TextQuoteSelector fuzzy fallback can silently re-anchor to a lookalike quote in a different location after the original text is edited

[antrixy]

Spec clause: W3C Web Annotation Data Model — Text Quote Selector. The selector is defined by exact, prefix, and suffix; prefix/suffix exist specifically to disambiguate among multiple occurrences of exact in the document.

Expected behavior: Hypothesis's own blog states the design intent directly: "Hypothesis already deals with minor changes to a document thanks to our fuzzy anchoring algorithm... But sometimes documents change a lot. If an annotation is anchored to a sentence in a paragraph and the whole paragraph is deleted, then even the smartest algorithm isn't going to help — the original text is no longer present and that annotation will fail to anchor. When annotations fail to anchor, we call them orphans." (web.hypothes.is/blog/showing-orphaned-annotations) This case should have produced an orphan. It didn't — it produced a confident, wrong match.

Minimal witness:

<!-- captured against -->
<div>
  <p id="alpha">Team Alpha reported that the new deployment reduced latency by twelve percent across all regions.</p>
  <p id="beta">Team Beta reported that the new deployment reduced latency by twelve percent across all regions.</p>
</div>
<!-- selector captured on Alpha's sentence: -->
{ "exact": "reduced latency by twelve percent across all regions",
  "prefix": "eported that the new deployment ",
  "suffix": ".Team Beta reported that the new" }

<!-- page later edited to: -->
<div>
  <p id="alpha">Team Alpha reported that the new deployment reduced latency by twenty percent across all regions after further optimization.</p>
  <p id="beta">Team Beta reported that the new deployment reduced latency by twelve percent across all regions.</p>
</div>

Expected vs. actual: Per Hypothesis's own stated design, this should orphan — the original text no longer exists at its captured location. Actual: toRange() returns a confident, non-null match — silently re-anchored inside <p id="beta">, a different paragraph with different authorship, based on a decoy occurrence. No signal is surfaced that the match moved.

Mechanism: the captured prefix ("eported that the new deployment ") is identical in both paragraphs — the disambiguating token (Alpha vs Beta) sits outside the fixed 32-character context window, so the fuzzy search can't distinguish them once the original location's exact text is gone.

Affected-implementation matrix:

Strategy	Result on this case
DOM-path (XPath+offset)	Fails at the correct location — wrong content, doesn't move elements
Text-position (global offset)	Same — wrong content, correctly contained
web-highlighter (tag+index+offset)	Same — wrong content, correctly contained
dom-anchor-text-quote (fuzzy text-quote)	Confidently anchors in a different element entirely

Related work, checked before filing:

• Bump eslint from 5.13.0 to 5.14.0 #954 in product-backlog — closed, resolved via Change anchored range representation to fix overlapping highlight issues #2720. Its example is a pre-existing duplicate at capture time plus a PDF cross-page-boundary case — different from this report, where the text is unique at capture and only becomes ambiguous after a later edit.
• Change anchored range representation to fix overlapping highlight issues #2720, "Change anchored range representation to fix overlapping highlight issues" — reviewed directly. Different mechanism, doesn't cover this: it fixes a cross-annotation side effect (a cached, resolved Range silently invalidated when painting a different annotation's highlight mutates the DOM), fixed by moving to just-in-time resolution. No changes to dom-anchor-text-quote or the fuzzy-matching algorithm. This report is about the matching algorithm's disambiguation logic when the page author edits the source text — a different layer and cause, no file overlap with Change anchored range representation to fix overlapping highlight issues #2720.

Oracle citation: TextQuoteSelector's own spec text (disambiguation is the stated purpose of prefix/suffix) plus Hypothesis's own orphans-design blog post (large changes should orphan, not silently reattach).

[robertknight]

I do not work at Hypothesis any more, but if I did I would close this as low-effort AI slop.

It is correct that quote anchoring can anchor to text that is similar to the original but doesn't semantically match. The problem with requiring the context to match exactly is that it would result in valid matches failing to anchor if the context has changed in a way that is not important. This happens often on real web pages where the extracted context may be text that comes before/after the selection in the HTML source, but isn't logically related to it.

Put another way, there is a trade-off between precision and recall when anchoring annotations. Any changes need to consider both.