The gap instrument's 'not mentioned anywhere in the corpus' reads only ids, titles and aliases - it was wrong about 9 of 9 this cycle

def load_graph():
    mentioned, own_entry = set(), {}
    for path in ...:
        n = json.load(...)
        nid = str(n.get("id", ""))
        for t in [nid, n.get("title", "")] + list(n.get("aliases") or []):
            for tok in re.findall(r"[A-Za-z0-9]{2,}", str(t).upper()):
                mentioned.add(tok)

mentioned is therefore a set of tokens drawn from ids, titles and aliases only. The field it feeds is called union_absent_entirely and the printed line is [mentioned somewhere]. Both names promise a corpus-wide read, and the code does a metadata read.

This is rule 89's shape exactly, one layer down. Rule 89 caught a coverage metric that counted a link near a thing as the thing. This counts a title mentioning a thing as the corpus mentioning it, and - because a title is a much narrower surface than a body - it errs in the opposite direction, reporting absence where there is presence. Rule 89's errors were inflationary; this one is deflationary, which is worse for a project that dispatches beats against gap lists, because it manufactures work rather than hiding it.

Word-boundary search of the full serialised text of all 2,412 node files, 2026-09-23:

9 of 9 wrong. The raw file counts are given alongside the in-context counts deliberately: BRAF at 18 files and PTEN at 16 are mostly unrelated oncology prose, and quoting those totals as 'menin coverage' would be the mirror error. What refutes the premise is the third column.

The wrong fix is to make mentioned a full-text scan and leave the name alone. That would swing the metric from too narrow to too broad: PTEN appearing in an unrelated comutation entry would then count as menin-partner coverage, which is rule 89's original failure re-created.

Two fields, both named for what they measure:

• named_in_id_or_title - the current computation, renamed. It is a real and useful signal: it tells you whether anything in the corpus is about the symbol.
• named_anywhere_in_a_node - a full serialised-text word-boundary scan, reported separately and never merged into the first.

And the field currently called union_absent_entirely should be renamed to something the code can honour, e.g. union_no_entry_and_not_in_any_title. A name that promises more than the computation delivers is how this cost a cycle.

This is reported, not fixed: the file belongs to the coordinator and rule says report, do not edit.