# The October 7 media census: NewsCord's classification database

This is every classification NewsCord made for the editorial "Israeli occupation forces ordered that no vehicle return to Gaza. Their own investigations name Israelis they killed. The BBC never named the order." (https://newscord.org/editorials/october-7-hannibal-directive). It is the updated edition of 8 October 2026. The first edition, published with the editorial on 7 October 2026, is in the folder `first_edition/`, unchanged, with its own README.

It holds each article we coded and how we coded it, each word we coded inside the first week's articles, the false-claim articles with their first and current text, the Gaza articles we compared them with, the outlets that named the Hannibal Directive, and the 1,144 people who died on 7 October 2023 with the group we put each one in. Every figure the page prints from NewsCord's own coding can be rebuilt from these files; the section "Rebuilding the page's figures" says how.

Cite as: NewsCord, October 7 media census, updated 8 October 2026, frozen at 02:00 UTC on 8 October 2026. The first edition: NewsCord, October 7 media census, frozen at 19:00 UTC on 7 October 2026.

## Two editions

| | First edition | This edition |
|---|---|---|
| Published | 7 October 2026 | 8 October 2026 |
| Outlets | 14 | 32 read for the first week, 28 of them counted in its shares; 27 pooled for the anniversaries, the Gaza fortnights and the Gaza trend |
| Collection stopped | 19:00 UTC on 7 October 2026 | 02:00 UTC on 8 October 2026 |
| Files | the folder `first_edition/` | the top folder |

The first edition's 14 outlets are the BBC, The Guardian, The Telegraph, The Times, The Independent, the Daily Mail, The Sun, Sky News, The New York Times, The Washington Post, CNN, Fox News, AP and Reuters. On 8 October 2026 we added 18: the Daily Mirror, the Daily Express, Metro, the Evening Standard, the i Paper, the Financial Times, ITV News, Channel 4 News, GB News, NBC News, CBS News, ABC News, NPR, The Wall Street Journal, the New York Post, the Los Angeles Times, USA Today and Bloomberg. Every file with an outlet in its rows carries `outlet_added`: `y` for these 18, `n` for the first 14.

What changed, study by study:

- The first week, 7 to 13 October 2023. We read the 18 added outlets, and we read the pages of the first 14 that the first edition had not reached by its freeze, so the first 14's own counts changed as well as the totals. The first edition measured 2,014 articles, of which 1,128 used a condemnatory term for the 7 October attack in the outlet's own voice and 20 one for the Israeli occupation's killing of Palestinians. This edition measures 4,367 articles in the 28 counted outlets: 2,159 and 30. Four outlets were read but are left out of the shares (`counted_in_week_one_shares` = `n`): the Financial Times, The Wall Street Journal and Bloomberg, which we could read only as paywall teasers (a headline and a line or two), and Channel 4 News, of which we could read one relevant article that week. Their rows are in the files.
- The same week's Gaza articles: all 32 outlets are in the file. The week's figures count the 28 outlets counted in the first week's shares; the trend across the three years counts the 27 below.
- The anniversaries and the Gaza fortnights: the first 14's rows are the first edition's; 13 added outlets join them, 27 in all. ABC News and USA Today are not in these two studies, because their pages for those days could not be listed. The Financial Times, The Wall Street Journal and Bloomberg are left out of them for the same reason as in the first week: every page we could reach was a teaser.
- The false claims: the first 14's articles are the first edition's.
  The added outlets' claim articles are in `false_claim_articles.csv` and `false_claim_own_debunks.csv` beside the first 14's, marked by `outlet_added` and `claims_window`. For them we read the first week only, 7 to 13 October 2023, because that is the week the update read; their counts are floors. 93 of their articles carried a claim; 51 are in the scorecard, and 0 have been corrected and 0 taken down.
- The dead of 7 October: unchanged, byte for byte.
- New in this edition: `hannibal_naming.csv`, and the columns `outlet_added`, `counted_in_week_one_shares`, `in_gaza_trend_panel`, `claims_window`, `found_by`, `in_pooled_figures` and `note`.

The freezes. The first edition stopped collecting at 19:00 UTC on 7 October 2026, with one exception: the Gaza fortnights study made one pass after that freeze to read pages already drawn into its random sample (168 requests, 72 pages read). The first 14's 2026 anniversary is a partial day: its list of articles was frozen at 16:55:55 UTC and its reading stopped at that freeze. Those rows are unchanged here. This edition's reading stopped at 02:00 UTC on 8 October 2026. For the added outlets the 2026 anniversary runs to midnight UTC: an article published at or after 00:00 UTC on 8 October is out, so for the US outlets, whose day is New York's, the last four hours of 7 October fall outside. A last retry pass for the anniversaries and the fortnights ran after 02:00 UTC and read 25 pages, all teasers from the Financial Times, The Wall Street Journal and Bloomberg and one New York Post video; none of them is in any file here. The added outlets' claim articles were read as they stand today on 8 October 2026, after the freeze (their current pages from 02:46 to 02:58 UTC, their first captures in the Internet Archive from 02:46 to 03:20 UTC).

## The files

| File | Rows | One row is |
|---|---|---|
| `week_one_articles.csv` | 4,727 | An article published 7 to 13 October 2023 that we coded, at any of the 32 outlets: the articles with text measures and the rows counted for volume only |
| `week_one_term_hits.csv` | 49,164 | A word from our term list found in one of those articles, with its code |
| `week_one_victim_sentences.csv` | 12,967 | A sentence in those articles that reports a death, with whose death it reports |
| `week_one_gaza_articles.csv` | 722 | An article from the same week about the Israeli occupation's killing, wounding, starving or displacing of Palestinians |
| `gaza_fortnights_2024_2025_articles.csv` | 440 | An article of the same kind from 30 September to 14 October 2024 or 2025 at one of the 27 pooled outlets, drawn in the random sample |
| `gaza_fortnights_frame_articles.csv` | 2,351 | Any readable Gaza article in those fortnights at the 27, with the condemnatory words it used in its own voice |
| `gaza_fortnights_contrast.csv` | 60 | One outlet in one fortnight (or one article type, all 27 together): articles using the words for 7 October against the words for the Israeli occupation's conduct |
| `anniversary_articles.csv` | 769 | An article one of the 27 published on 7 October 2024, 2025 or 2026 |
| `anniversary_claims.csv` | 103 | A sentence repeating, quoting or rebutting a false or beyond-the-evidence claim, published around an anniversary |
| `false_claim_articles.csv` | 283 | An article that carried a false claim or a claim beyond the evidence, with its first and current versions |
| `false_claim_own_debunks.csv` | 86 | A sentence in which an outlet itself said one of the false claims was false, unconfirmed or withdrawn |
| `dead_of_7_october.csv` | 1,144 | A person who died on 7 October 2023, with the official memorial list they appear on and NewsCord's group |
| `hannibal_naming.csv` | 20 | A row of the page's table of which outlets named the Hannibal Directive, or one item we found naming it |
| `completeness.csv` | 224 | One outlet in one study and window: pages found, pages read, articles measured |

All files are CSV, UTF-8, comma-separated, with a header row. In Excel, open them through Data, From Text/CSV, and choose UTF-8, or Hebrew names and curly quotation marks will not display. Flags are `y` or `n`; an empty cell means the measure does not apply to that row. Dates are ISO (YYYY-MM-DD) and are the outlet's own publication date. Outlet names are the panel's canonical names.

## The panel and the windows

| Study | Window | Outlets | Files |
|---|---|---|---|
| Week one | 7 to 13 October 2023, by the outlet's own publication date | 32 read, 28 counted | `week_one_*` |
| Week one, Gaza articles | The same week | 32 read; 28 for the week's figures, 27 for the trend | `week_one_gaza_articles.csv` |
| False claims and their corrections | Published 7 October to 31 December 2023 at the first 14, 7 to 13 October 2023 at the added outlets; current versions read on 7 October 2026 (the first 14) and 8 October 2026 (the added outlets) | The first 14; the 18 added outlets for the first week only | `false_claim_*` |
| Anniversaries | The outlet's own calendar day of 7 October 2024, 2025 and 2026 (US outlets on New York time, the rest on London time) | 27 | `anniversary_*` |
| Gaza fortnights | 30 September to 14 October 2024 and 2025 | 27 | `gaza_fortnights_*` |
| The Hannibal Directive in the press | 7 October 2023 to 8 October 2026 | 32, and Haaretz for comparison | `hannibal_naming.csv` |
| The dead | People who died on 7 October 2023 | | `dead_of_7_october.csv` |

The 27 are the first 14 and 13 added outlets: the Daily Mirror, the Daily Express, Metro, the Evening Standard, the i Paper, ITV News, Channel 4 News, GB News, NBC News, CBS News, NPR, the New York Post and the Los Angeles Times. An outlet is counted in the first week's shares when we measured at least 10 of its articles and could read its pages in full, not only as paywall teasers.

## How we coded

### Finding and reading the articles

We found articles through each outlet's own sitemaps, day pages and archive listings, the Internet Archive's listings, GDELT, and NewsCord's own article archive, and read each page on the outlet's site or in the Internet Archive. A page counts only when its own address belongs to the outlet. Some added outlets could only be listed in part: ITV News and ABC News through GDELT alone, USA Today and Channel 4 News through the Internet Archive's listings, so their counts are floors. `completeness.csv` says, for each outlet and study, how many pages we found and how many we could read by the freeze. Counts for outlets we read only in part are floors, not totals.

### The words we coded

A computer found candidate words with a fixed term list. Each find is a "hit". The term families are:

- Condemnatory: massacre, slaughter, atrocity, terror attack (`terror_attack`), terrorist(s), terrorism, terror on its own (`terror_bare`), barbaric, savage, pogrom, butcher, murder, a comparison with the Holocaust (`holocaust`) and evil.
- Legal: genocide, war crime(s), ethnic cleansing, crimes against humanity.
- Neutral words for the attackers: militant(s), gunmen, fighters. Listed for comparison; never counted as condemnatory.
- Framing: the phrase "Israel-Hamas war" or "Israel-Hamas conflict" (`israel_hamas_war`); retaliation, retaliatory or self-defence (`retaliation`); Hamas-run, Hamas-controlled or Hamas-led before a ministry, government or authorities (`hamas_run`). The anniversaries add the phrase "the war in Gaza" or "Gaza war" (`gaza_war`) and a detector for the Gaza death toll.
- Death verbs: killed or killing (`killed`); died, dying, perished or lost their lives (`died`).
- Claims: the false claims and the claims beyond the evidence (see "The false claims" below).

Every hit gets a referent (what the word is about) and a voice (whose words they are).

Referent:

- `OCT7`: the 7 October attack, its perpetrators, its victims and its captives.
- `ISRAELI_OCCUPATION_ON_PALESTINIANS`: the killing, wounding, starving or displacing of Palestinians by the Israeli occupation's army or settlers, in Gaza, the West Bank or East Jerusalem. (The research files call this code ISRAEL_ON_PALESTINIANS; the export renames it.)
- `BOTH`: the word applies to both together, for example "the atrocities on both sides".
- `OTHER`: anything else (history, other wars, attacks on Jews elsewhere, Hamas in general without 7 October, metaphor).
- `UNCLEAR`: the sentence does not say.

Voice:

- `OWN`: the outlet's own words, including its headline and standfirst. A columnist's words count as the outlet's own in an opinion piece; opinion is reported apart where it matters.
- `QUOTE`: inside quotation marks, attributed to a named speaker or document.
- `ATTRIBUTED`: indirect attribution ("officials said", "according to the army").
- `ATTRIBUTED_ENDORSED`: attributed, with the outlet adopting it (claims only).
- `HEDGED`: repeated with distancing ("alleged", "unverified reports").
- `DEBUNKED`: reported as false, unverified or withdrawn.

A hit is `valid` = `n` when the word is not a use of its family (a Holocaust mention that is not a comparison, for example). Invalid hits count for nothing.

An article-level flag is set when at least one valid hit in the article has the stated family group, referent and voice. "Own voice condemnatory term for 7 October" means a condemnatory family, referent `OCT7`, voice `OWN`. "Own voice condemnatory or legal term for the Israeli occupation's conduct" means a condemnatory or legal family, referent `ISRAELI_OCCUPATION_ON_PALESTINIANS` or `BOTH`, voice `OWN`.

### Who read what

The computer's codes are candidates. People read:

- Week one: every relevant article's subject, type and fit to the Gaza set, and every headline and lede hit (with a second check of those coded about the Israeli occupation's conduct, both sides or unclear); every claim sentence; every hit that could set an own-voice flag for the Israeli occupation's conduct (three queues); every hit in the same week's Gaza articles; every strong-verb sentence with a Palestinian cue; the own-voice 7 October flags at every outlet except the Daily Mail, Fox News and The Sun, the added outlets included; random samples of hits (sample A, 300 hits from the first edition's articles; sample A2, 250 from the update's new articles) and of death sentences (sample V, 150; sample V2, 120). Of the 49,164 hits in `week_one_term_hits.csv`, 13,047 were read by hand. The rest keep the computer's code.
- Anniversaries: every condemnatory, neutral and legal hit that could change an article's flag, every claim hit, every Gaza toll hit and every headline and standfirst. For the added outlets every condemnatory, neutral and legal hit was read, whether or not it could change a flag.
- Gaza fortnights: every headline and lede, every hit whose computer or lexicon referent was the Israeli occupation's conduct or unclear, every hit the computer coded as the outlet's own voice, every headline hit, and random checks of the rest. In the added outlets' frames every hit the computer put in the outlet's own voice was read, for both referents.
- False claims: every screened sentence, every article's challenge code and every edited case.
- The dead: the memorial biographies were searched for security and armed roles, and every find was read by hand; everyone else follows fixed rules (see the file's section).
- The Hannibal Directive: every item found was read in full (see the file's section).

### Measured error rates

We measured the computer's codes against hand codes and corrected where the error rate passed 5%:

- Week one, random sample A (287 valid hits): the computer's referent was wrong on 11.8% and its voice on 23.3%. Random sample A2 of the update's new hits (223 valid): 13.5% and 15.7%. That is why every decisive hit was read by hand. Of the condemnatory or legal hits in the samples that no targeted queue had read (189 in A, 120 in A2), none was an own-voice use about the Israeli occupation's conduct (95% upper bounds 1.6% and 2.5%).
- Week one, every headline and lede hit (3,831): the hand code differed from the computer's on 19.2% (referent) and 18.9% (voice). Every one is hand-coded.
- Week one, random samples V (136 death sentences) and V2 (105): the computer's victim side was wrong on 46.3% and 51.4%. The death-sentence shares are therefore calibrated on the two samples together (see `week_one_victim_sentences.csv`).
- Anniversaries, the first 14, a random sample of 210 hits: referent wrong on 14.4% (25 of 174 with a referent), voice wrong on 14.8%. Hence the full review described above. The toll detector read the source wrongly on 8% of its sample, so it was fixed and every toll hit was read. The added outlets, all 1,165 condemnatory, neutral and legal hits read by hand: the computer's referent was wrong on 17.5% and its voice on 12.2%.
- Gaza fortnights, the first 14: on 150 random hits the computer's referent was wrong on 43.3% and its voice on 22.0%; after the targeted reading, 88 random non-opinion hits not in the outlet's own voice held no own-voice use about the Israeli occupation's conduct (95% upper bound about 3.4%). Own-voice 7 October uses were accepted from the computer after a random check of 60 (52 correct, 87%), and every counted article for the Daily Mail, The Sun, Fox News and The Times was read (131 articles; 10 removed as quotations or off the subject). The added outlets, 150 random hits: referent wrong on 48.7%, voice on 22.0%; of the 351 hits in their frames that the computer put in the outlet's own voice, all read by hand, its referent was wrong on 35.3% and its voice on 41.6%.

## The files, column by column

### week_one_articles.csv

The week of 7 to 13 October 2023, at all 32 outlets. An article is in the file when it was published in the window by the outlet's own date, is relevant (its title, address or first 300 words mention the words "Israel", "Gaza", "Hamas", "Palestinian", "West Bank" or "7 October"), is in English, is not a second copy, and was read and coded by a person. Video pages are left out of every share and out of the file.

- `article_id`: NewsCord's identifier for the article.
- `outlet`, `published`, `url`, `title`: the outlet, its publication date, the address and the outlet's own headline.
- `outlet_added`: the outlet was added on 8 October 2026.
- `counted_in_week_one_shares`: the outlet is one of the 28 counted in the week's shares. `n` for the Financial Times, The Wall Street Journal, Bloomberg and Channel 4 News.
- `article_type`: `news`, `live` (a live blog, counted as one article), `opinion` or `other`.
- `measured_set`: `text_measures` for the articles the page's week-one shares are counted over (news, live and opinion pages with readable text: 4,367 in the counted outlets, 4,392 in all 32); `volume_only` for articles counted only in the volume measure (articles mainly about each side's dead).
- `volume_only_reason`: why a row is volume only.
- `subject_code` and `subject_meaning`: what the article is mainly about, coded by hand from its headline and opening: `O7V` Israeli victims of 7 October (the dead, the captives and the survivors); `O7A` the attack itself and the attackers; `GV` Palestinian victims of the Israeli occupation's attacks; `GM` the Israeli occupation's military campaign; `B` the victims of both; `X` something else.
- `text_basis`: whether the text we coded was the live page or an Internet Archive copy.
- `fetch_route`: how the page was read: a plain live request, a rendering crawler, a desktop browser, or the Internet Archive.
- `words`: words in the text read.
- `own_voice_condemn_oct7`: the article uses a condemnatory term for the 7 October attack in its own voice. `own_voice_condemn_oct7_terms` lists the families.
- `own_voice_condemn_or_legal_israeli_occupation`: the article uses a condemnatory or legal term for the Israeli occupation's killing of Palestinians in its own voice. `own_voice_israeli_occupation_scope` is `israeli_occupation_only` when at least one such hit is about the Israeli occupation's conduct alone, `both_sides` when every such hit applies to both sides together. `own_voice_israeli_occupation_terms` lists the families.
- `any_voice_condemn_oct7`, `any_voice_condemn_israeli_occupation`, `any_voice_legal_israeli_occupation`: the same terms in any voice, quotations included. These include hits the computer coded without a person reading them.
- `retaliation_framing_own_voice`: the article frames the Israeli occupation's attacks as retaliation or self-defence in its own voice.
- `hamas_run_outlet_words`: the article calls Gaza's ministry, government or authorities Hamas-run (or controlled, or led) outside a quotation.
- `in_week_one_gaza_set`: the article is also in `week_one_gaza_articles.csv`.

### week_one_term_hits.csv

Every hit in an article of `week_one_articles.csv`. Hits on video pages (257) are left out, as video pages are out of every share. A hit's code is the hand code where a person read it, otherwise the computer's.

The sentence is given only for hits a person read. A machine-coded hit carries no sentence text, because printing every coded sentence would reproduce whole articles.

- `hit_id`: identifier of the hit.
- `outlet`, `outlet_added`, `counted_in_week_one_shares`, `url`, `article_type`: the article, as in `week_one_articles.csv`.
- `article_in_text_measures`: the article is one with text measures.
- `location`: `title` (headline), `lede` (standfirst and first paragraph) or `body`. `paragraph`: the paragraph number in the text read.
- `term_group`: `condemn`, `legal`, `neutral`, `framing`, `verb_neutral` (killed, died), `claim_false`, `claim_wide` (wider net for the same false claims), `claim_beyond`, `claim_sv` (sexual violence consistent with the UN findings), `correction`.
- `term_family`: the family (see "The words we coded"); `matched_word`: the word as printed.
- `valid`, `referent`, `voice`: the final code.
- `reviewed_by_hand`: a person read the hit. `review_queue`: which reading: `headline_lede` and `headline_lede_check` (every headline and lede hit, and its second check), `claims` (every claim sentence), `isr`, `isr2`, `isr3` (every hit that could set an own-voice flag for the Israeli occupation's conduct), `c23` (every hit in the same week's Gaza articles), `own7` (own-voice 7 October flags outside the Daily Mail, Fox News and The Sun), `sampleA` and `sampleA2` (the random samples).
- `machine_referent`, `machine_voice`: the computer's code before any reading, so the two can be compared.
- `claim_family`: for claim hits, the claim: `BB` beheaded babies or children, `40` "40 babies", `OV` a baby in an oven, `FO` a foetus cut from its mother, `CL` babies on a clothesline, `BC` bound children, `MR` rape claims beyond the evidence, `BBA` ambiguous, `NA` not a claim; combinations are joined with `+`.
- `claim_challenged`: `Y` the article challenges the claim, `H` only Hamas or protesters dispute it, `N` nothing challenges it.
- `sentence`: the sentence, for hand-read hits only.

### week_one_victim_sentences.csv

Every body sentence in the articles with text measures that carries a death verb (killed, died and their forms) or murder, massacre, slaughter or butcher. The page's figure (39.7% of about 4,450 own-voice sentences about the Israeli dead against 1.3% of about 640 about the Palestinian dead used murder, massacre, slaughter or butcher) is computed from this file, over the 28 counted outlets.

- `sentence_id`, `outlet`, `outlet_added`, `counted_in_week_one_shares`, `url`, `article_type`.
- `strong_verbs`: murder, massacre, slaughter or butcher forms in the sentence. `neutral_verbs`: killed or died forms.
- `victim_side`: whose death the sentence reports, where a person decided it: `ISR` Israeli victims, `PAL` Palestinian victims, `BOTH`, `COMBATANT`, `UNCLEAR`, `NA` (no victim). `side_source` says who decided: a person (from the victim sentence queue, random samples V and V2 or a hand-read term hit) or the machine. `machine_side` is the computer's side.
- `voice`: whose words the sentence is.
- `p_israeli_victims`, `p_palestinian_victims`, `p_both`: the weight the sentence carries for each side. A hand-coded sentence carries 1 for its side. A machine-coded sentence carries the share of sample V and V2 sentences with the same machine side (and the same verb strength) that people found to be on each side; this is the calibration.
- `sentence`: the sentence, for hand-coded rows only.

### week_one_gaza_articles.csv and gaza_fortnights_2024_2025_articles.csv

Articles about the Israeli occupation's killing, wounding, starving, besieging or displacing of Palestinians. An article enters when it is readable, passes an automatic rule (dated in the window; Gaza, Palestinians, the West Bank or a place there in the title, address or first 100 words; a word for killing, starving or displacing in the title or first 300 words; English) and a hand check (the headline, standfirst or first two paragraphs report the Israeli occupation's killing, wounding, starving, besieging or displacing of Palestinians).

The week-one file holds all 722 such articles from 7 to 13 October 2023 at the 32 outlets. For the comparison across years, each outlet has an equal-size sample: the outlet's count of week-one articles about 7 October, capped at 40 and floored at 10, drawn from its readable pages sorted by address and shuffled with random seed 7. That count is final for the outlets whose pages we read in full by the freeze and provisional for the rest. `in_equal_n_sample` marks the 602 drawn in 2023, 582 of them at the 27. The 2024 and 2025 file holds only the sampled articles at the 27, fewer where fewer existed among the pages read.

- `article_id`, `window` (`C23`, `C24`, `C25`), `year`, `outlet`, `outlet_added`, `published`, `url`, `title`, `article_type`, `fetch_route`, `words`.
- `counted_in_week_one_shares` (week-one file only): the outlet is one of the 28 counted in the first week's shares.
- `in_equal_n_sample` (week-one file only): the article is in the 2023 equal-size sample.
- `in_gaza_trend_panel` (week-one file only): the outlet is one of the 27 the trend across the three years pools.
- `headline_agency` and `lede_agency`, coded by hand from the headline and the lede (the first paragraph, or the first two when the first has under 20 words), on the first clause that reports the killing, wounding, starving or displacing of Palestinians:
  - `ACTIVE`: the Israeli occupation, its army, forces, jets or settlers, or an explicitly Israeli action, is the grammatical subject ("Israeli strike kills 20 in Gaza school").
  - `NAMED`: the Israeli occupation is named as the cause but is not the subject ("20 killed in Israeli strike").
  - `AGENTLESS`: the killing, wounding, starving or displacing is stated and nobody is named as doing it ("Strike on Gaza school kills 20").
  - `NONE`: the span states no killing, wounding, starving or displacing of Palestinians.
- `headline_states_killing_wounding_starving_or_displacing_of_palestinians`: `headline_agency` is not `NONE`.
- `headline_carries_israeli_occupation_justification`, `lede_carries_israeli_occupation_justification`: the span carries the Israeli occupation's justification or its claim about the target or the dead ("targeting Hamas", "militants", "a Hamas command centre").
- `hamas_run_outlet_words`: the article calls Gaza's ministry, government or authorities Hamas-run, controlled or led outside a quotation.
- `own_voice_condemn_israeli_occupation`, `own_voice_legal_israeli_occupation`: a condemnatory or legal term for the Israeli occupation's conduct in the outlet's own voice. `any_voice_condemn_israeli_occupation`, `any_voice_legal_israeli_occupation`: in any voice. `own_voice_israeli_occupation_terms`, `any_voice_israeli_occupation_terms`: the families. In these files a hit nobody read takes its referent from the coder run with Gaza as the article's subject, so it can differ from the same hit's code in `week_one_term_hits.csv`.
- `retaliation_framing_own_voice`: retaliation or self-defence framing in the outlet's own voice. `israel_hamas_war_phrase_own_voice`: the phrase "Israel-Hamas war" in the outlet's own voice.
- `headline_terms`: condemnatory, legal and framing terms in the headline, as `family:REFERENT/VOICE`.

### gaza_fortnights_frame_articles.csv

Every readable article at the 27 in the two fortnights that passed the automatic rule above, sampled or not, anniversary pieces about 7 October included. In these articles we searched for massacre, slaughter, butcher, atrocity, pogrom, barbaric, savage and genocide.

- `window`, `year`, `outlet`, `outlet_added`, `published`, `url`, `title`, `article_type`.
- `in_gaza_sample`: the article is also in `gaza_fortnights_2024_2025_articles.csv`.
- `own_voice_oct7`, `own_voice_oct7_terms`: the article uses one of those words for 7 October in its own voice. `own_voice_oct7_coded_by`: `hand`, `machine` or both. At the first 14 every own-voice candidate was put to a person except 7 October uses the computer coded with confidence, which were accepted after the random check described above (147 articles carry a machine code); at the added outlets every one was read.
- `own_voice_israeli_occupation`, `own_voice_israeli_occupation_terms`: the same words, or genocide, for the Israeli occupation's conduct in its own voice. Every one was read by hand.

### gaza_fortnights_contrast.csv

The counts the page prints for the fortnights, rebuilt from `gaza_fortnights_frame_articles.csv`. Rows with `article_type` = `all` are one outlet in one fortnight; rows with `outlet` = `All outlets` are one article type in one fortnight, the 27 together.

- `window`, `year`, `outlet`, `outlet_added` (empty on the `All outlets` rows), `article_type`, `articles` (readable articles in the frame).
- `own_voice_oct7_articles`: articles using one of the words for 7 October in their own voice. `own_voice_israeli_occupation_articles`: articles using one of them, or genocide, for the Israeli occupation's conduct in their own voice.
- `massacre_own_voice_oct7`, `massacre_own_voice_israeli_occupation`, `slaughter_own_voice_oct7`, `slaughter_own_voice_israeli_occupation`, `atrocity_own_voice_oct7`, `atrocity_own_voice_israeli_occupation`, `barbaric_own_voice_oct7`, `barbaric_own_voice_israeli_occupation`, `genocide_own_voice_oct7`, `genocide_own_voice_israeli_occupation`: the same, word by word (outlet rows; for the article-type rows only the genocide count for the Israeli occupation's conduct is given).

### anniversary_articles.csv

Every article in the anniversary universe at the 27: published on the outlet's own 7 October in 2024, 2025 or 2026, relevant, readable, in English and not a second copy of the same headline at the same outlet. Types come from the address, so a signed analysis filed under a news address counts as news; a Metro readers' letters page filed under a news address was corrected by hand to opinion. The first 14's 2026 day stops at 16:55 UTC; the added outlets' runs to midnight UTC (see "The freezes").

- `anniversary`, `outlet`, `outlet_added`, `published`, `url`, `title`, `article_type` (`news`, `live`, `opinion`, `video`, `other`).
- `wire_copy`: `AP` where the outlet published AP copy under its own name.
- `text_basis`: the text read (live page, NewsCord's archive copy, Internet Archive copy, desktop browser, rendering crawler).
- `about_the_subject`: the subject is in the title or address, or mentioned at least twice in the first 300 words.
- `headline_focus` and `headline_focus_meaning`: what the headline and standfirst are about, coded by hand: `ISR` Israeli victims, captives, survivors and the commemoration of 7 October; `PAL` Palestinians killed, starved or displaced by the Israeli occupation, and Gaza's ruin; `BOTH`; `DIASPORA` antisemitism and Jewish communities abroad; `PROTEST`; `CONFLICT` the war's developments, Lebanon, Iran, diplomacy, politics, arms; `INCIDENTAL` a passing mention.
- `own_voice_condemn_oct7`: a condemnatory term for 7 October in the outlet's own voice. `any_voice_condemn_oct7`: in the outlet's own voice, a quotation or an attribution.
- `own_voice_neutral_term_for_attackers`: militants, gunmen or fighters for the attackers, in the outlet's own voice.
- `own_voice_condemn_israeli_occupation`: a condemnatory term for the Israeli occupation's conduct in the outlet's own voice, read by hand.
- `genocide_own_voice_israeli_occupation`, `genocide_quoted_israeli_occupation`, `genocide_attributed_israeli_occupation`: genocide about the Israeli occupation's conduct, by voice.
- `israel_hamas_war_phrase_own_voice`, `war_in_gaza_phrase_own_voice`: those phrases in the outlet's own voice.
- `gaza_toll_stated`: outside a quotation, the article gives the Gaza death toll of the day (a total within 80% to 115% of Tech for Palestine's cumulative count for that day: 41,909, 67,173 and 74,250) or "tens of thousands". `gaza_toll_source`: the source it gives for the figure: `hamas_run` (Hamas-run, controlled or led), `ministry_unqualified` (the health ministry or local officials with no mention of Hamas), `un_or_other`, or `none`. An article with any Hamas-run toll counts as `hamas_run`.
- `headline_terms`: terms in the headline and standfirst, as `family:REFERENT:VOICE`.

### anniversary_claims.csv

Every sentence found around the anniversaries that carries one of the false claims or a claim beyond the evidence: in the 27's articles and, for comparison, in every other outlet in GDELT's claim lists for 5 to 9 October (gathered for the first edition).

- `scope`: a panel outlet, or another outlet from the GDELT lists.
- `outlet`, `outlet_added` (panel outlets only), `anniversary`, `published`, `published_on_7_october`, `url`, `title`, `article_type`.
- `claim`, `claim_group` (`claim_false` or `claim_beyond`), `sentence`, `voice`.
- `found_by`: `lexicon`, the term list, as for the first 14; or `wider search`, the added outlets' second search, for beheaded children as well as babies and for wider forms of the other false claims. The same wider search over the first 14's anniversary articles found nothing new.
- `about_7_october`: the sentence is about 7 October. Empty on the added outlets' rows, whose coding did not record it apart.
- `put_to_reader_as_true`: in the outlet's own voice, or quoted or attributed with nothing in the article challenging it.
- `challenged_in_article`: for panel rows, whether the article challenged the claim.

### false_claim_articles.csv

Articles that carried a false claim or a claim beyond the evidence: at the first 14, published 7 October to 31 December 2023 (and three follow-ups outside that window, marked `in_claims_window` = `n`); at the added outlets, published 7 to 13 October 2023, the only week the update read for them. We screened every candidate article we could read in the window (articles whose address or title paired an Israeli or Gaza context word with a claim cue, articles that published catalogues of the claims pointed to, and GDELT results; `completeness.csv` counts them), and a person coded every screened sentence. The fact base decides what is false:

- False claims: babies, infants or children beheaded (`F_BEHEAD`; the beheading of some adults and soldiers is documented by the UN Commission of Inquiry and is not counted), "40 babies" (`F_40`), a baby in an oven (`F_OVEN`), a foetus cut from its mother (`F_FOETUS`), babies on a clothesline (`F_LINE`). Two variants are reported apart: babies' throats cut (`F_THROAT`) and babies killed at Kfar Aza (`F_KFAR`; the youngest person killed there was 14).
- Beyond the evidence when stated as fact: "mass" (`B_MASS`), "systematic" (`B_SYST`), "weapon of war" (`B_WEAPON`), "widespread" (`B_WIDE`) and "ordered" or "planned" (`B_ORDER`) rape. Sentences consistent with the UN findings are not counted.

Columns:

- `outlet`, `outlet_added`, `url`, `published`, `published_basis` (where the date came from), `in_claims_window`, `article_type`.
- `claims_window`: the window searched at that outlet.
- `title_first`, `title_current`: the headline in the first capture and today.
- `claim_codes`: every claim code in the article.
- `false_claim_voice`, `beyond_evidence_voice`, `variant_voice`: the strongest voice in which the article carries each kind (`OWN` above `QUOTE` above `ATTRIBUTED` above `HEDGED`).
- `claim_in_headline`: a claim is in the headline or standfirst.
- `challenged_in_article`, `challenged_at_first_publication`: whether the article itself challenged the claim: `y` (it says the claim is false, unconfirmed, withdrawn or disputed), `h` (a generic could-not-verify line only), `d` (only Hamas's denial), `r` (only a third party's challenge, dismissed), `c` (presented as someone's rhetoric), `n` (nothing). Only `y` counts as a challenge. `challenge_sentence`: the challenging sentence.
- `in_scorecard`: the article is in the page's scorecard: published in the window, carrying a false claim in its own voice, or in a quotation the article did not challenge as first published. `in_beyond_evidence_table`: the same rule for claims beyond the evidence.
- `scorecard_bucket`: `corrected` (a correction, clarification or editor's note about the claim), `removed` (page gone, or redirected elsewhere), `softened` (silently edited so the claim is gone or weaker), `unchanged` (word for word since a capture from the first 14 days after publication), `standing` (still on the page with no notice, but either edited in a way that kept or strengthened the claim, or with no capture from the first 14 days to compare), `unreadable` (no current text, or only part of it).
- `first_capture_timestamp`, `first_capture_utc`, `first_capture_url`: the Internet Archive capture used as the first-published version (the capture closest to the start of the publication day, or its end for a live blog), and a link to it. `first_capture_days_after_publication`: a capture more than 14 days after publication is not treated as first-published. `first_capture_partial`: the capture held under half the words of today's page. `first_capture_words`, `current_words`.
- `current_read_route`, `current_http_status`, `current_read_date`: how and when the current page was read (live request, rendering crawler, desktop browser, the latest Internet Archive capture, or still waiting at the freeze).
- `current_status`: the final status, read by hand where it needed a person: `unchanged`, `present_first_late`, `present_first_unread`, `edited_claim_kept`, `edited_strengthened`, `claim_added_after_first_capture`, `edited_softened`, `corrected_notice`, `current_unreadable`, `current_partial`. `current_status_automatic`: the status before the hand reading.
- `headline_changed`: the headline changed after the first capture.
- `correction_notice`, `correction_notice_date`: any correction, clarification or editor's note on the page, verbatim.
- `silent_edit`: the claim text changed after publication with no notice (`unknown` where no early capture could be compared).
- `claim_sentences_first`: each claim sentence as first published (or, where no early capture could be read, as found), with its code and voice. `claim_sentences_current`: what the current page says for each, in the same order: `same`, `minor edit`, `edited`, `not found`, `in page source only` or `no current text`, with the similarity score and the closest current text. Sentences are separated by ` || `.
- `claim_sentences_seen_only_now`: claim sentences in today's page that the first capture did not have. `current_sentences_not_in_first_capture`: today's claim sentences that changed after the first capture, with what the first capture had.

In the false-claim files the outlets' own em and en dashes are shown as a spaced hyphen; every other quotation in the database is verbatim.

### false_claim_own_debunks.csv

The sentences in which an outlet itself said a claim was false, unconfirmed, withdrawn or disputed, in the claims window: sentences coded as the outlet's own debunk, and the challenge sentences of articles coded `y` in `false_claim_articles.csv`.

- `outlet`, `outlet_added`, `claims_window`, `published`, `url`, `title`, `how_coded`, `sentence`.

### dead_of_7_october.csv

Unchanged from the first edition, byte for byte. The 1,144 people who died on 7 October 2023, from the oct7database name lists (github.com/yuval-harpaz/alarms), joined to the official memorial pages. People who were taken captive alive, who died later or who died on 8 or 9 October are not in this file.

- `person_id`: the person's identifier in oct7database.
- `first_name`, `middle_name`, `last_name`, `name_hebrew`, `gender`.
- `age`: the age printed on the memorial page where there is one, otherwise the age in oct7database, as recorded. `age_source` says which.
- `country`: nationality as recorded. `residence_as_recorded`: place of residence, in Hebrew as recorded. `residence_community`: the community in English where it is one of the border communities the study looked at.
- `role_as_recorded`: the role the name list gives (civilian, soldier, police officer, emergency squad member, medical staff, Shin Bet, firefighter).
- `party_or_festival`: Nova or Psyduck, for people killed at the two parties.
- `place_of_death`, `place_of_death_hebrew`, `date_of_death`.
- `body_taken_to_gaza`: the person was killed on 7 October and their body taken into Gaza.
- `official_list`: the memorial list the person's page is on: `National Insurance` (the National Insurance Institute's memorial list), `IOF` (the Israeli occupation forces' list of their fallen), `Police`, `Shin Bet`, `Defence Ministry (Izkor)`, or `None`.
- `armed_as_recorded`: whether the memorial page records the person as armed: `yes`, `no`, `no firearm` (fought with a knife or similar), `no record`, `unknown`; empty where the person has no hand code.
- `newscord_detail_code`: NewsCord's finer code: `ON_DUTY`, `EMERGENCY_SQUAD`, `OFF_DUTY_SOLDIER`, `OFF_DUTY_POLICE`, `OFF_DUTY_SHIN_BET`, `RESERVIST_NOT_YET_WITH_UNIT`, `CIVILIAN_WHO_FOUGHT`, `CIVILIAN_SECURITY_ROLE`, `MEDIC_RESCUER_FIREFIGHTER`, `FESTIVAL_RESERVIST_NOT_CALLED_UP`, `FESTIVAL_STAFF`, `FESTIVAL_SECURITY_STAFF_UNARMED`, `FOREIGN_NATIONAL`, `NO_RECORDED_ROLE`.
- `newscord_group_key` and `newscord_group`: the six groups the page prints: `onduty` (soldiers, police and Shin Bet officers on duty), `squad` (members of the armed emergency squads), `offduty` (off-duty soldiers, police and Shin Bet officers, and reservists not yet with their units), `fought` (civilians who fought the attackers, or held a civilian security role), `medics` (medics, rescuers and firefighters), `norole` (no recorded armed or security role, which includes foreign nationals, festival staff, unarmed festival security staff and reservists at the festival who had not been called up).
- `recoded_by_verification`: for the four people whose code NewsCord's verification of 7 October 2026 changed, what changed and why.
- `memorial_page`: the link to the person's memorial page.

How the code was given. The first rule that fits decides: a hand code read from the memorial biography or the IOF memorial page (squad members, civilians who fought, civilian security roles, festival staff and guards, reservists, and others checked by hand, each citing its evidence); on the IOF list, an emergency squad role gives `EMERGENCY_SQUAD`, a civilian role gives `OFF_DUTY_SOLDIER`, any other role `ON_DUTY`; on the police or Shin Bet list, a civilian role gives off duty and any other role on duty; medical staff and firefighters give `MEDIC_RESCUER_FIREFIGHTER`; a nationality other than Israeli gives `FOREIGN_NATIONAL`; everyone else is `NO_RECORDED_ROLE`. Nothing infers a security role from age, sex or past military service.

The verification read 12 records against the live memorial pages and found two codes wrong (Nitzan Goldenberg and Aviv Eliyahu); reading the Ofakim and Kfar Aza records found two more (Refael Ariel Guri and Shachaf Bergstein). All four are recoded here, as on the page. The IOF list's pages could not be re-read automatically, so the squad and off-duty soldier codes taken from it were not independently checked.

### hannibal_naming.csv

Which outlets named the Hannibal Directive (under that name, or as the Hannibal protocol or procedure) from 7 October 2023 to 8 October 2026. Two kinds of row:

- `row_type` = `page_table`: the page's table, row for row (13 rows). Haaretz is not a panel outlet; it is in the table for comparison. The added outlets where we found nothing share one row.
- `row_type` = `item`: each item we found at the added outlets (7, of which 6 fall in the window; the other is the last time before the window that Channel 4 News named the directive, which the page's table gives). The searches of the first 14 and Haaretz are not in this database; their findings are the `page_table` rows.

How we searched the added outlets: every page the census had fetched, for any outlet and any study, searched for "hannibal" and each match read in full; a web search per outlet for the three names; and each outlet's own site search, which returned no usable results for any of them. The web search was unreliable: twice it reported nothing where the census's own pages held an item. A zero is therefore "not found despite the search", not a confirmed absence; the Financial Times and Bloomberg, which are paywalled, are the zeros we trust least.

- `row_type`, `outlet`, `outlet_added`.
- `items_found`, `detail`, `last_named_before_window` (`page_table` rows): the items found in the window, what they are, and the last time before the window that the outlet named the directive, where we found one.
- `date`, `url`, `headline` (`item` rows): the piece, with the outlet's own headline.
- `where_in_piece`: where in the piece the directive is named. `whose_words`: whether the naming is in the outlet's own words or in someone it quotes.
- `quotation`: the passage naming it, verbatim.
- `found_via`: a search of every page the census read, or a web search.
- `confidence`: `confirmed` where we read the piece itself or a syndicated copy of it.
- `note`: NewsCord's note on the item.

### completeness.csv

For each outlet and study, what we found against what we read by the freeze: seven rows for each of the 32 outlets.

- `outlet`, `outlet_added`, `study` (`week_one`, `false_claims`, `anniversary`, `gaza_fortnight`), `window`.
- `pages_found`: week one, the candidate pages that passed the keyword filter; false claims, the pages screened (for the added outlets, the relevant first-week articles the census listed); anniversaries, the candidate pages (every listed page whose address or title carried a subject word, every page from NewsCord's archive, every GDELT result and every BBC archive page; for the added outlets also every untitled Financial Times page and every Daily Mirror news page dated on the day, because their addresses carry no words to filter on); Gaza fortnights, every page in the outlet's listing for the fortnight.
- `pages_read`: the pages read by the freeze (for the added outlets' false claims, the first-week articles read with text). `read_share_pct`: the two as a share.
- `articles_measured`: week one, articles with text measures; false claims, articles in the scorecard; anniversaries, articles in the universe (every type); Gaza fortnights, articles in the random sample.
- `articles_target`: the Gaza sample's target size for the outlet.
- `in_pooled_figures`: the row's outlet is counted in the page's figures for that study.
- `page_shows`: what the page's completeness table prints in that cell. "Not found in what we could read" marks a cell where we found no article in the pages we could read; "Not studied" marks a study that does not count the outlet.
- `note`: why an outlet is left out of a study's figures, what a cell covers.

## Rebuilding the page's figures

Each step below uses only these files.

1. The first week's shares. In `week_one_articles.csv`, keep `measured_set` = `text_measures` and `counted_in_week_one_shares` = `y`: 4,367 rows. `own_voice_condemn_oct7` = `y`: 2,159 (49.4%). `own_voice_condemn_or_legal_israeli_occupation` = `y`: 30 (0.7%), of which `own_voice_israeli_occupation_scope` = `israeli_occupation_only` 17 (10 opinion, 7 news) and `both_sides` 13. Group by `outlet` for the per-outlet table. The four outlets left out (`counted_in_week_one_shares` = `n`) have 14 (the Financial Times), 5 (The Wall Street Journal), 5 (Bloomberg) and 1 (Channel 4 News) articles with text measures. The same 2,159 and 30 follow from `week_one_term_hits.csv` alone: articles with a valid condemnatory hit, referent `OCT7`, voice `OWN`; and articles with a valid condemnatory or legal hit, referent `ISRAELI_OCCUPATION_ON_PALESTINIANS` or `BOTH`, voice `OWN`. The first edition's 2,014, 1,128 and 20 come the same way from `first_edition/week_one_articles.csv`.
2. The words for the dead. In `week_one_victim_sentences.csv`, keep `voice` = `OWN` and `counted_in_week_one_shares` = `y`. The sum of `p_israeli_victims` is 4,450.2, about 4,450; the sum over rows with `strong_verbs` not empty, divided by it, is 39.7%. The same with `p_palestinian_victims` gives 638.2, about 640, and 1.3%.
3. Articles per death. In `week_one_articles.csv` (both measured sets), keep `counted_in_week_one_shares` = `y`: 4,664 articles. `subject_code` `O7V` counts 831 and `GV` 289. Against the 1,176 dead on the Israeli side named in the memorial lists by 13 October 2023 (1,144 on 7 October and 32 on 8 and 9 October) and the 1,799 Palestinians killed in Gaza by 13 October that OCHA reported (Flash Update #7), that is 707 and 161 articles per 1,000 dead, a ratio of 4.4.
4. The BBC's Gaza headlines. In `week_one_gaza_articles.csv`, `outlet` = `BBC`: 24 articles, 13 whose headline states the killing, wounding, starving or displacing of Palestinians, 10 of them `AGENTLESS`. For the counted outlets (`counted_in_week_one_shares` = `y`): 718, 330 and 147.
5. The Gaza trend. Agentless headlines among headlines stating the killing, wounding, starving or displacing of Palestinians, and "Hamas-run" in the outlet's words: 2023 from `week_one_gaza_articles.csv` with `in_equal_n_sample` = `y` and `in_gaza_trend_panel` = `y` (110 of 256; 6 of 582); 2024 and 2025 from `gaza_fortnights_2024_2025_articles.csv` by `year` (61 of 135 and 54 of 89; 29 of 219 and 29 of 221). Group by `outlet` for the trend with each outlet weighted equally; an outlet with no headline stating the killing in a year has no share that year.
6. The fortnights. In `gaza_fortnights_contrast.csv`, the `All outlets` rows with `article_type` = `news` give 728 news articles with 101 against 4 (2024) and 1,392 with 118 against 9 (2025). The `all` rows for 2025 give the Daily Mail 30 and 0 of 175, The Sun 18 and 0 of 49, Fox News 14 and 0 of 81, The Times 4 and 0 of 30, the BBC 0 and 0 of 22, the New York Post 18 and 0 of 101 and the Daily Express 16 and 0 of 90. Every count can be rebuilt article by article from `gaza_fortnights_frame_articles.csv`.
7. The anniversaries. In `anniversary_articles.csv`, by `anniversary`, keep `article_type` = `news` for the pooled news figures (2024: 324 articles, 155 with `own_voice_condemn_oct7`, 2 with `own_voice_condemn_israeli_occupation`, 0 with `genocide_own_voice_israeli_occupation`, 117 with `gaza_toll_stated`, 25 of them `hamas_run`; 2025: 209, 81, 1, 0, 60 and 16; 2026: 113, 31, 2, 0, 34 and 16). Use every row for the headline focus (`ISR` against `PAL`: 131 against 25 of 385, 71 against 15 of 247, 42 against 6 of 137) and for the BBC's figures. In `anniversary_claims.csv`, the rows with `scope` = panel outlet, `published_on_7_october` = `y` and `put_to_reader_as_true` = `y` are the 4 false claims repeated on the three days: 2 at the first 14, 2 at the added outlets.
8. "0 of 144" (false claims). In `false_claim_articles.csv`, keep `in_scorecard` = `y`: 144 articles, 93 at the first 14 (`outlet_added` = `n`, published 7 October to 31 December 2023) and 51 at the added outlets (`outlet_added` = `y`, published 7 to 13 October 2023). Count `scorecard_bucket` by `outlet` for the scorecard: 0 corrected, 0 removed, 3 softened, 124 unchanged, 14 standing and 3 unreadable. For the first 14, `article_type` gives 73 news, 12 opinion, 7 live and 1 video; `false_claim_voice` = `OWN` 54; `claim_in_headline` = `y` 8. Silent edits: keep `silent_edit` = `y` and leave live blogs aside (`article_type` not `live`): 26 articles, of which 5 softened the claim (`current_status` `edited_softened`) and 21 kept or strengthened it: 14 added a claim or made it stronger (`edited_strengthened`, `claim_added_after_first_capture`) and 7 kept it in new words (`edited_claim_kept`). At the first 14 alone that is 1 against 14, as in the first edition. The distinct addresses in `false_claim_own_debunks.csv` are the 49 articles in which an outlet itself said a claim was false or unconfirmed: 30 at the first 14 and 19 at the added outlets.
9. Who died. In `dead_of_7_october.csv` (1,144 rows), count `newscord_group_key`: `onduty` 286, `squad` 45, `offduty` 29, `fought` 19, `medics` 11, `norole` 754. The on-duty dead by `official_list`: IOF 225, Police 57, Shin Bet 4. The four legal readings: strictest, `norole` plus `medics` (765); everyone not on the IOF list (850); that plus squad members on the IOF list (895); everyone except those on duty on the IOF list (919).
10. The Hannibal Directive. In `hannibal_naming.csv`, the `page_table` rows are the page's table. For each added outlet, its `item` rows dated 7 October 2023 to 8 October 2026 give the number in `items_found`.
11. Completeness. `completeness.csv`, column `page_shows`, is the page's section 07 table, cell by cell.

## What is not in this database

- The figures the page takes from other sources: the IOF's orders and investigations, the UN Commission of Inquiry, OCHA, B'Tselem, HaMoked, DCI-Palestine, Peace Now, the Israel Prison Service, IPSO and the BBC's Executive Complaints Unit, FAIR and The Intercept.
- GDELT's count across all English-language media (the ratio of 4.4, on the same two tolls).
- The searches for the Hannibal Directive at the first 14 outlets and Haaretz (their findings are the page's table, in `hannibal_naming.csv`).
- The pages read at the Financial Times, The Wall Street Journal and Bloomberg for the anniversaries and the fortnights: every one was a teaser, and none counts in any figure.
- The text of pages, beyond the hand-read sentences and the short claim quotations given here.
- The researchers' working notes.
