Skip to content
Read • Question • ConnectAnthony’s notes & researchReturn with a new question
Anthony’s NotesA personal field guide
From the notebook

Open Questions

Open Questions

Navigation

Hub: 00 - Fact-Check MOC · Previous: 05 - Findings Register · Next: 07 - Query Design Audit

The three structural problems that outweigh any individual error, and the empirical questions still open.

The three structural problems

These matter more than any individual error in 05 - Findings Register.

Duration mismatch. The thesis covers October 2023 to October 2024. Of roughly twenty figures, three span the full year and two state no window at all; the rest cover 30 to 180 days, with 100 days modal. Partially defensible, since the claim is specifically that the frame was set early, but not as written.

The unaddressed objection. Emotive terms may track violence form — close-range, deliberate, against the defenseless — rather than victim identity. October 7 fits that form; aerial bombardment and siege do not. The book contains no counterargument section, no engagement with this anywhere in 257 pages. Tested directly in 04 - Results Stage 1 with a mixed result.

Johnson has the material to answer this and does not use it: he reports that the NYT used “massacre” 124 times for October 7 and zero times across hundreds of discrete Israeli strikes. Event-form explains a large gap; it strains to explain zero.

Measurement versus assertion. Word-frequency counts can establish that coverage was asymmetric. They cannot establish that it “created the conditions” for anything, or that it followed from deliberate choices rather than newsroom routine, source access, or wire-copy inheritance. There is no measurement anywhere of audience belief, policymaker behavior, or public opinion. The causal link is argued, not evidenced — and no improvement in sampling fixes it. This is the widest gap between the book’s data and its title, and the “beyond a reasonable doubt” framing invites exactly the scrutiny it cannot survive.

Open empirical questions

  1. Does the Nation CSV shortfall have an innocent explanation? Almost certainly row-capped export versus aggregate endpoint. First question for Ali.
  2. Was the “genoicde” query actually run misspelled? Re-run both spellings against GDELT and compare. Resolvable in minutes.
  3. Does the editorial-versus-attributed distinction hold under coding? The single most important open question. Preliminary clip inspection suggests Nuseirat’s massacre usages are attributed to others while October 7’s include substantial editorial-voice usage. If that holds across a coded sample it substantially vindicates Johnson’s method. Requires manual coding of 100+ clips per event.
  4. Where does the book’s year-long eight-outlet print data come from? Not in either repository.

Next actions

  1. Manual coding pass on GDELT clips with VICTIM_CODE and IS_QUOTE columns, at least 100 per event. Blocks any reporting of stage-1 numbers.
  2. Paginate past the 50-clip cap so October 7 and Bucha yield exact counts rather than floors.
  3. On receipt of the full Intercept dataset, audit victim-attribution coding against sentence text.
  4. Extract remaining figure-linked numbers (Figs. 1.1, 1.2, 2.2a/b, 2.3, 2.4, 4.1, 7.1, 7.2, 8.1–8.6) into a claim table and run the same internal-consistency checks that produced findings 1 through 4.
  5. Citation integrity pass across all 476 endnotes: resolve each to a live URL, verify author, title, outlet, date.

A note on method

Two temptations to resist, given that this touches an actively contested subject.

The first is treating a discrepancy count as a verdict. Ten errors in a 476-note apparatus is unremarkable for a trade book and says little on its own; what says something is their direction and whether they touch load-bearing claims.

The second is letting the audit’s framing do the work. Prompted to find bias, an analysis finds bias everywhere; prompted to verify, it rubber-stamps. Where interpretation is required, run the question both ways and treat disagreement between the two passes as the signal.

A thesis can be under-evidenced and still true. Establishing that the method does not support the conclusion is not the same as establishing that the conclusion is false, and this audit is scoped to the former.