# K16c – independent fact-check of the seven English articles (30.10.2026)

Articles: `kontroll/reanalyser/artikler/{norat2002,larsson2006,wcrf2007,cup2010,cup2016,update2026,evidence_1997_iarc114}.md` (versions on disk 30.10, update2026.md modified 13:19).

Method: (1) every "x.xx (x.xx–x.xx)" triple in each article was extracted by script and looked up in an index built from `all.json`, `update2026.json`, the run `.md` files, `estimates.csv`, `synthesis_extraction.csv`, `derived_per100g.csv`, `cup_extraction.csv` and `kontroll/evidens/`. Every triple not found there was traced by hand to the primary/synthesis text (all traced; listed in the tables). (2) Every quotation in double quotes was searched (normalised) in the primary texts under `drive/`; the ones that did not match automatically were read by hand in the two-column text (all were verbatim apart from what is noted). (3) Characterisations, k, I², weights, LOO, case counts and page references were checked against the run files, `estimates.csv`/`studies.csv` and the primary texts. (4) Cross-article consistency. (5) Overreach. Pools I quote that are not in the run files were recomputed with my own DerSimonian–Laird script (scratchpad), which reproduces the run files exactly (e.g. Norat 1.27 (1.11–1.45); update size subgroups 1.09 (1.03–1.16) / 1.15 (0.99–1.33)).

## Summary of verdicts

| Article | FEIL | OVERDREVET | IKKE VERIFISERBAR | OK (checked) |
|---|---|---|---|---|
| norat2002 | 4 | 1 | 0 | 24 |
| larsson2006 | 2 | 1 | 0 | 22 |
| wcrf2007 | 3 | 2 | 1 | 24 |
| cup2010 | 2 | 1 | 0 | 22 |
| cup2016 | 2 | 1 | 0 | 27 |
| update2026 | 7 | 2 | 3 | 22 |
| evidence_1997_iarc114 | 5 | 1 | 0 | 18 |

**Most serious finding:** the case counts in update2026 Table 2 are hard-coded in `scripts/update2026.py` (not read from the database), and 8 of 20 are not the counts in the primary papers. Several belong to a different publication or a different analysis (UK Biobank 2,609 is Bradbury 2020's count, not Knuppel 2020's). The total is about 45,100, not 46,450. The two study-size subgroups change too (FEIL P1–P2).

---

## 0. Cross-article checks (point 4 of the brief)

| # | Item | Result |
|---|---|---|
| X1 | 46 233 cases | **Superseded.** The rettelogg R80 (K16a) corrected the BWHS count 347 → 564, so 46 233 → 46 450. All md articles say 46,450 consistently. But (a) 46,450 is itself wrong (see P1; ≈45,123), and (b) the HTML has not been rebuilt: `html/artikler/update2026.html` (lines 15 and 622), `html/index.html` (l. 118) and `html/oversikt.html` (l. 161) still say 46 233, even though R80 says the HTML was corrected. The HTML table total already says 46,450, so the HTML page contradicts itself. |
| X2 | 20 cohorts | OK in update2026 (k = 20, json). |
| X3 | 16 731 screened | OK (SEARCH_LOG: 8,119 + 9,242 − 630 = 16,731). |
| X4 | 1 152 multiverse | OK twice, but these are **two different multiverses that happen to have the same size**: update2026 (2×2×3×2×2×2×2×3×2) and CUP 2016 (768 colorectal + 384 colon). Neither article says so. Recommend one clarifying clause in update2026 §3.6 (replacement P14). |
| X5 | Islam/Wada double count fixed (colon 81 % / 44 %) | OK in cup2016 (§2.5, §2.6, §3.5, §4.4 item 3) and in `multiverse_report.md`. **But** `database-samledom.md` §3, which cup2016 and update2026 cite as a source, still has the stale figures: colon "signifikant i 85 %", Islam "54 %", and "Kolon med kohortversjoner (Takachi → Islam 2019, Oba → Wada 2017) 1,14 (1,02–1,27)". That last row is the double-counted version. The samledom needs updating (not an article; reported for completeness). |
| X6 | Database size | Three different versions in three articles. cup2016: 10,247/10,217/18/12. update2026 §2.6: 10,247 from 38 publications in 30 cohorts, 6,055/18/4. update2026 §4.4: 10,217 of 10,247/18/12. evidence: 10,247/38/30, 6,055/18/4. Current `estimates.csv`: **10,383 rows, 53 publications (studies.csv), 10,353 verified, 18 corrected, 12 disputed**. FEIL in three articles (U1, P3, E2). |
| X7 | Rettelogg range | norat2002 and larsson2006 cite "R1–R60". The log now runs to R80. |
| X8 | Name of the Tiemersma 2002 cohort | Five different names. larsson: "Dutch MONICA cohort". wcrf2007 and cup2010: "Dutch Monitoring Project". cup2016: "Dutch MONICA". update2026: "Netherlands MORGEN". evidence: "MORGEN". The source (studies.csv; Tiemersma 2002; IARC text l. 5684) is the **Monitoring Project on Cardiovascular Disease Risk Factors, 1987–1991**. MORGEN (1993–97) is a different, later project, and MONICA is the WHO project. Use "Dutch Monitoring Project (MP)" everywhere (U2, P5, E4; larsson Table 1, see L3). |
| X9 | Shared rows | OK across articles: Bernstein 2015 HvL rows 0.96 (0.76–1.21) / 0.98 (0.75–1.27) (norat, larsson; Suppl. S1/S2 = bernstein2015_229/_337). S5 1.00 (0.89–1.11) → 1.00 (0.87–1.13) per 120 g (wcrf). Papadimitriou 1.06 (0.95–1.14) (cup2010, cup2016, update). Ollberding unprocessed 0.97 (0.83–1.13). Lee 0.80 vs 0.81. Harvard 150 + 205 colon cases; Wei 1,139 + 339 = 1,478. Chan/Larsson colon 1.68 (1.11–2.55). SMC combined 1.69 (1.11–2.56). English ×5.833. |
| X10 | `followups.csv` | Named in norat2002 (§2.5, §5) and larsson2006 (§5). **The file does not exist.** The follow-ups are coded in `scripts/reanalyses.py` (Bernstein rows bernstein2015_229/_337 chosen there). |

---

## 1. norat2002

### FEIL (listed first)

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "with the plotted value instead of the tabulated one, only the upper confidence limit of the pool changes, by 0.01 (R63)" | §3.1, last bullet | own DL recomputation; R63 says only "≤ 0,01" | **FEIL** | With Singh 1.49 (0.94–2.35): 1.274 (1.116–1.455) = **1.27 (1.12–1.46)**. Both limits move by 0.01. |
| NHS/HPFS follow-ups "from `followups.csv`" | §2.5, §5 | file system | **FEIL** | No such file. The follow-ups are in `scripts/reanalyses.py`. |
| "correction log (`rettelogg.md`, entries R1–R60)" | §2.7 | rettelogg | **FEIL** | Outdated; R1–R80. |
| "K11 found no transcription error in any of the 18 Norat rows" | §3.1 | K11_rapport l. 20 ("norat2002 17"); R59 | **FEIL** (minor) | K11 checked 17 rows. The 18th, the Thun figure-read row, was added by K11 (R59). |

### OVERDREVET

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "Adding it gives 1.27 (1.11–1.45), an exact match"; "The reproduction is therefore robust to reading error" | Abstract; §4.4 item 2 | own recomputation | **OVERDREVET** | The exact match needs K11's Thun reading (0.93–1.39) **and** the tabulated Singh value. With the author's high-resolution Thun reading (0.91–1.38) the pool is 1.27 (1.11–**1.46**). With both high-resolution readings (Thun 0.91–1.38; Singh as plotted, 1.49) it is **1.28 (1.12–1.46)**. "Within 0.01" is supported; "exact" and "robust" are not. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| Published 1.27 (1.11–1.45), n 9, p-het 0.45; 1.22 (1.05–1.41), p 0.17; all designs 1.35 (1.21–1.51), 1.24 (1.08–1.41); 1.49 (1.26–1.77), 1.28 (1.11–1.47) | Abstract, §1, §4.2 | 27_Norat2002.txt l. 228–258 | OK |
| 34 case–control + 14 cohort studies, English, 1973–99 | §1 | Norat l. 131–133, 187 | OK ("in English" is on p. 242, counts on p. 243) |
| CUP SLR quotes 1.22 (1.05–1.41) "nine studies" in §2.5.1; 1.19 (0.91–1.55) 8 studies in §2.5.1.3 | §1 | slr2010.txt l. 4829, 4947–4949, 7736, 7810–7814 | OK |
| IARC "a moderate, but significant, increase" (§2.2.3, p. 142) | §1 | 36_IARC114.txt l. 7205 | OK |
| No citation of Norat 2002 in WCRF 2007 | §1 | 32_WCRF2007.txt (only Norat 2004/2005) | OK |
| All Norat method quotes (§2.2) | §2.2 | Norat l. 78–145 (p. 242) | OK, verbatim. One quote ending "…was present." stops mid-sentence (the source continues "58 to incorporate…"); acceptable. |
| 18 `norat2002` rows in synthesis_extraction | §2.3 | csv | OK |
| Thun: 40 rows in db, K13 0 errors; Q5 "fatty meats" 1.21/1.05, no CI, crude, matched on age, race, sex; weekly frequency index; 2,757 vs 1,150 cases; includes liver | §2.3, Table 1, §3.1 | work/thun1992/NOTES.md; R65 | OK |
| Pool values 1.38 (1.16–1.64), 1.27 (1.11–1.45) Q 7.34, 1.35 (1.13–1.62), 1.26 (1.06–1.50) I² 27 %, 1.27 (1.02–1.59) I² 34 %, FE 1.23 (1.08–1.41)/1.28 (1.07–1.54), 1.01 (0.88–1.16) | Table 2, §3.2–3.4 | all.json / norat2002.md | OK |
| Thun 43.5 %, Giovannucci 11.1 %; Bernstein together 62.1 % | §3.2, §3.4 | run file | OK |
| All LOO statements (§3.5) | §3.5 | run file | OK |
| Bernstein 0.96 (0.76–1.21) S1, 0.98 (0.75–1.27) S2 | §2.5, §3.4 | bernstein2015_229/_337 | OK |
| Pietinen beef/pork/lamb 0.80 (0.50–1.20), 99 vs 35 g | §3.3 | estimates; Larsson Table I | OK |
| Fig 2 readings 1.77 (1.09–2.89) etc., Singh 1.49 (0.94–2.35), 140.3 px/unit, eight ticks | §2.4, §3.1 | R63 (the individual example triples are in the R63 reading only; not in a run file) | OK (traceable to R63) |
| "Eighteen studies (9 case-control and 9 cohort)" vs Table II n 17/8 | §3.6 | Norat l. 298, 248 | OK |
| Larsson took the 120 g portion from Norat and checked its references (pp. 2657, 2660) | §1 | 31_Larsson2006.txt l. 17–18, 212–214 | OK |

### Replacement texts (English)

1. §3.1, last bullet: "The choice of Singh value barely matters: with the plotted value instead of the tabulated one, the pool is 1.27 (1.12–1.46). The point estimate is unchanged and each confidence limit moves by 0.01 (our recomputation; R63)."
2. Abstract, Results, sentence 3: "Adding it gives 1.27 (1.11–1.45), matching the published estimate; it carries 43.5 % of the weight. With the calibrated high-resolution readings of Thun (0.91–1.38) and Singh (1.49) instead, the pool is 1.28 (1.12–1.46), within 0.01 of the published value."
3. §4.4, item 2, last two sentences: "Two independent readings, one from a 220-dpi rendering (K11) and one calibrated on a high-resolution scan supplied by the author (R63), agree to within 0.02 on each limit. With the high-resolution readings the pool is 1.27–1.28 (1.11–1.12 to 1.46). The reproduction therefore holds to within 0.01, but the exact two-decimal match depends on K11's reading and the tabulated Singh value. It remains a reproduction from a figure."
4. §2.5, "Follow-ups" lead-in: "We substituted later publications of the same cohorts, as coded in `scripts/reanalyses.py`:". §5, Database line: delete "`followups.csv`, ".
5. §2.7: "Every change is recorded in a correction log (`rettelogg.md`, entries R1–R80)."
6. §3.1, Transcription: "K11 found no transcription error in the 17 Norat rows it checked; the eighteenth row, the figure-read Thun entry, is K11's own reading (R59)."

---

## 2. larsson2006

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| Database includes `followups.csv` | §5 | file system | **FEIL** | No such file (see X10). |
| "Corrections are recorded in `rettelogg.md` (R1–R60)" | §2.6 | rettelogg | **FEIL** | R1–R80. |

### OVERDREVET

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "The highest-versus-lowest analysis is a model of reproducibility for its time." | §4.1 | – | **OVERDREVET** | A comparative claim; no comparison with other syntheses of the period is made. |

### Notes (not counted)
- L3 (naming): Table 1 calls Tiemersma "Dutch MONICA cohort". This should be "Dutch Monitoring Project" (X8).

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| 1.28 (1.15–1.42) 15 studies, 7,367 cases; DR 1.28 (1.18–1.39) 14 studies, Q 8.82, I² 0 | Abstract, §1, Table 2 | Larsson l. 17, 327, 357 | OK |
| Reproduction 1.28 (1.15–1.42), Q 4.86; all corrected runs, Q values, I² 0 % throughout | Table 2, §3.4 | all.json | OK |
| WCRF 2007 quote "are supported by a recently published meta-analysis of 15 prospective studies… 1.28 (95% CI 1.18–1.39) per 120 g/day" (p. 121) | §1 | 32_WCRF2007.txt l. 10289–10292 | OK (source begins "These data are supported") |
| SLR 2010 "1.28 (1.18-1.39, 14 studies) for each 120g/day increase in red (and processed) meat" | §1 | slr2010.txt l. 4941–4942 | OK |
| CUP 2011 "a published dose response meta-analysis of 7367 cases from 14 studies…(60)", p. 13 | §1 | report2011.txt l. 669–671, footer 13 | OK |
| Method quotes (inclusion, Bostick vs Sellers, greatest degree of control, fn 4 rule, fn 6, amplitude, Stata 9.0) | §2.2 | Larsson l. 10–60, 183–186, 200–221 | OK (two-column text; verbatim) |
| Table II fn 4: both definitions refs 5, 8, 18, 23; total only 9, 11, 19, 20; fresh only 4, 10, 13, 24 | Table 1, §3.3 | Larsson l. 290–292 | OK |
| 50 rows in synthesis_extraction | §2.3 | csv (21 + 15 + 14) | OK |
| Case column sums to 7,367 | §3.1 | own sum | OK |
| Weights 18.5/14.3/10.8/9.8 %; Chao 7.6 → 22.1 %; Bernstein 13.1/10.2 % | §3.2, §3.4 | larsson2006.md | OK |
| Subgroups rectum 1.56 (1.25–1.95), proximal 1.15 (0.91–1.47), distal 1.31 (0.62–2.79), colon 1.21 (1.05–1.40) Q 2.35, Singh left out | §3.2 | Larsson l. 240–243; K11 L24; R58 | OK |
| Sensitivity analyses 1.27 (1.14–1.42) refs 7, 20 (Kojima NA); 1.28 (1.15–1.42) nested | §3.3 | Larsson l. 230–231, 148 | OK |
| Fresh only 1.22 (1.08–1.37) n 9; total 1.24 (1.09–1.42) n 8 | §4.2 | Larsson l. 267–268 | OK |
| DR study refs 4, 5, 8–13, 18–20, 23, 24; Kato excluded "because red meat consumption could not be quantified" | §3.6 | Larsson l. 352–355 | OK |
| Chao colon 1.15 (0.90–1.46), rectum 1.71 (1.15–2.52); FE 1.28 (1.04–1.58) | Table 1, §3.3 | run file; estimates | OK |
| Bostick main dish includes ham | §3.3 | bostick1994.txt l. 449 | OK |
| Wei top category "5 or more per week"; main dish | §4.3 | wei2004.txt l. 496 | OK |
| LOO ranges (1.27–1.32; 1.20–1.24; 1.13–1.18; lower ≥ 1.03) | §3.5 | run file | OK |
| Effect of definition ≈ 0.06; follow-ups a further 0.07 | §4.3 | 1.28 → 1.22 → 1.15 | OK |

### Replacement texts

1. §5, Database: delete "`followups.csv`, ".
2. §2.6, Correction log: "Corrections are recorded in `rettelogg.md` (R1–R80)."
3. §4.1, lead sentence: "**The highest-versus-lowest analysis is unusually transparent.**"
4. Table 1, Tiemersma row: "Tiemersma 2002 (Dutch Monitoring Project, nested c–c) [20]".

---

## 3. wcrf2007

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "the 2006 SLR" (11 occurrences; also §2.1 "the 2006 SLR is not on disk") | throughout | slr2010.txt Tables 35–36 (l. 7865–7881) and K-files: "2005 SLR" | **FEIL** (label) | The only source that describes the review, the 2010 SLR, calls it the **"2005 SLR"**. The protocol (§7) says "2006 SLR". No source in the files supports 2006. |
| "Without them, the summary is about 1.10 and its interval includes 1" | Abstract, Conclusions | wcrf2007.md | **FEIL** | Removing the two superseded reports alone gives **1.14 (0.88–1.49)**. 1.10 (0.89–1.36) also needs the Chen strata combined. This is the same error K14 corrected in the overview article (R67–R78: "WCRF 2007 1,10 krever også Chen-sammenslåing"). |
| Bostick "four categories" (Table 1) vs "Bostick's three categories" (deviation table in run file, cited in §3.3 D5) | Table 1; run file | estimates bostick1994_005–008 | **FEIL** (minor, internal) | Four categories (<1, 1, 2–3, >3/week), three above the reference. Fix the run-file text in `reanalyses.py`. |

### OVERDREVET

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "No single row dominates the run, but two rows drive it." | §3.6 follow-up | run file: Bernstein 54.1 % | **OVERDREVET**/self-contradictory | One row carries more than half the weight. |
| "about six times narrower than a Greenland–Longnecker slope from Bostick's categories … can yield" | §3.3 D5 | derived_per100g: Bostick GL 0.94 (0.43–2.04) per 100 g | **OVERDREVET** (mild) | Log width 1.56 per 100 g (4.4×), or 1.87 per 120 g serving (5.3×), against 0.35. "Four to five times" is what the database supports. |

### IKKE VERIFISERBAR

| Claim | Location | Note |
|---|---|---|
| "The 2010 SLR's summary of the same earlier review says nine of eleven did, and one showed no association (Table 35)" | §3.8 | The Table 35 cell is in the red-meat table but reads "Nine studies reported an increased risk for the highest versus lowest **processed meat** comparison" (slr2010 l. 7868–7871; flagged in synthesis_extraction_NOTES l. 76). Whether "nine of eleven" refers to red meat cannot be determined; say so. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| "A substantial amount of data … Red meat is a convincing cause of colorectal cancer." (p121) | §1 | 32_WCRF2007.txt | OK |
| 1.43 (1.05–1.94) "per times/week", 7 studies; 1.29 (1.04–1.60) text; 1.29 (1.05–1.59) figure; Larsson "1.43 (1.05–0.95)" | §1, §3 | WCRF txt l. 10195ff; R66 | OK |
| Box 3.3 quote "Unless indicated otherwise, random effects models … were used to generate the forest plots presented in this Report" (p53) | §2.3 | WCRF txt l. 5327–5336 | OK |
| SLR 2010: "if not available colon cancer", "run separated by exposure measurement unit (gram or time/day)", "1.43 (95% CI = 1.05-1.94, seven studies) per time/day increase", Willett/Giovannucci "replaced by the most recent publication by Wei", Chen under red and processed meat, EPIC observed chosen (p156) | §2.3, §3.3, §3.4 | slr2010.txt l. 7797–7808 | OK |
| Table 36: 1.29 (0.94–1.78) per 100 g, I² 32 %; I² 58 % p 0.01; sex strata men 1.79 (0.98–3.27), women 1.27 (0.74–2.19) | §3.3, §3.4 | slr2010 l. 7886–7905 | OK |
| Reproduction 1.43 (1.05–1.94) I² 58 %, FE 1.15 (1.00–1.32), Q 19.00, τ² 0.092 | §3.1 | all.json | OK |
| Weights 24.4/20.1/17.3/13.7 %; 47.1/30.2 %; 56.1/28.0 %; 54.1/32.3 %; Singh 1.5 % | §3 | wcrf2007.md | OK |
| All corrected runs, LOO tables, per-100 g run 1.29 (1.04–1.60)/(1.05–1.59) I² 6 %, weights 45.3/36.1/18.6 % | Table 2, §3.6–3.7 | run file | OK |
| English 1.03 (0.98–1.08) per time/week; 1.03⁷ = 1.23 (0.87–1.71) | Table 1, §3.4 | english2004_005; own arithmetic | OK |
| Willett ≥1/day 2.49 (1.24–5.03) Table 3; Giovannucci ≥5/wk 3.57 (1.58–8.06) Table 2; Bostick >3 vs <1 1.21 (0.75–1.96) | Table 1 | estimates | OK |
| Willett 150, Giovannucci 205 colon cases; Wei 1,139 + 339 | Table 1, D3 | studies.csv; wei2004.txt l. 164 | OK |
| Wei 1.12 (0.99–1.27) "one additional serving per week" (Table III, p441) vs Methods servings per day (p435); disputed, R15 | §2.5, §4.3 | wei2004.txt l. 637, 162–164; R15 | OK |
| Bernstein S5 1.00 (0.89–1.11) → 1.00 (0.87–1.13) | §2.5 | bernstein2015_488; arithmetic | OK |
| 24 numeric rows matched by K11 | §2.7 | K11_rapport l. 83 (25 rows with RR; Larsson 4.3.3 lacks the upper limit) | OK |
| Fig 4.3.1: "Sixteen cohort studies" with refs 8–24; Philips 1975 2.70 (1.60–4.55); "Chen 2003" 1.48 (0.85–2.59) = Jiashan pork 1.481 (0.846–2.593) in Chan S2; "Wei 2003 Men/Women"; "Slattery 2001 Women"; Phillips called case-control (Table 35 fn) | §3.8 | WCRF txt l. 10195–10221; chan2011_supp_s002 l. 336–344; slr2010 l. 7881, 8244 | OK |
| Chan 2011 abstract opens with the 2007 "convincing" judgement | §1 | chan2011.txt l. 11–12 | OK |

### Replacement texts

1. First occurrence (§1): "Because the underlying earlier SLR (labelled the "2005 SLR" in SLR 2010 Tables 35–36) is not available in our files, …". Then "the 2005 SLR" in all ten other places. Note in §2.1: "(the protocol calls it the 2006 SLR)".
2. Abstract, Conclusions, sentence 2: "Without them the summary is 1.14 (0.88–1.49), and 1.10 (0.89–1.36) when the Chen strata are also combined; both intervals include 1. The corrected runs lean heavily on one row (Bostick 1994) whose printed interval is itself doubtful."
3. Table 1, Bostick row: "Beef, pork or lamb as main dish, four categories (three above the reference); >3 vs <1 serving/week 1.21 (0.75–1.96) (Table 3, p43)." In `reanalyses.py` / run file deviation table: "from Bostick's four categories (212 cases)".
4. §3.6, follow-up, last sentence: "The run is driven by these two rows, and above all by the Bernstein row."
5. §3.3 D5, bullet 1: "The printed CI, 0.80–1.14, spans 0.35 on the log scale. That is four to five times narrower than a Greenland–Longnecker slope from Bostick's categories can yield (212 cases; per 100 g/day 0.94 (0.43–2.04), log width 1.56; `derived_per100g.csv`)."
6. §3.8, "Claim of uniformity": "… (p120). The 2010 SLR's summary of the same earlier review (Table 35, p157) says that nine of eleven studies showed increased risk and one no association, but that cell, in the red-meat table, refers to the "processed meat comparison", so the count may not concern red meat."

---

## 4. cup2010

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "All published summaries reproduce: … colon … 1.17 (1.02–1.34) in Chan; rectum 1.18 (0.98–1.43)" | Abstract, Results | Chan T1; SLR p155; run file | **FEIL** | These are our reproduced values. The published values are 1.17 (1.02–**1.33**) and 1.18 (0.98–**1.42**). An abstract that says "published summaries" should give the published numbers. |
| "K11 confirmed all 69 numeric rows" | §2.7 | K11_rapport l. 20 ("cup2010 76"); csv: all 76 rows have RR and CI | **FEIL** | 69 is not traceable. K11 checked all 76. |

### OVERDREVET

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "The run is therefore driven by the two largest cohorts, but no single row dominates it." | §3.6, dominance check | run file: EPIC 53.2 % | **OVERDREVET** | One row carries more than half the weight. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| 1.17 (1.05–1.31) 8 studies; colon 1.12 (0.97–1.29) 9; rectum 1.18 (0.98–1.42) 7; I² 0 | §1 | slr2010 l. 7781, 7917, 8523; rm2010 | OK |
| Chan abstract quote "RR for 100 g/day increase = 1.17, 95% CI = 1.05–1.31" | §1 | 33_Chan2011.txt | OK |
| CUP 2011 quote "The evidence that red meat, processed meat, … are causes of colorectal cancer is convincing." | §1 | report2011.txt | OK |
| IARC "may have increased the measurement error" (pp143–144) | §1 | IARC l. 7311–7312 | OK |
| SLR method quotes (common scale, 120 g; 100 g increment; energy conversion) p155 | §2.3 | slr2010.txt | OK |
| Chan quote on Larsson proximal/distal (ref 34) | §2.3 | Chan txt | OK |
| Sensitivity 1.21 (1.06–1.37), 1.12 (0.97–1.31), 1.20 (0.99–1.47); "did not provide details on the definition of red meat" | §2.3 | slr2010 l. 7781 | OK |
| English mutually adjusted 1.02 (0.98–1.08), "modeled together", → 1.12 (0.89–1.57) | §2.5, D4 | english2004.txt; ×5.833 arithmetic | OK |
| Papadimitriou per SD 36.2 g → 1.06 (0.95–1.14); Ollberding 0.97 (0.83–1.13) | §2.5 | update2026.json; run file | OK |
| Table 1 values, weights and Test 2 re-derivations | §3.2 | run file; test2_report | OK |
| Järvinen cut-points men <94/94–141/142–206/>206 vs <101…>220; ratios 91–94 % | D3 | jarvinen2001.txt Table 3 (men 93–94 %, women 91–92 %) | OK |
| Nöthlings 1,009 cases / 1,522 controls; blood from incident cases; SLR search end Dec 2009 (+ to May 2010); Chan to March 2011 | D2 | nothlings2009.txt l. 18, 152–153; slr2010 l. 981, 4833–4838; chan2011 l. 16 | OK |
| Case sums 4,313/4,314; 3,126/3,172 (gap 46 = 109 − 63); 1,415/1,477 | §3.3 | synthesis_extraction_NOTES §4 | OK |
| Chan S2 quote on Larsson 2005; Chan colon 10 rows, 3,561 cases, 389 = Larsson proximal + distal (234 + 155) | §3.4 | chan2011_supp_s002; R14 | OK |
| All runs, LOO, weights (44.4 → 8.9 %; recomputed weights within 0.39 pp) | §3.5–3.6 | cup2010.md | OK |
| Proximal/distal HvL unresolved; Chao S1 red and processed | §3.4 | K11 l. 47; R64 | OK |

### Replacement texts

1. Abstract, Results, sentence 1: "All published summaries reproduce, two of them within 0.01 in one limit: colorectal 1.17 (1.05–1.31); colon 1.12 (0.97–1.29) in the SLR and 1.17 (1.02–1.33) in Chan (reproduced 1.17 (1.02–1.34)); rectum 1.18 (0.98–1.42) (reproduced 1.18 (0.98–1.43)); highest versus lowest 1.10 (1.00–1.21)."
2. §2.7, Synthesis extraction: "The CUP 2010/Chan 2011 extraction (76 rows) was checked by a separate controller (K11, `kontroll/K11/K11_rapport.md`) against the printed pages and the rendered Chan Fig 4. K11 confirmed the RR, CI and printed weight in all 76 rows. It flagged citation and labelling issues, which were corrected (correction log R57–R58)."
3. §3.6, Dominance check: "The two substituted rows carry 53.2 % (EPIC) and 25.1 % (MEC) of the weight, so the run is driven by these two cohorts, EPIC above all. Without the EPIC row the result is 1.04 (0.90–1.20); without the MEC row it is 1.07 (0.97–1.17)."

---

## 5. cup2016

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| "of 10,247 rows, 10,217 were verified, 18 corrected and 12 left disputed … (`database-samledom.md`, §2" | §2.7 | estimates.csv: 10,383 / 10,353 verified / 18 corrected / 12 disputed. The cited samledom §2 says 6,055 of 6,077. | **FEIL** | Outdated, and the cited source gives yet another figure. |
| "Dutch MONICA (Tiemersma 2002)" | Tables 1A, 1B | studies.csv; Tiemersma | **FEIL** (label) | Dutch Monitoring Project, not WHO MONICA (X8). |

### OVERDREVET

| Claim | Location | Verdict | Note |
|---|---|---|---|
| "Few diet–cancer syntheses meet that standard." | §4.1 | **OVERDREVET** | A comparative claim with no data behind it. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| "probably a cause"; "beef, pork, lamb, and goat from domesticated animals" (p. 8 fn 10); "showed no significant association (RR 1.12 (95% CI 1.00–1.25))" (p. 34); Panel quote (p. 36); search to 30 April 2015 (p. 14); "largely unchanged" | Abstract, §1, §2.2 | cup/report.txt l. 351, 571, 2126, 2430; page footers checked | OK |
| Norat listed as Principal Investigator (p. 88) | §4.1 | report.txt l. 6958–6960 (footer 88) | OK, factual and hedged |
| SLR p. 335 exposure rule (spelling as printed); "Weights are from random effects analysis"; Table 79 phrases | §2.2 | rm_334_352.txt (Table 79 phrases split over the column layout) | OK |
| Sensitivity quotes p. 334/335: 1.09 (0.96–1.25)/1.19 (1.06–1.34); colon w/o Larsson 1.09 (1.03–1.16); rectum w/o EPIC 1.05 (0.94–1.17); "No association … was reported in all studies" | §3.1, §4.3 | rm_334_352.txt l. 26–27, 43, 51, 44–46 | OK |
| Our LOO 1.18 (1.05–1.34) (SMC combined) and 1.08 (0.87–1.33) (rectum w/o EPIC); FE intervals 1.08–1.37 and 0.96–1.34; the logic that no subset can be narrower than the full FE interval | §3.1 | cup2016.md l. 140, 226; all.json | OK |
| All Table 2 runs (k, RR, I², FE), weights MEC 35.0/35.10, EPIC 24.9/24.84, I² 23/23.6, 12/11.7 | Table 2, §3.1 | all.json | OK |
| 21/26 within ±0.03, 18 with limits within 0.02; Wei/English exceptions; Table 1A values | §3.2 | test2_report.md; cup_extraction | OK |
| Takachi men total-meat medians 20–117 vs red-meat 15–102 (Table 2); men colon 481 cases; corrected 1.39 (1.08–1.81), 0.95 (0.63–1.44) | §3.2, Table 3 | takachi2011.txt l. 285–304; run file | OK |
| Larsson medians 1.5/2.5/4.0/5.5 × 120/7 | §3.2 | database-samledom l. 25; dose_conversions_cup | OK |
| MEC "Total red meat" 1.02 (0.91–1.16) vs unprocessed 0.98 (0.87–1.10) | §3.3 | estimates; IARC row 20 | OK |
| Case totals 6,662 (with Järvinen 63) / 6,708; colon 4,081, Takachi 259 vs 481 → 222; English 452 vs 451 | §3.3 | cup_extraction sums | OK |
| Calibrated EPIC 1.49/1.36/1.75; English adjusted per 100 g 1.12/1.00/1.48 | §3.3 | run file | OK |
| Corrected and follow-up runs: EPIC 58.9 %; Bernstein 0.99 (0.87–1.12) 42.7 %; Islam 1.05 (0.88–1.26) 29.1 %; all LOO ranges | §3.4 | cup2016.md l. 69, 193, 210 | OK |
| Multiverse 768/123 (16 %), median 1.08 (1.03–1.16), EPIC 43/2/3 %, Järvinen 25/7, Tiemersma 21/11, Larsson 9/11/13; colon 384/312 (81 %), 1.22 (1.11–1.28), Takachi 100 %, EPIC 96/67, Islam 1.13/44 % | §3.5 | multiverse_report.md | OK (double count fixed) |
| Unreproduced inputs 8.0 %/4.5 %/4.35 % of weight | §4.4 | Lee 5.17 + Tiemersma 2.82; Lee 3.49 + Singh 1.05; Lee 4.35 | OK |
| English Fig 138 2.30 (1.20–4.20) | §4.3 | rm_354_375 l. 348 | OK |
| K11 did not cover CUP 2016; independent check 20.10.2026 | §2.7 | kontroll_rapport.md l. 3 | OK |

### Replacement texts

1. §2.7, sentence 2: "Every row of the estimates database was re-checked against full text or supplement by independent agent instances that had not seen the extraction. Of 10,383 rows, 10,353 were verified, 18 corrected and 12 left disputed because the papers' own numbers are internally impossible. No numerical error was found in the 13 CUP studies (`database/estimates.csv`, status 30 October 2026; corrections in `kontroll/rettelogg.md`)."
2. Tables 1A and 1B: replace "Dutch MONICA" with "Dutch Monitoring Project".
3. §4.1, paragraph 1: delete the sentence "Few diet–cancer syntheses meet that standard."

---

## 6. update2026

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| Case counts in Table 2: NLCS 2,290; UK Biobank 2,609; DANSDA 644; Israel 1,114; JPHC + Takayama 2,622; Finnish cohorts 1,259; Korea 119; Alberta 232 | Table 2 | Primary texts / studies.csv / estimate rows. Gilsing 2015 l. 349: **437**. Knuppel 2020 category rows sum to **3,201** (studies.csv 3,201); **2,609 is Bradbury 2020's count** (bradbury2020.txt l. 36). Mejborn 2021 l. 190: **127**. Dankner 2024: **813** in the analysis (839 overall). Islam 2019 quantity pool (JPHC-5y + Takayama): **3,550** (l. 61; multiverse_report says 2,749). Tammi 2024: **1,124**. Wie 2014 Table: colorectal n **53**. Al Rajabi 2021: **257** in Model 2 (116 + 141; 265 overall). | **FEIL** | The values are hard-coded in `scripts/update2026.py` (l. 60–96) and not read from the database. The other 12 cohorts are correct. |
| "20 cohorts and 46,450 cases" | Abstract; Table 2 total; §4.1 | as above | **FEIL** | With the counts above, the total is **45,123**, or 45,157 with Dankner 839 and Al Rajabi 265. Say "about 45,100". |
| Size subgroups: ≥ 1,000 cases k 10, 1.09 (1.03–1.16) I² 38 %; < 1,000 k 10, 1.15 (0.99–1.33) I² 0 % | Table 4; §4.2 item 4 | own DL on the json rows, with corrected counts | **FEIL** | NLCS (437) and Israel (813) move to < 1,000. Then ≥ 1,000: k 8, **1.10 (1.03–1.18), I² 49 %**; < 1,000: k 12, **1.11 (0.99–1.23), I² 0 %**. The small-study argument in §4.2 item 4 largely disappears. Rerun the script after fixing the counts. |
| Database: "10,247 estimates from 38 publications in 30 cohorts … 6,055 rows were verified; 18 corrected; 4 disputed" (§2.6) vs "10,217 of 10,247 rows verified; 18 corrected, 12 disputed" (§4.4) | §2.6, §4.4 | estimates.csv: 10,383 rows; studies.csv: 53 publications; 10,353/18/12 | **FEIL** | The two sections contradict each other, and both are outdated. |
| "It recomputed the 2026 main pool and all 45 re-analysis runs … (`kontroll/meta_check_reanalyser.txt`)" | §2.6 | all.json has 46 runs. K14 l. 45: 45 of 46 checked (WCRF Fig 4.3.3 not). The cited file (regenerated 13:19) now holds the synthesis_extraction meta-check only, with no run names and no 1.0967. The main pool 1.0967 (1.0447–1.1513) is confirmed in K16a/RAPPORT.md l. 41. | **FEIL** | Wrong count and wrong file reference. |
| "Netherlands MORGEN (Tiemersma 2002)" | Table 2, §2.3, §3.6, §3.7 | studies.csv; IARC l. 5684 | **FEIL** (label) | Dutch Monitoring Project 1987–1991, not MORGEN (X8). |
| NLCS route "case-cohort" | Table 2 | gilsing2015.txt l. 331: "The full-cohort approach is used for analyses of NLCS-MIC" | **FEIL** (minor) | |
| "six entered unchanged (ATBC, Netherlands MORGEN, MCCS, SMC, MEC, SWHS)" | §2.3 | update2026.json vs cup2016 | **FEIL** (misleading) | MEC enters with the unprocessed row, 0.97 (CUP: 1.01). The Monitoring Project enters with our GL value, 1.92 (CUP: 1.69). Only the publication is unchanged. |

### OVERDREVET

| Claim | Location | Verdict | Note |
|---|---|---|---|
| "The update shows that this fragility was a feature of the 2016 evidence base and not of the association reported." | §4.1 | **OVERDREVET** | "Not of the association" is a claim about the true association, which the data cannot support; the paper itself says stability is not validity (next paragraph). |
| "Adding the post-2015 cohorts removes the dependence of statistical significance on any one extraction choice." | §3.6 | **OVERDREVET** (mild) | True only for the choices in the multiverse. |

### IKKE VERIFISERBAR

| Claim | Location | Note |
|---|---|---|
| "no estimator, single omission or combination of extraction alternatives moves the lower limit below 1.03" | §4.1 | Estimators and LOO: OK (lowest 1.03). Multiverse: the json stores only the count of significant combinations (1,152/1,152) and the point-estimate range, not the lower limits. |
| Negative controls 0.89 (0.74–1.08) and 0.76 (0.65–0.90) | Abstract, §3.8 | Consistent with update2026.md. K16a could not reproduce poultry (0.90) because the per-study rows are not stored in the json (R80 text). |
| Linear slope 1.24 (1.06–1.44) | Abstract, §3.7 | Consistent with the run file. K16a got 1.23 (1.05–1.44); the inputs are not in the json. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| Main 1.10 (1.04–1.15), k 20, I² 13 %, τ² 0.0014, Q 21.8; REML τ² 0.0016; HKSJ 1.10 (1.04–1.16); FE 1.10 (1.06–1.14); Egger 0.33, p 0.40 | §3.3, Table 3 | update2026.json main | OK |
| All Table 2 RRs and weights; "55 %" (25.6 + 16.3 + 13.4) | Table 2, §3.2 | json set | OK |
| LOO range 1.09–1.11, lowest 1.09 (1.03–1.15) without NIH-AARP, highest 1.11 (1.06–1.16) without NHS/HPFS; MEC/EPIC raise to 1.11 | Table 3, §3.4 | json loo | OK |
| Sensitivity runs (MWS, NIH-AARP, Tiemersma, two largest, Vulcan, Wang) | Table 3 | json sensitivity | OK |
| Subgroups other than size | Table 4 | update2026.md | OK |
| Multiverse 1,152, median 1.11 (1.09–1.13), 1,152/1,152 | §3.6 | json multiverse | OK |
| Spline sets, knots 9/38/84, Wald 0.38 p 0.54, Q 20.0/22 df, curve values | §3.7 | update2026.md | OK |
| PRISMA counts (8,119; 9,242; 520/110/630; 16,731; 15,703; 11,541; 155 and 4 rescued; 1,028; 478/261/256/6/5; 4 extra excluded; 22 = 18 + 4; 16/18 sensitivity; Hoang/Jin via (b) only) | §2.2, §3.1, Table 1 | SEARCH_LOG.md l. 17–134 | OK |
| Vulcan 728 cases, 1.02 (0.75–1.38); Beslay 1.42 (1.03–1.94); Wang 1.08 (0.98–1.19) / 1.01 (0.95–1.07), Suppl Table 4, 51.4 g/day, ^1.944 | §3.1 | studies.csv; wang2020_075/_086; json | OK. The Wang strata are age strata, as the article says; the json prov label "sex strata combined" is wrong (script label only). |
| CUP 2016 1.12 (1.00–1.25), I² 24 %, 8 studies (Vieira p. 7); fish 0.89 (0.80–0.99) 11 studies; poultry 0.81 (0.53–1.25) 7 studies | §1, §4.3, §4.5 | vieira2017.txt l. 239, 356, 419 | OK |
| Han 2019 1.00 (0.92–1.09), 5 studies, 322,502 participants, low certainty | §4.5 | han2019.txt l. 257 | OK |
| Ungvari 60 studies; HR 1.15 (1.10–1.21) (abstract); 1.18 (1.13–1.23), 71 cohorts, I² 29 %; "generally referred to beef, pork, lamb, mutton, and veal" | §4.5 | ungvari2025.txt l. 28–34, 209–222, 153 | OK |
| fish ≈ 3× red meat on log scale (0.274/0.095) | §4.3 | arithmetic | OK |
| 1.10 per 100 g ≈ 5 % at 50 g | §4.1 | 1.10^0.5 = 1.049 | OK |
| CUP 2016 multiverse 16 % of 768 | §3.6 | multiverse_report | OK |

### Replacement texts

1. Table 2, Cases column: NLCS **437**; UK Biobank (Knuppel 2020) **3,201**; DANSDA **127**; Israel **813**; JPHC + Takayama **3,550**; Finnish cohorts **1,124**; Korea **53**; Alberta's Tomorrow **257**. Total "**45,123**". Make the same change in `scripts/update2026.py`, rerun, and rebuild the HTML. Footnote: "Cases are those in the analysis that produced the estimate used (Dankner 2024: 813 of 839; Al Rajabi 2021: 257 in Model 2 of 265)."
2. Abstract, Results: "The main pool had 20 cohorts and about 45,100 cases." §4.1, first sentence: "With 20 cohorts and about 45,100 cases, the CUP 2016 method gives 1.10 (1.04–1.15) per 100 g/day."
3. Table 4, size rows (provisional, from my recomputation of the json rows; confirm with the script): "≥ 1,000 cases | 8 | 1.10 (1.03–1.18) | 49 %" and "< 1,000 cases | 12 | 1.11 (0.99–1.23) | 0 %".
4. §4.2, item 4: "**Small-study composition.** Within the main pool, the cohorts with fewer than 1,000 cases give 1.11 (0.99–1.23) and the larger ones 1.10 (1.03–1.18), so study size explains little of the gap. Egger's test in the main pool shows no asymmetry."
5. §2.6, paragraph 1: "It holds 10,383 estimates from 53 publications. Every row was checked against the full text or supplement by instances that had not seen the extraction: 10,353 rows were verified; 18 were corrected; 12 are disputed because the source paper's own numbers are internally impossible."
6. §4.4, AI-agent paragraph: "The independent row-level control (10,353 of 10,383 rows verified; 18 corrected, 12 disputed) and the correction log are the safeguards …"
7. §2.6, paragraph 2: "… with an independent second script (`kontroll/meta_check.py`, written by the K11 controller without sight of `pool.py`). It recomputed 45 of the 46 re-analysis runs to four decimals (K14). A further independent recomputation (K16a, `kontroll/K16a/RAPPORT.md`) reproduced the 2026 main pool, subgroups, sensitivity runs, leave-one-out, Egger, HKSJ, REML, the multiverse and the WCRF 2007 per-100 g run. The main pool is 1.0967 (1.0447–1.1513) in both."
8. Table 2 and §2.3, §3.6, §3.7: replace "Netherlands MORGEN" / "MORGEN" with "Dutch Monitoring Project".
9. Table 2, NLCS route: "per 50 g^2; NLCS Meat Investigation Cohort, full-cohort analysis".
10. §2.3, last paragraph: "For the cohorts CUP 2016 pooled for colorectal cancer, six entered through the same publication (ATBC, Dutch Monitoring Project, MCCS, SMC, MEC, SWHS). For MEC we used the paper's unprocessed row, and for the Monitoring Project our own Greenland–Longnecker value, rather than CUP's rows."
11. §4.1, paragraph 2, sentence 3: "The update suggests that this fragility reflected the small 2016 evidence base: with the post-2015 cohorts the pooled estimate is stable at about 1.10."
12. §3.6, last sentence: "Adding the post-2015 cohorts removes the dependence of statistical significance on any one of the extraction choices tested."
13. §4.1, paragraph 1, last sentence: "In the update, no estimator and no single omission moves the lower limit below 1.03, and every multiverse combination keeps it above 1."
14. §3.6, first sentence: "All 1,152 combinations of this update's multiverse (not to be confused with the 1,152 = 768 + 384 combinations of the CUP 2016 multiverse) had a lower confidence limit above 1 …"

---

## 7. evidence_1997_iarc114

### FEIL

| Claim | Location | Source | Verdict | Note |
|---|---|---|---|---|
| Thun 1992 "not in database / cannot check"; "Four of the seven cohorts are in the database. The report prints numbers for only two of them … and both match"; "Three WCRF cohorts (Phillips 1985, Thun 1992, Knekt 1994)" not in db; abstract "Both WCRF 1997 numbers we could check match" | Abstract; Table 1; §3.1; §3.5; §4.4 | R65/K13: thun1992 40 rows in estimates.csv; Thun Table 4 "fatty meats" Q5 men 1.21, women 1.05, crude, no CI | **FEIL** | Out of date. Thun is in the database, and the WCRF values 1.21/1.05 match it. So five of the seven cohorts are in the database, and three have printed numbers, all matching. The evidens files (wcrf1997_cohorts.*) are equally out of date. |
| "database of 10,247 estimates"; "10,247 estimates from 38 publications in 30 cohorts … 6,055 rows verified, 18 corrected and 4 disputed" | Abstract Methods; §2.1 | estimates.csv; studies.csv | **FEIL** | 10,383 / 53 publications / 10,353 / 18 / 12. |
| "If 'positive' means a highest-vs-lowest point estimate above 1, exactly seven cohorts qualify" | §3.2 | Table 2 rows 13, 15 (iarc114_cohorts) | **FEIL** | By the stated rule SCHS (Butler 2008, 1.01) and Takayama men (1.03) also qualify. The seven only come out with a threshold of about 1.05. |
| Tiemersma cohort "MORGEN" | §3.2 list; Table 2 row 6 | IARC l. 5684 ("Monitoring Project on Cardiovascular…") | **FEIL** (label) | X8. |
| "IARC Monograph 114 (2015)" | Abstract | reference [3]: 2018 | **FEIL** (minor) | Evaluated 2015, published 2018. |

### OVERDREVET

| Claim | Location | Verdict | Note |
|---|---|---|---|
| "For a 16-page hand-built table this is a good record." | §4.1 | **OVERDREVET** (mild) | How IARC built the table ("hand-built") is not documented. |

### OK (selection)

| Claim | Location | Source | Verdict |
|---|---|---|---|
| IARC quotes: "limited evidence" (§6.1 l. 24324–24327); Group 2A (l. 24346–24347); "the substantial epidemiological data … strong mechanistic evidence" (§6.4); inclusion rule (l. 5660–5664); "Fourteen cohort studies and two cohort consortia…" (l. 5660–5663, right column); §5.2.1 "seven studies…" (l. 23964–23971); "chance, bias, or confounding could not be ruled out…" (l. 23999–24002); "a meta-analysis including data from 10 cohort studies" (l. 23978); "8 studies" (l. 7252); Alexander "1.02 (95% CI, 1.00–1.14; 10 studies) for 1 serving/day increase" (l. 7266–7267) | §1, §3 | 36_IARC114.txt (two columns; line ranges correct within the stated ±5) | OK |
| Alexander 2015 prints 1.02 (1.00–1.04) per serving per week; I² 88.62 for 70 g | Table 3 | alexander2015_jacn.txt l. 722–725 | OK |
| Singh: Table 4 subset 1.40 (0.87–2.25), 1.90 (1.16–3.11); full cohort ≥1/week 1.41 (0.90–2.21); abstract quote | Table 3, §4.1 | singh1998.txt l. 19–21, 392–394, 472 | OK (the paper prints "≥1/week" and ">1 time/week" in different places; quoted verbatim) |
| Bouvard: "seven" of 15 case-control; "ten cohort studies"; "residual confounding … difficult to exclude" | §3.5, §4.1, §4.2 | bouvard2015.txt l. 93–106, 140–141 | OK |
| 178 triples: 166 match / 11 not in db / 1 NO_MATCH; 23 tabulated + 14 text-only rows | §3.3, Data availability | iarc114_table_values_check.csv; iarc114_cohorts.csv | OK |
| Chan 2011 numbers in IARC all match Chan Table 1 | §3.4 | IARC l. 7250–7254; run files | OK |
| Bernstein rectal rows labelled "Processed red meat (1 serving/day)" with unprocessed values 1.05, 1.14 | Table 3 | IARC l. 8040–8044; bernstein2015_132 | OK |
| DCH twice (Sørensen 2008, Andersen 2009) | Table 3 | IARC l. 7846ff | OK |
| Wei 2004 colon ≥5/week vs <1/month 1.43 (1.00–2.05); per unit 1.12 (0.99–1.27) | Table 3 | estimates | OK (the 1.12 row is disputed for its unit, R15; consider adding "(unit disputed)") |
| WCRF 1997 quotes (p. 246, p. 243), tally 16/8/8/0, 26 case-control studies (23 + 3), Willett 1.77 (1.09–2.88) and 2.49 (1.24–5.03) (16 cases), Giovannucci 1.71 (1.15–2.55), Goldbohm 0.84 (0.51–1.37) and 0.96 (0.87–1.06), Bostick 1.04 / 1.21 | §3.1, Table 1 | 25_WCRF1997.md; wcrf1997_cohorts.md; estimates | OK (from snippets, as stated) |
| "which IARC said only of processed meat" | §4.2 | IARC §6 (processed meat Group 1, sufficient evidence) | OK |

### Replacement texts

1. Table 1, Thun row: "| Thun 1992 | CPS-II mortality | "no difference"; men 1.21, women 1.05 | "fatty meats" Q5: men 1.21, women 1.05 (Table 4, p. 1495); unadjusted, no CI | match. Fatal colon cancer; crude estimates; the report's own 1.21 for men is called "no difference" |".
2. §3.1, after Table 1: "Five of the seven cohorts are in the database. The report prints numbers for three of them (Willett 1990, Giovannucci 1994 and Thun 1992), and all match exactly."
3. Abstract, Results, last sentence: "Every WCRF 1997 number we could check (Willett 1990, Giovannucci 1994, Thun 1992) matches."
4. §3.5, bullet 2: "**Two WCRF cohorts** (Phillips 1985, Knekt 1994) and **five IARC Table 2.2.1 rows** …". §4.4, bullet 2: "Five IARC Table 2.2.1 studies and two WCRF cohorts are outside the database."
5. Abstract, Methods: "… matched to a double-checked database of 10,383 estimates from the primary papers …". §2.1, Primary papers: "The database holds 10,383 estimates from 53 publications. Every row was checked against the full text or supplement by an extraction instance that had not built it: 10,353 rows verified, 18 corrected and 12 disputed, with all corrections logged [7]."
6. §3.2, sentence before the list: "If "positive" means a highest-vs-lowest point estimate clearly above 1 (here ≥ 1.1), exactly seven cohorts qualify:". Add after the list: "A strict "> 1" rule would also admit SCHS (1.01) and Takayama men (1.03)."
7. §3.2 cohort list and Table 2 row 6: replace "MORGEN" with "Dutch Monitoring Project (MP)".
8. Abstract, Background: "IARC Monograph 114 (evaluated 2015, published 2018) found …".
9. §4.1: "For a 16-page table this is a good record."

---

## Other observations (outside the seven md files; for the main session)

- `kontroll/meta_check_reanalyser.txt` was regenerated at 13:19 and now contains the synthesis_extraction meta-check (PART A/B). It no longer contains run-name recomputations of `all.json`. Any article citing it for the reanalysis recomputation (update2026 §2.6) points to the wrong content.
- `kontroll/evidens/wcrf1997_cohorts.{csv,md}` and `NOTES.md` still list Thun 1992 as `not_in_db`.
- `database-samledom.md` still holds the pre-fix colon multiverse (85 %/54 %) and the double-counted "Oba → Wada" colon row 1,14 (1,02–1,27), plus 6,055/6,077 verification counts.
- The `update2026.json` provenance label for the Wang sensitivity row says "sex strata combined"; the rows are age-at-onset strata (wang2020_075 onset_lt60, _086 onset_ge60).
- HTML (`html/artikler/update2026.html`, `html/index.html`, `html/oversikt.html`) still says 46 233; it should be regenerated after the case-count fix (P1).