| |
| provenance:design:existing_datasets [2026/08/28 19:04] – Provenance for design:existing_datasets: populations, the three-field disagreement, the fold and its full 3,192-string residue, the residue hand-audit, quote checks, external sources verified and rejected, judgement calls, the four-pass review log, and th karel.kubicek.claude | provenance:design:existing_datasets [2026/08/28 19:05] (current) – Add the mechanical-guard run (check_wrap, check_tables, check_page_numbers) and the hand-read of its 29 unaccounted figures to the review log. Authored by Claude karel.kubicek.claude |
|---|
| * A section on **reusing another paper's released artifact** was missing, even though the item this page answers names "last year's crawl" explicitly. Added, built on two figures the page already has (9 of 60 residue strings name a prior paper's dataset by citation; 57.7% artifact release) plus {[demir2022_reproducibility]}. The Demir citation was first written as "measured how much a repeated crawl of the same list differs from itself", which is **not** what that paper did; corrected to its actual design (4.5M pages, 24 setups) and its own sentence. | * A section on **reusing another paper's released artifact** was missing, even though the item this page answers names "last year's crawl" explicitly. Added, built on two figures the page already has (9 of 60 residue strings name a prior paper's dataset by citation; 57.7% artifact release) plus {[demir2022_reproducibility]}. The Demir citation was first written as "measured how much a repeated crawl of the same list differs from itself", which is **not** what that paper did; corrected to its actual design (4.5M pages, 24 setups) and its own sentence. |
| * The dataset-family table originally showed 24 of the fold's 35 Internet/web/security families, cutting the tail silently — including a 34-paper family. Replaced with the complete table generated from ''--wiki'', plus a sentence naming the counts in the four scopes that are **not** tabulated, so nothing is dropped without a number. | * The dataset-family table originally showed 24 of the fold's 35 Internet/web/security families, cutting the tail silently — including a 34-paper family. Replaced with the complete table generated from ''--wiki'', plus a sentence naming the counts in the four scopes that are **not** tabulated, so nothing is dropped without a number. |
| | |
| | ==== Mechanical guards, after every fix ==== |
| | |
| | Run against the final page text, not against a draft: |
| | |
| | <code> |
| | node scripts/check_wrap.mjs pages/draft/design_existing_datasets.txt pages/draft/provenance_design_existing_datasets.txt |
| | node scripts/check_tables.mjs pages/draft/design_existing_datasets.txt pages/draft/provenance_design_existing_datasets.txt |
| | node scripts/check_page_numbers.mjs pages/draft/design_existing_datasets.txt scripts/report_existing_datasets-output.txt |
| | </code> |
| | |
| | ''check_wrap'' and ''check_tables'' pass on both pages. ''check_page_numbers'' reports **29 unaccounted figures**, and all 29 were read by hand: every one is from a cited paper (Huang et al.'s 95 / 52.07 / 85.71 / 75.00 / 80.00, Wu et al.'s 1,407 / 49 / 72 / 24.69 / 19.17, Izhikevich et al.'s 443, Demir et al.'s 4.5, the IMC 2016 and IMC 2018 exemplars' 09, 03, 4.1 and 489,580,002), an external primary source (27,616,457 / 308 / 13.6 / 200 / 405 / 86,372 / 6962 / 2019), a hypothetical in the page's own prose (a 10,000-host validation sample), or plain arithmetic on a figure the report **does** print (61.9% = 100 − 38.1%). **No corpus figure on the page is missing from the report.** That is the only thing this guard can prove, and it is worth saying explicitly: it traces digits, so it cannot tell a right number from a wrong one, and it never sees a quoted phrase. |
| |
| ==== Pass D — generic (Fable) ==== | ==== Pass D — generic (Fable) ==== |
