. Record when only an abstract or project README was accessible. A repository can describe a newer manuscript than the linked preprint; preserve that distinction.

Decision criteria and a useful review record

Revise dependent explanations when the underlying meaning changes, and leave a short note in the updates ledger. Do not imply full-methods review when the catalog entry is based on an abstract. Separate a conference acceptance statement from independent verification of the experiments. A careful source trail is more useful than an impressive count of links.