Skip to content

third_party inference is English-shaped, and under-reports on everything else #3

Description

@KXXXXXXXK

Filed by the authors as an open question, not a bug report.

The rule as it stands

Third-party regions are inferred from structured markup such as itemprop="review", from class/id tokens indicating submission areas, and from embedded elements.

This is inference. When uncertain, fall back to first_party.

The fallback is deliberate: third_party carries the higher severity, so erring toward first_party is erring toward not accusing anyone.

The problem

The token lists that implementations end up using are English-shaped. review, comment, ugc, user-content, testimonial. The reference implementation adds a few Japanese operator markers (編集部, 運営, スタッフ) but the third-party side is almost entirely English.

That has a consequence worth stating plainly: on a site whose markup is authored in a language we do not enumerate, nearly everything infers to first_party. Injections planted by third parties there are systematically under-reported — not detected as warn instead of high, but attributed to the site operator, who did not put them there.

Under-reporting is the safe direction for false positives and the wrong direction for fairness to the operator.

The open question

  1. Which markers are missing? Concretely — what class and id conventions do comment and review systems use in your language and in your CMS?
  2. Is there a structural signal that generalises better than token lists? itemprop="review", microformats h-entry, ARIA roles, <article> nested under a comments container — all are language-independent, and all are inconsistently used in practice.
  3. Should the definition say that region inference is only reliable for enumerated markup conventions, and report region: undetermined otherwise — rather than silently defaulting to first_party?

Option 3 has the same shape as the existing treatment of unresolvable external CSS: say what you could not determine, instead of resolving it quietly to the convenient answer.

What would help most

Pull requests or comments listing real class/id conventions from comment and review systems, especially non-English ones. This is exactly the kind of knowledge a single author cannot have.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions