FIGSHARE:26877211

As of May 2024, the final dataframe consists of 6552 rows. The RDS file contains a structured data frame to reflect the fields described in Table 1 of the manuscript

Data Records, paragraphs 1 and 2

jecfa.rds has 6552 rows and 17 columns, one for each field of Table 1, named as in R/process_df.R in lower case with "." replaced by "_". JECFA.csv holds the same 6552 rows and 17 columns, with no differing cell. The MD5 of both files read matches the archive's computed_md5. The columns R/add_metadata.R adds are absent, so the deposit is the output of process_df.R. The "May 2024" date cannot be checked from the files, which carry no scrape date.

We validated the effectiveness of this method by identifying 6552 entries, which were then cleaned to remove duplicates of (E)-N-[2-(1,3-BENZODIOXOL-5-YL)ETHYL]-3- (3,4-DIMETHOXYPHENYL)PROP-2-ENAMIDE ... and test pages, resulting in 6549

Methods, Web-scraping, paragraph 2

D1 (distinct URLs) gives 6552, 3 from 6549 and outside the tolerance of 2, as no URL is repeated. D2 (distinct JECFA names less page 5883) and D3 (rows less page 5883 and less the second row of page 6316's chemical) give 6550, within tolerance. Page 6316's chemical appears twice, at pages 6316 (name only) and 6319 (full record); page 5883 is named TEST and carries nothing else. The paper's 6549 is reached, for example, by removing both rows of that chemical with the test page (the one variant tested).

How the test differed from the paper’s method

• The criterion's readings D1 to D3 are all run; measured is the first that matches (D2). D1 does not match.

Not quoted: Numbers from Table 1, column Availability (%), titled "Common content of most JECFA pages"; no sentence of the text states them.

Over all 6552 rows (U1 and U2 are the same, as no URL repeats), 16 of 17 shares are within tolerance with the deposited fas column (reading A), FAS being 81.38% against 93% (lower bound 83.7%), and 17 of 17 with fas after the step of R/add_metadata.R (reading B, 86.74%). Over the 6550 rows left after cleaning, no share moves by more than 0.05 points. Rounded to whole numbers 12 of 17 equal Table 1 (Chemical names 83, Synonyms 78, Report and its link 98 differ by one point, FAS by 12 under A). Counting the 752 empty fas strings as present gives 92.86%, which suggests Table 1 counted every non-NA value; the claims file's reading E counts them as absent. The outcome rests on 16 of 17 under A, or 17 under B, and does not depend on how the empty strings are counted.

How the test differed from the paper’s method

• The criterion names the FAS column of both process_df.R and add_metadata.R; the deposit holds the first. Both are run (A and B), and both give reproduced.

For the evaluation year, from 1958 to 2000 there were 2,356 evaluations, representing 37% of the total. Between 2000 and 2010, the number of evaluations increased significantly to 3,099 (48% of the total)

Data Records, paragraph 3; Table 2, Evaluation year

The four bins and the unknowns equal Table 2 exactly; shares are over the 6434 rows with a year. Counts are compared with the claims file's tolerances (236, 310, 78, 20 and 12). All years lie between 1958 and 2023. The text's 2,356 for the first bin differs by one from Table 2 and from the data.

In terms of functional class, the majority of the evaluated substances were flavoring agents, with 4,635 substances comprising 72% of the total

Data Records, paragraph 3; Table 2, Functional Class

The counts equal Table 2 exactly; shares are over the 6442 rows with a class. Every functional_class string starts with one of the four class names (the data spell "Flavouring Agent"), so none is left unmatched. Flavouring agents are the largest class.

Although almost all 85 FAS codes are referenced, the ones collectively accounting for 50% of the records are 40, 42, 44, 46, 48, 50, and 52

Data Records, paragraph 4

The seven named codes hold 3036 rows, 46.34% of all 6552 rows (F1, within 5 points of 50%) under both FAS readings; over the rows with a FAS code (F2) 56.94% with the deposited column (A, outside tolerance) and 53.42% after add_metadata.R (B, within). They are not the seven most frequent; FAS 59 (351 rows) and FAS 54 (287) are more frequent than FAS 52 (286) under both readings, and FAS 60 (353) too under B; the named codes rank 1 to 6 and 9 (A) or 10 (B). The deposited column has 81 distinct non-empty values, which meets the 77 the criterion takes for "almost all 85", and 73 after add_metadata.R, which does not. The 81 include malformed values (eight such as "77 JECFA 86", five decimals such as "67.29", and "617") and "86", above 85. Counting only whole numbers from 1 to 85 gives 66 codes (A) and 71 (B), both below 77, so "almost all 85" does not hold under a cleaned reading; codes 2, 3, 4, 7, 9, 11, 61, 63, 66, 76, 79, 81, 84 and 85 are absent under both. The criterion sets the outcome by the share and the ranking, so this does not change it.

How the test differed from the paper’s method

• The criterion names the FAS column of both process_df.R and add_metadata.R; the deposit holds the first. Both are run (A and B), and both give partially reproduced.

The unknowns have only the JECFA name without associated compound names or additional information.

Data Records, paragraph 3

95 rows carry only a URL and a JECFA name, and each of them has both class and year unknown. They are 95 of 110 rows with an unknown class (86.4%) and 95 of 118 with an unknown year (80.5%), both below the criterion's 90%. The criterion states only when the claim is reproduced, so the outcome is not reproduced (changed from partially reproduced after verification round 1). The other 15 rows with an unknown class carry a year, a report and identifiers; 13 of them lie between pages 6385 and 6450 and 13 were evaluated in 2008. The other 23 with an unknown year all carry a functional class (13 food additives, 9 food contaminants, 1 flavouring agent), 15 of them chemical names and 14 synonyms; 11 of them are pages 6476 to 6533. The statement holds exactly for the 95 rows unknown on both class and year. Whether "the unknowns" means those rows is asked of a reader in questions/figshare-26877211.yaml; the claim is tested as the claims file has it.

The search functionality of the portal for chemicals by the first character yielded 6,547 items. This number aligns with the number of websites containing at least a JECFA name (n = 6,551 minus a duplicate and a test page)

Technical Validation, paragraphs 2 and 3

The portal's 6547 is outside data, not public in its May 2024 state, and is not tested. In the dataset all 6552 rows carry a JECFA name, within the tolerance of the paper's 6551 but one more. After the cleaning of unique-pages, D2 and D3 give 6550, 3 more than 6547 and outside the margin of 2 the criterion takes for "aligns"; D1 gives 6552. The paper's own numbers (6551 less 2 is 6549) are within 2 of 6547; the one extra row in the deposit is what puts the count outside. The data do not explain the gap between 6552 and 6551; leaving out the test page, named "TEST", gives 6551, but the paper then removes the test page again. Removing both rows of page 6316's chemical with the test page gives 6549, within 2 of 6547, but that is none of the readings D1 to D3 and is not used. The criterion states only when the dataset's part is reproduced and gives no partial level; one of its two conditions fails, so the dataset's part is not reproduced, read as for unknowns-name-only. The portal part stays untested. The outcome was partially reproduced until verification round 3, which agreed with the change. The margin of 2 is the claims file's choice and puts the paper's own 6549 against 6547 at its edge, so this outcome rests on one row. No measured value is given, as the first expected value (the portal count) cannot be measured.